Changelog

Changelog#

Full Changelog#

Changelog#

All notable changes to MassGen will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]#

Recent Releases#

v0.1.87 (May 15, 2026) - Documentation: Framework Comparisons & llms.txt Documentation release adding three “MassGen vs …” comparison pages (CrewAI, LangGraph, AutoGen/AG2), a curated llms.txt index plus full-corpus llms-full.txt dump (per llmstxt.org spec), and small README/landing-page pointers so AI agents and crawlers can discover the docs. Also ships a one-line refine=False fix for the bootstrap_subagent discriminator that was being shadowed by the orchestrator’s default max_new_answers_per_agent.

v0.1.86 (May 13, 2026) - bootstrap_subagent Discriminator + Codex MCP Approval Fix Variant B (criteria_mode: bootstrap_subagent) is now functional: the orchestrator runs an in-process critic between rounds, merges critic-proposed criteria into the accumulator, and augments the next round’s checklist. This release also fixes Codex MCP tool calls under codex exec by writing the approval bypasses needed for non-interactive runs.

v0.1.85 (May 11, 2026) - Discriminative Criteria Emergence (criteria_mode) New orchestrator.coordination.criteria_mode option lets evaluation criteria emerge from observed gaps across rounds instead of being pre-authored. bootstrap_inline variant is fully functional on all backends with checklist tool support — agents emit proposed_criteria alongside submit_checklist, the accumulator dedupes/caps, and the next round’s checklist is augmented automatically.

v0.1.84 (May 8, 2026) - TUI Consensus Map A compact visual map below the agent status ribbon during multi-agent runs. Shows agent nodes with latest answer labels, vote arrows, current vote leader, winner state, and waiting/working indicators — driven by existing coordination events without backend schema changes. Hidden on welcome and single-agent runs.

[0.1.87] - 2026-05-15#

Added#

Framework Comparison Pages (#1094): Three new “MassGen vs …” pages under docs/source/reference/comparisons/ — crewai.rst, langgraph.rst, autogen.rst. Each page positions MassGen’s parallel-refinement-with-voting model against the target framework’s coordination shape and lists when to reach for one versus the other
llms.txt Index (#1094): Curated llmstxt.org-spec index published at the docs site root via Sphinx html_extra_path (docs/source/_extra/llms.txt) — gives AI agents a small, hand-picked map of the docs
llms-full.txt Corpus (#1094): Concatenated full-docs dump (~1 MB across 59 files), generated by a Sphinx build-finished hook in docs/source/conf.py and shipped alongside llms.txt for crawlers that want the complete corpus
Docs Landing Page Update (#1094): “How Does MassGen Compare?” section on docs/source/index.rst now lists all four comparisons (LLM Council + the three new ones), with the parent docs/source/reference/comparisons.rst losing its “coming soon” note and gaining a toctree
README Pointers (#1094): One-line pointers in README.md (and synced README_PYPI.md) directing AI agents to llms.txt / llms-full.txt

Fixed#

bootstrap_subagent Discriminator Single-Shot (#1094): Orchestrator._run_bootstrap_discriminator_step now passes refine=False to SubagentManager.spawn_subagent. This is the canonical single-shot knob that SubagentManager actually respects at the orchestrator level — without it, the orchestrator’s max_new_answers_per_agent: 3 default shadowed the coordination-dict overrides, letting the discriminator refine instead of single-shot. Found via live log inspection (log_20260513_095921_816676)
- massgen/orchestrator.py:1298 — refine=False added to spawn_subagent call
- massgen/tests/test_bootstrap_criteria.py — new assertion that discriminator must pass refine=False to spawn_subagent for single-shot

Documentations, Configurations and Resources#

Comparison pages: docs/source/reference/comparisons/{crewai,langgraph,autogen}.rst
Sphinx build-finished hook: docs/source/conf.py — generates llms-full.txt from the source tree at build time
README pointers: README.md, README_PYPI.md — AI agents are directed to llms.txt / llms-full.txt

Notes#

Originally-planned Image/Video Edit Capabilities (#959) and Discriminative Criteria Refinements deferred to v0.1.88.
Closes #1082 (publish llms.txt + llms-full.txt) and #1083 (CrewAI / LangGraph / AutoGen comparison pages).

Technical Details#

Major Focus: Make MassGen discoverable to AI agents and crawlers, and give human readers structured “MassGen vs …” comparisons against the three frameworks most often asked about
PRs Merged: #1094
Issues Closed: #1082, #1083
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.86] - 2026-05-13#

Added#

Functional bootstrap_subagent Variant: orchestrator.coordination.criteria_mode: bootstrap_subagent now runs a between-rounds LLM critic via Orchestrator._run_bootstrap_discriminator_step(). The critic reads the task and each agent’s latest answer, emits proposed_criteria as JSON, and the orchestrator merges them into the accumulator for the next round’s checklist.
Discriminator De-Duping Gate: _maybe_run_bootstrap_discriminator runs the critic once per unique answer snapshot, avoiding repeated critiques when the visible answer set has not changed.
Session-End Criteria Drain: Orchestrator._drain_at_session_end forces a final drain before final presentation so late stdio JSONL emissions are not stranded after the last checklist resolution pass.

Fixed#

Codex MCP Approval Bypass: CodexBackend._write_workspace_config now writes both top-level approval_policy = "never" and per-MCP-server default_tools_approval_mode = "approve" for non-interactive approval modes. This prevents external MCP tools such as submit_checklist, create_task_plan, new_answer, and read_media from failing immediately with “user cancelled MCP tool call” under codex exec.

Documentations, Configurations and Resources#

Updated Config: massgen/configs/coordination/bootstrap_subagent_criteria.yaml now documents the v0.1.86+ active critic-driven flow.

Tests#

massgen/tests/test_bootstrap_criteria.py — expanded to 35 tests covering session-end drain, mocked discriminator spawning and merge behavior, static/inline no-op paths, and empty-answer no-op behavior.
massgen/tests/test_codex_native_hook_adapter.py::TestCodexWorkspaceApprovalPolicy — covers Codex workspace approval policy output across approval modes.

Notes#

Image/Video Edit Capabilities (#959) remain deferred to v0.1.87.

Technical Details#

Major Focus: Complete the discriminative criteria emergence story by making the dedicated critic-driven path functional, and restore Codex MCP tool-call reliability for non-interactive automation.
PRs Merged: #1090
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.85] - 2026-05-11#

Added#

Discriminative Criteria Emergence: New orchestrator.coordination.criteria_mode option lets evaluation criteria emerge from observed gaps across rounds, instead of requiring them to be authored upfront via --eval-criteria or --checklist-criteria-preset. Two variants:
- bootstrap_inline (fully functional on all backends with checklist tool support — SDK and stdio): each agent emits a short proposed_criteria list alongside its submit_checklist call — criteria a stronger answer would satisfy that the current answers do not. Proposals are deduped by exact text, FIFO-capped (bootstrap_max_total, default 30), persisted to bootstrap_criteria_accumulator.json in the session log dir, and merged into the next round’s effective checklist via the existing EvaluationSection machinery. SDK path (Claude Code) gets the field directly in the in-process tool schema; stdio backends (gemini, codex, response, chat_completions, claude, grok) get a JSONL emission channel — proposed_criteria.jsonl next to the checklist specs, drained by the orchestrator on each criteria resolution.
- bootstrap_subagent (wired, LLM step deferred): same accumulator pipeline but criteria are intended to come from a between-rounds critic rather than the agents. The accumulator still propagates seeded entries; the in-process LLM discriminator pass is queued for v0.1.86.
massgen/bootstrap_criteria.py (new module): houses merge_proposals, augment_with_accumulator, is_bootstrap_mode, and validate_criteria_mode — pure helpers shared between orchestrator and tests.
Coordination Config Fields: CoordinationConfig.{criteria_mode, bootstrap_max_per_agent_per_round, bootstrap_max_total} — parsed in cli.py:_parse_coordination_config, validated in CoordinationConfig._validate_criteria_mode, excluded from API params in backend/base.py:get_base_excluded_config_params.
SDK Path Wiring (Orchestrator._init_checklist_tool_sdk): the submit_checklist schema gains an optional proposed_criteria array in bootstrap_inline mode only; static-mode agents see the historical schema unchanged. Parsed proposals land on AgentState.criteria_proposals and are drained into the orchestrator’s accumulator on each criteria resolution.
Stdio Path Wiring (massgen/mcp_tools/checklist_tools_server.py): the FastMCP submit_checklist tool conditionally adds proposed_criteria to its inspect.Signature when state["criteria_mode"] == "bootstrap_inline". Emissions are appended to proposed_criteria.jsonl in the specs directory; Orchestrator._drain_pending_criteria_proposals reads and truncates the file each pass.

Why This Matters#

Removes a cold-start friction: users no longer need to pre-author criteria for a new task. The first round produces both answers and the criteria the second round must rise to.
Anti-Goodhart by construction — criteria come from observed gaps, not priors that may not match the task.
Uses MassGen’s multi-round/multi-agent shape directly; the cross-agent channel (workspace sharing) already existed, so no new transport was needed.

Documentations, Configurations and Resources#

New Configs: massgen/configs/coordination/bootstrap_inline_criteria.yaml and bootstrap_subagent_criteria.yaml (forked from features/fast_iteration.yaml) — runnable examples for both variants
Updated docs/modules/coordination_workflow.md: new section documenting criteria_mode, accumulator semantics, and the two variants

Tests#

massgen/tests/test_bootstrap_criteria.py — 30 new tests (476 lines) covering merge/dedup/cap, config validation, AgentState.criteria_proposals field, _resolve_effective_checklist_criteria augmentation across criteria sources, EvaluationSection rendering gating, _drain_pending_criteria_proposals behavior, and round-N → round-N+1 propagation end-to-end

Notes#

Originally-planned Image/Video Edit Capabilities (#959) deferred to v0.1.86.
bootstrap_subagent LLM discriminator pass queued for v0.1.86.

Technical Details#

Major Focus: Let evaluation criteria emerge from the run rather than be pre-authored — anti-Goodhart, anti-cold-start, and natively shaped by MassGen’s multi-round refinement loop
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.84] - 2026-05-08#

Added#

TUI Consensus Map (#1085): Compact visual map mounted below the agent status ribbon during multi-agent runs that summarizes coordination state without replacing the timeline. Shows one node per agent with latest answer labels, vote direction arrows, current vote leader, winner state, and waiting/working indicators. Hidden on welcome screen and single-agent runs
- massgen/frontend/displays/textual_widgets/consensus_map.py — new ConsensusMapState, ConsensusAgentState, ConsensusMapSnapshot, and ConsensusMap widget
- massgen/frontend/displays/textual_widgets/__init__.py — widget export
- massgen/frontend/displays/textual_terminal_display.py — mounting below status ribbon, event/status wiring, visibility logic
- massgen/frontend/displays/tui_event_pipeline.py — event routing for the map
- massgen/frontend/displays/textual_themes/base.tcss — consensus map theme styling
Event-Driven State Updates (#1085): The Consensus Map subscribes to existing structured coordination events (answer_submitted, vote, agent_stopped, winner_selected, final_presentation_start, agent_restart, phase_change, context_received) — no backend schema changes required
Direct-Callback Fallback (#1085): When direct TUI callbacks update agent status or votes (without the unified event pipeline), the map remains accurate for the same visible state

Documentations, Configurations and Resources#

OpenSpec Change Proposal: openspec/changes/add-tui-consensus-map/{proposal,tasks}.md and specs/textual-tui/spec.md — full design proposal, scenario coverage, and validation tasks

Tests#

massgen/tests/frontend/test_consensus_map.py — unit tests for state transitions and Textual widget compact rendering / visibility (244 lines)
massgen/tests/frontend/test_timeline_snapshot_scaffold.py — runtime TUI snapshot coverage for answer/vote/winner state (+68 lines)
massgen/tests/frontend/__snapshots__/test_timeline_snapshot_scaffold/test_timeline_snapshot_real_tui_consensus_map.svg — golden TUI snapshot

Notes#

Originally-planned Image/Video Edit Capabilities (#959) deferred to v0.1.85.

Technical Details#

Major Focus: Make the physical shape of multi-agent collaboration visible at a glance — convergence, votes, and leader without scanning timelines and toasts
PRs Merged: #1085
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.83] - 2026-05-01#

Added#

In-Session Standalone Checkpoint MCP (#1079): The standalone checkpoint MCP server (originally for external hosts like Claude Code) can now be exposed inside a normal MassGen run, so a single-agent session can call its richer init + checkpoint tools and have its own reviewer team evaluate plans
- massgen/mcp_tools/standalone/checkpoint_mcp_server.py — wired into in-session orchestration
- massgen/orchestrator.py — orchestrator integration with affordance gating
- massgen/system_message_builder.py, massgen/system_prompt_sections.py — standalone checkpoint prompt section
coordination.standalone_checkpoint Config Block (#1079): New YAML block under orchestrator.coordination with fields:
- enabled (bool, default false) — opt-in gate
- team_config (path) — team YAML the standalone server runs
- mode (generate | verify, default generate) — invalid values fall back to generate with a warning
- single_checkpoint (bool, default false) — one-shot checkpoint per session
- include_workspace_context (bool, default false) — mount parent workspace read-only for reviewers
- massgen/agent_config.py — CoordinationConfig fields and to_dict serialization
- massgen/cli.py — _parse_standalone_checkpoint parser with mode validation
Enhanced Checkpoint Tool Card (#1079): Tool card visualization distinguishes primary operations from system tasks with improved context and result display
- massgen/frontend/displays/textual_widgets/tool_card.py
Example Configs (#1079):
- massgen/configs/checkpoint/standalone_mcp/fast_iteration.yaml — fast-iteration single-agent run with in-session standalone checkpoint
- massgen/configs/checkpoint/standalone_mcp/reviewers.yaml — reviewer team config for the standalone server

Changed#

Single-Agent-Only Affordance Gating (#1079): When standalone_checkpoint.enabled: true is set on a multi-agent parent, the system skips the standalone server with a warning (the standalone server runs its own reviewer panel)
Workspace Metadata Exclusions (#1079): Updated _metadata_dirs in filesystem manager constants to keep standalone-checkpoint metadata out of final snapshots
- massgen/filesystem_manager/_constants.py
Backend & API Param Exclusion Lists (#1079): New coordination keys excluded from forwarded backend/API params
- massgen/backend/base.py, massgen/api_params_handler/_api_params_handler_base.py

Documentations, Configurations and Resources#

New Checkpoint Module Section: docs/modules/checkpoint.md — added “Standalone Checkpoint MCP (in-session)” subsection with config schema, behavior table, and sample config reference
Configuration Examples: massgen/configs/checkpoint/standalone_mcp/{fast_iteration,reviewers}.yaml — runnable examples for in-session standalone checkpoint

Tests#

massgen/tests/test_standalone_checkpoint_config.py — config parsing & defaults
massgen/tests/test_standalone_checkpoint_mcp_config.py — MCP server config wiring
massgen/tests/test_standalone_checkpoint_injection.py — orchestrator-level injection
massgen/tests/test_standalone_checkpoint_prompt.py — prompt section rendering across modes
massgen/tests/test_standalone_checkpoint_backend_parity.py — backend parity coverage
massgen/tests/frontend/test_standalone_checkpoint_tool_card.py — TUI tool card visualization

Notes#

Originally-planned features deferred: Checkpoint Safety Mode for Irreversible Actions (#1026) and Round Evaluator over-indexing fix (#994) deferred to a future release.

Technical Details#

Major Focus: Bringing the standalone checkpoint MCP’s richer planning affordance into single-agent in-session use, with explicit single-agent-only gating
PRs Merged: #1079
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.82] - 2026-04-29#

Added#

TUI Copy Mode (#1076): New Ctrl+Shift+S toggle that releases terminal mouse tracking so users can drag-select text natively and copy with the terminal’s built-in shortcut; press again to restore Textual’s normal mouse behavior. Auto-restores mouse capture on exit if copy mode is active
- massgen/frontend/displays/textual_widgets/copy_mode_banner.py — banner widget and set_terminal_mouse_capture helper
- massgen/frontend/displays/textual_terminal_display.py — action_toggle_copy_mode and CopyModeBanner integration
Checkpoint Workspace Context Option (#1076): New include_workspace_context config field for the standalone checkpoint MCP server — optionally mounts the executor’s workspace directory as read-only context for reviewer agents (default false)
- massgen/mcp_tools/standalone/checkpoint_mcp_server.py
Checkpoint Plan Quality Criteria (#1076): New _build_checkpoint_plan_quality_criteria produces mode-aware quality criteria (single vs. multi-checkpoint) that score selective branch depth and fallback handling in generated plans
Checkpoint Agent Recovery Guidance (#1076): Single-checkpoint mode continuation workflow added to checkpoint_instructions.md — detailed recovery steps for agents when a plan branch resolves to terminate without requiring a re-checkpoint

Changed#

TUI Ribbon Dividers (#1076): Visual separators in the agent status ribbon changed from │ (pipe) to · (dot) for a cleaner look
- massgen/frontend/displays/textual_widgets/agent_status_ribbon.py
Checkpoint “Better Means” Safety Guidance (#1076): Extended checkpoint planning prompt with four axes for recognizing when a cheaper path becomes unsafe: scarcity/contention, external visibility, authority substitution, and scope expansion
- massgen/mcp_tools/standalone/checkpoint_mcp_server.py
Checkpoint Workspace Section Templated (#1076): Workspace section in checkpoint planning prompt now uses a {workspace_section} template variable, with content injected based on include_workspace_context setting

Fixed#

TUI Copy Mode Exit Cleanup (#1076): Mouse tracking is correctly restored before the driver tears down when the user exits while copy mode is active

Documentations, Configurations and Resources#

Updated Checkpoint Instructions: massgen/mcp_tools/standalone/checkpoint_instructions.md — single-checkpoint continuation workflow with agent recovery steps for terminate branches

Technical Details#

Major Focus: TUI copy mode for easier text selection and checkpoint quality/safety improvements
PRs Merged: #1076
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.81] - 2026-04-27#

Added#

Multi-Region Circuit Breaker Failover (Phase 6) (#1072): LLM circuit breaker fails over to backup regions when the primary trips OPEN, with automatic recovery when the primary returns to healthy

Technical Details#

Major Focus: Multi-region failover for production-grade circuit breaker resilience — completes the circuit breaker series (Phase 1-6)
PRs Merged: #1072
Contributors: @amabito, @HenryQi and the MassGen team

[0.1.80] - 2026-04-22#

Added#

Circuit Breaker Adaptive Thresholds (Phase 5) (#1065): Self-tuning thresholds that respond to each backend’s actual failure patterns
Single Checkpoint Mode (#1070): New standalone checkpoint mode — no recheckpointing within a single operation
Draft Plan Verify Mode (#1070): New standalone checkpoint mode — verify a draft plan before executing

Changed#

Effective Threshold Helpers: Extracted helper functions for cleaner threshold computation
Benign Case Clarity (#1070): Clearer benign-case handling in checkpoint flow

Fixed#

Force-Open Metrics (#1065): Gated force_open metrics and log on actual state transition
Preserve _open_until (#1065): Preserved _open_until on force_open with intent comments for clearer semantics

Documentation, Configurations and Resources#

Updated Standalone MCP README: Updated massgen/mcp_tools/standalone/README.md with new checkpoint modes
Updated Checkpoint Instructions: Updated massgen/mcp_tools/standalone/checkpoint_instructions.md

Technical Details#

Major Focus: Adaptive circuit breaker thresholds and new standalone checkpoint modes
PRs Merged: #1065, #1070
Contributors: @amabito, @ncrispino, @HenryQi and the MassGen team

[0.1.79] - 2026-04-20#

Added#

Better Fast Mode Options: New options to control coordination speed — fine-grained speed vs. quality tradeoff

Changed#

Broader Checkpoint Framing: Checkpoint mode framing broadened from safety-only to high-stakes and coordinated phases — use for deploys, deletions, financial ops, AND coordinated planning steps
Checkpoint Instructions Clarity: More clarity in trust settings for checkpoint agents

Documentation, Configurations and Resources#

Updated Checkpoint Module: Updated docs/modules/checkpoint.md with broadened framing
Updated Fast Iteration Config: Updated massgen/configs/features/fast_iteration.yaml with new speed options
Updated Standalone MCP README: Updated massgen/mcp_tools/standalone/README.md
Updated Checkpoint Instructions: Updated massgen/mcp_tools/standalone/checkpoint_instructions.md with trust setting clarity

Technical Details#

Major Focus: Fast mode speed control and broader checkpoint framing
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.78] - 2026-04-17#

Added#

Circuit Breaker Distributed Store (Phase 4) (#1061): Pluggable state store for the LLM circuit breaker. Previously each process kept its own CB state, so one worker tripping OPEN did not stop siblings from hammering a rate-limited upstream. CB state (failure counts, open/half-open/closed, cooldown timers) can now be shared across workers and processes. Default (store=None) keeps the existing single-process path unchanged.
- CircuitBreakerStore Protocol: the interface the CB uses to persist state
- InMemoryStore (CB state store): thread-safe, zero-deps — useful for single-process and tests
- RedisStore (distributed CB state store): shares CB state across processes via Redis (redis>=4.0, lazy-imported); available through the optional redis-store extra
- Atomic atomic_record_failure / atomic_record_success so CB state transitions are linearizable when workers race on the same backend
Optional Redis Dependency Group: New redis-store extra for the Redis-backed CB store — install with pip install massgen[redis-store]

Tests#

CB Store Unit Tests (#1061): New massgen/tests/test_cb_store.py covering InMemoryStore and RedisStore behavior, Protocol contract, and metrics integration
CB Store Adversarial Tests (#1061): New massgen/tests/test_cb_store_adversarial.py covering TOCTOU races, Redis eviction, corrupted state handling, CAS semantics, probe-claiming, and TTL edge cases

Documentation, Configurations and Resources#

New Roadmap: New ROADMAP_v0.1.79.md for the next release
Updated pyproject.toml: Added redis-store optional dependency group

Technical Details#

Major Focus: Distributed circuit breaker state — completes the CB observability stack started in v0.1.72 / v0.1.76
PRs Merged: #1061
Contributors: @amabito, @ncrispino, @HenryQi and the MassGen team

[0.1.77] - 2026-04-15#

Added#

Answer Now Button (#1062): New “Answer Now” button lets agents submit answers more quickly, both within a round, and bypassing additional refinement rounds when quality is already sufficient

Changed#

Updated Checkpoint Instructions: Refined agent memory instructions for checkpoint MCP
Updated Coordination Workflow Docs: Clarified coordination workflow documentation

Technical Details#

Major Focus: Answer Now Button — faster answers when quality is sufficient
PRs Merged: #1062
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.76] - 2026-04-13#

Added#

Exa AI Search Tool (#1057): New Exa AI-powered search tool added to MCP server registry with example config
Circuit Breaker Observability (Phase 3) (#1056): Observability module with probe ownership, lock release mechanisms, and per-attempt latency regression tracking
Checkpoint Agent Instructions (#1058): Copyable custom instructions for agent memory files with checkpoint MCP information

Fixed#

Docker Dependencies (#1058): Fixed Dockerfile installs for reliable container builds
Circuit Breaker Strengthening (#1056): Strengthened observability across all backends

Documentation, Configurations and Resources#

Updated MCP Server Registry: Updated docs/source/reference/mcp_server_registry.rst with Exa search tool
Updated MCP Integration Guide: Updated docs/source/user_guide/tools/mcp_integration.rst
Updated Standalone MCP README: Updated massgen/mcp_tools/standalone/README.md with checkpoint instructions
New Checkpoint Instructions: New massgen/mcp_tools/standalone/checkpoint_instructions.md
New Config: New massgen/configs/tools/web-search/exa_search_example.yaml

Technical Details#

Major Focus: Exa AI Search & Circuit Breaker Observability (Phase 3)
PRs Merged: #1056, #1057, #1058
Contributors: @amabito, @HenryQi, @ncrispino, @teocollazo and the MassGen team

[0.1.75] - 2026-04-10#

Added#

Codex Native Hooks (#1053): Hybrid hook system for Codex backend combining native hooks and MCP capabilities
Checkpoint WebUI Auto-Launch (#1053): Checkpoint workflows now auto-launch the WebUI with configurable host/port for visual monitoring
Standalone MCP Server Documentation: Guide for massgen-checkpoint-mcp with setup, examples, troubleshooting, and safety policy integration

Changed#

Checkpoint Planning Improvements (#1053): Precondition validation and recovery tree support; user/system prompt and eval criteria pass-through to checkpoint agents
Safety Policy Update: Updated safety policy for checkpoint based on Claude Code safe mode

Fixed#

WebUI Automation Redirect (#1053): Fixed erroneous setup redirect during automation mode

Documentation, Configurations and Resources#

Updated Coordination Workflow: Updated docs/modules/coordination_workflow.md with hook architecture and delivery rules
Updated Injection Guide: Updated docs/modules/injection.md
Standalone MCP README: New comprehensive massgen/mcp_tools/standalone/README.md

Technical Details#

Major Focus: Codex Hooks & Checkpoint WebUI — deeper Codex integration and visual checkpoint monitoring
PRs Merged: #1053
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.74] - 2026-04-08#

Changed#

Checkpoint MCP Improvements (#1050): Major enhancements to standalone checkpoint MCP server (massgen/mcp_tools/standalone/checkpoint_mcp_server.py) — refinements to subprocess execution, isolation, workspace handling, and event relay
Pre-collab Criteria Refinements (#1050): Improvements to evaluation criteria generation in precollab_utils.py

Fixed#

Duplicate Tool Calls (#1050): Resolved duplicate tool call issues in base_with_custom_tool_and_mcp.py, chat_completions.py (including for MiniMax on OpenRouter), and response.py backends

Documentation, Configurations and Resources#

Updated Checkpoint Module: Updated docs/modules/checkpoint.md with checkpoint MCP improvements
OpenSpec Updates: Updated openspec/changes/update-checkpoint-coordination-objectives/ design, spec, and tasks

Technical Details#

Major Focus: Checkpoint MCP improvements and stability fixes
PRs Merged: #1050
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.73] - 2026-04-06#

Added#

Eval Criteria Evolver Subagent (#1047): New subagent type that evolves evaluation criteria across rounds — sharper, more opinionated criteria as the run progresses
Checkpoint Objective Mode (Initial Draft) (#1047): Initial draft of checkpoint MCP with objective mode for safety planning of irreversible actions (deletions, deployments, financial operations); returns ordered plan with per-step constraints and recursive recovery trees

Changed#

Improved Eval Criteria Visibility: See what criteria agents are working against, more clearly
Trace Analyzer Improvements: Refinements to trace analyzer subagent behavior

Fixed#

Evolver Fixes: Stability fixes for the criteria evolver subagent

Documentation, Configurations and Resources#

Updated Checkpoint Module: Updated docs/modules/checkpoint.md with objective mode documentation
OpenSpec Change: New openspec/changes/update-checkpoint-coordination-objectives/ proposal and spec for objective mode

Technical Details#

Major Focus: Eval Criteria Evolver & Checkpoint Objectives — self-improving criteria and safety planning
PRs Merged: #1047
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.72] - 2026-04-03#

Changed#

Grok Backend Update (#1044): Updated Grok backend with latest improvements

Added#

Circuit Breaker Phase 2 (#1038): LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only in v0.1.68); Gemini also handles 503 errors
Config Plumbing Smoke Tests (#1038): Smoke tests verify circuit breaker wiring and API call timing for all backends

Fixed#

Response API Timing (#1038): Added start/end API call timing to ResponseBackend non-MCP path

Technical Details#

Major Focus: Circuit Breaker Phase 2 — rate limit protection across all major backends
PRs Merged: #1038, #1044
Contributors: @amabito, @HenryQi, @ncrispino and the MassGen team

[0.1.71] - 2026-04-01#

Changed#

Better Evaluation Criteria: Improved criteria generation for higher-quality, more opinionated output
System Prompt Tuning: Adjusted system prompts for better agent performance across coordination rounds

Fixed#

Final Injection Fix: Corrected injection behavior at the final stage
Eval Criteria GPT Pre-Collab Fix: Resolved evaluation criteria issues with GPT models during pre-collaboration phase
Execution Trace Analyzer Launch Fix: Trace analyzer now starts correctly
Trace Memory Fix: Corrected memory handling in execution traces
Auto Round Memory Fix: Fixed automatic round handling for memory

Documentation, Configurations and Resources#

Updated Log Analyzer Skill: Updated massgen/skills/massgen-log-analyzer/SKILL.md
Updated Execution Trace Analyzer: Updated massgen/subagent_types/execution_trace_analyzer/SUBAGENT.md

Technical Details#

Major Focus: Stability and polish for v0.1.70’s evaluation criteria system
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.70] - 2026-03-30#

Added#

Evaluation Criteria Redesign (#1035): Three-tier categorization (primary, standard, stretch) with anti-pattern definitions per criterion and aspiration statements
Improved Checklist-Gated Evaluation (#1035): Tighter iterative submission cycles — improved scoring, gap analysis, and improvement proposals drive more meaningful iteration before final voting
Fast Iteration Mode (#1035): Streamlined multi-round submission phases via fast_iteration.yaml config
WebUI Review Modal (#1035): Approve and comment on outputs directly in the browser when working in git
Background Trace Analysis (#1035): Execution trace analyzer starts automatically from round 2

Changed#

Improved Evaluation Criteria Generation (#1035): Criteria generation now produces opinionated, task-specific criteria with aspiration statements
Enhanced Workspace Cleanup (#1035): Improved isolation between rounds
Refined Per-Round Token Tracking (#1035): More accurate per-round token usage tracking

Fixed#

Subagent Fixes (#1035): General fixes for subagent behavior and path issues

Documentation, Configurations and Resources#

Updated Coordination Workflow: Updated docs/modules/coordination_workflow.md with checklist-gated workflow documentation
Updated Subagents Guide: Updated docs/modules/subagents.md with background trace analysis
New Injection Guide: New docs/modules/injection.md for injection documentation
Updated Concepts Guide: Updated docs/source/user_guide/concepts.rst with evaluation criteria redesign
Updated YAML Schema: Updated docs/source/reference/yaml_schema.rst with new configuration options
Updated MassGen Skill: Updated massgen/skills/massgen/SKILL.md with opinionated criteria format
Updated Criteria Guide: Updated massgen/skills/massgen/references/criteria_guide.md with three-tier system
New Config: New massgen/configs/features/fast_iteration.yaml for fast iteration mode

Technical Details#

Major Focus: Evaluation Criteria Redesign — three-tier categorization with anti-patterns and checklist-gated workflow
PRs Merged: #1035
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.69] - 2026-03-27#

Added#

WebUI Automation Auto-Start (#1032): Automation mode now auto-starts coordination runs without browser interaction — open the URL at any point to monitor progress, even mid-run
MassGen Skill Redesign (#1032): Increased usability and integration with the WebUI; skill now launches the WebUI for live session tracking
Quickstart Wizard Rework (#1032): New WelcomeStep, SkillsStep, ApiKeyStep redesign, DockerStep expansion, and SetupModeStep restructure for smoother onboarding
Workspace Browser Expansion (#1032): WorkspaceModal and improved workspace connection

Changed#

Flexible Evaluation Criteria Fields (#1032): Criteria JSON now accepts description or name as alternatives to text field for more flexible criterion authoring
Automatic Config Resolution (#1032): Automation mode auto-resolves config when none is specified (same as CLI without --web)

Fixed#

Web Automation Skill Lifecycle (#1032): Web automation now correctly auto-ends when a skill completes
WebUI Version Default (#1032): Fixed WebUI defaulting to v2

Documentation, Configurations and Resources#

Updated WebUI Guide: Updated docs/source/user_guide/webui.rst with automation mode flags, auto-start behavior, and interactive examples
MassGen Skill: Updated massgen/skills/massgen/SKILL.md with WebUI wrapper and monitoring instructions
Advanced Workflows: Updated massgen/skills/massgen/references/advanced_workflows.md with skill WebUI integration patterns
Config Setup: Updated massgen/skills/massgen/references/config_setup.md with updated quickstart guidance

Technical Details#

Major Focus: WebUI Automation & Improved Skill — seamless integration between the skill workflow and WebUI monitoring
PRs Merged: #1032
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.68] - 2026-03-25#

Added#

Checkpoint Coordination Mode (#1028): New delegator pattern — main agent plans solo then calls checkpoint() to delegate execution to fresh agent instances with clean backends and cloned workspaces
WebUI Checkpoint Support (#1028): Checkpoint mode display integrated into the modernized WebUI
LLM API Circuit Breaker (#1024): Automatic 429 rate limit handling with circuit breaker pattern for Claude backend

Fixed#

LiteLLM Supply Chain Fix (#1025): Pinned litellm<=1.82.6 and committed uv.lock to prevent dependency attacks

Technical Details#

Major Focus: Checkpoint Mode — delegator pattern for multi-agent coordination
PRs Merged: #1028, #1025, #1024
Contributors: @ncrispino, @amabito, @HenryQi and the MassGen team

[0.1.67] - 2026-03-23#

Added#

Modernized WebUI (#1016): Complete UI redesign with inline final answers, keyboard shortcuts, and Zustand state management (message, mode, tile, agent, theme stores)
RoundBudgetGuardHook (#1013): Per-round cost enforcement with configurable warning thresholds (50%, 75%, 90%) and graceful termination on budget overrun
Unified Pre-Collab Phases (#1016): Persona generation, evaluation criteria, and prompt improvement now run in parallel with unified TUI batch display
Regression Guard (#1016): Blind A/B verification subagent before submitting revisions to catch silent regressions

Technical Details#

Major Focus: Modernized WebUI and quality improvements
PRs Merged: #1016, #1013
Contributors: @ncrispino, @amabito, @HenryQi and the MassGen team

[0.1.66] - 2026-03-20#

Added#

Step Mode (#1011): New --step CLI flag runs a single agent for one iteration then exits, loading/writing state from a session directory — building block for external orchestrators like massgen-refinery
Console Text Sanitization (#1010): Reusable sanitize_console_text utility for safe TUI and logger rendering

Fixed#

Codex Windows UTF-8 (#1010): Ensure UTF-8 encoding when writing files in Codex backend
TUI Event Pipeline (#1010): Console safety features for logger and text sanitization in event pipeline

Technical Details#

Major Focus: Step Mode — building block for external orchestrators
PRs Merged: #1011, #1010
Contributors: @ncrispino, @praneeth999, @HenryQi and the MassGen team

[0.1.65] - 2026-03-18#

Added#

Quality Server (#1007): Standalone massgen_quality_tools MCP server with session-based checklist evaluation, configurable scoring thresholds, improvement proposals, and coverage validation
Workflow Server (#1007): Standalone massgen_workflow_tools MCP server with multi-round answer submission, automatic deliverable snapshots, and vote support
Media Server (#1007): Standalone massgen_media_tools MCP server with image/video/audio generation and critical-first media analysis

Technical Details#

Major Focus: MassGen Refinery Plugin — standalone MCP servers for Claude Code
PRs Merged: #1007
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.64] - 2026-03-16#

Added#

Gemini CLI Backend (#999, #952): New subprocess-based backend for Google’s Gemini CLI with session persistence, MCP tools via .gemini/settings.json, and Docker support
WebSocket Mode (#990): Persistent WebSocket transport for OpenAI Response API with auto-reconnection and real-time event streaming
Execution Trace Analyzer (#1002): New subagent type for mechanistic analysis of agent execution traces with 7-dimension evaluation framework
Copilot Docker Mode (#999): Containerized tool execution for Copilot backend with sudo and network configuration

Fixed#

Response API Duplicates (#1000): Prevent duplicate item errors in recursive tool loops

Technical Details#

Major Focus: Gemini CLI Backend
PRs Merged: #999, #990, #1002, #1000
Contributors: @praneeth999, @ncrispino, @HenryQi, @db-ol and the MassGen team

[0.1.63] - 2026-03-13#

Added#

Ensemble Pattern Defaults (#996): disable_injection and defer_voting_until_all_answered now default to true for ensemble-style subagent orchestration
Transformation Pressure (#996): Round evaluator applies transformation pressure to push agents toward meaningful structural changes
Success Contracts (#996): Explicit quality gates that agents must satisfy before the round evaluator allows convergence

Changed#

Lighter Refinement (#996): Subagents use lighter refinement prompts to reduce token overhead and latency
Killed Agent Handling (#996): Graceful management of agents that time out or fail mid-round
Verification Replay (#996): Evaluation consistency across rounds via replayed verification context

Fixed#

Timeout Fallback (#996): More robust coordination when agents hit timeout boundaries

Technical Details#

Major Focus: Ensemble & Contracts — ensemble pattern defaults, transformation pressure, success contracts, lighter refinement
PRs Merged: #996 (dev/v0.1.62-p1)
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.62] - 2026-03-11#

Added#

MassGen Skill (#992): New general-purpose multi-agent skill with 4 modes (general, evaluate, plan, spec) for Claude Code and other AI agents
Session Viewer (#992): New massgen viewer command for real-time observation of automation sessions with interactive session picker and web mode
Headless Quickstart (#992): Non-interactive setup via --quickstart --headless for CI/CD integration
Web Quickstart (#992): Browser-based setup flow via --web-quickstart
Skill Auto-Sync (#992): GitHub Actions workflow to auto-sync MassGen Skill to separate repository for easy installation

Changed#

Claude Code Backend (#992): Background task execution support and SDK MCP integration
Codex Backend (#992): Native filesystem access, JSONL event streaming, and MCP tool support
Copilot Model Discovery (#992): Runtime model fetching with metadata caching
Planning & Evaluation (#992): Better planning prompts with thoroughness support, removed should/could criteria to reduce output similarity
CLI Enhancements (#992): --print-backends table, viewer subcommand, multi-agent quickstart via --quickstart-agent

Fixed#

Skill Viewer (#992): Fixed skill viewer display and added convenience shell script
Correctness Prompts (#992): Updated correctness prompts for improved accuracy

Technical Details#

Major Focus: MassGen Skill & Viewer — general-purpose skill, session observation, backend improvements
PRs Merged: #992 (evaluator-skill)
Contributors: @ncrispino (6 commits), @HenryQi (2 commits) and the MassGen team

[0.1.61] - 2026-03-09#

Added#

Round Evaluator Subagent Type (#986): New round_evaluator subagent type that delegates evaluation to specialized evaluator subagents for deeper quality assessment
round_evaluator_example.yaml Config (#986): New example config for the round evaluator paradigm

Changed#

Orchestrator Refactoring (#986): Major orchestrator refactoring (+1,189 lines) to support the round evaluation workflow
Evaluation Prompts (#986): Improved evaluation prompts for clearer, more actionable feedback with task plan injection
Simplified Config (#986): Simplified config handling for evaluation parameters
SUBAGENT.md Generality (#986): Improved SUBAGENT.md for broader subagent compatibility

Fixed#

Session Resumption (#986): Fixed resumption from already-resumed logs
Round Evaluation Prompts (#986): Improved round evaluation prompt clarity

Technical Details#

Major Focus: Round evaluator paradigm — delegated evaluation to specialized subagents
PRs Merged: #986 (improve_verification_time)
Contributors: @ncrispino (8 commits), @HenryQi (1 commit)

[0.1.60] - 2026-03-06#

Added#

read_media Rewrite (#978): Rewritten with clearer schema, better error handling, and improved naming
MediaCallLedgerHook (#978): New MediaCallLedgerHook for tracking read/generate media tool calls via the hook framework
GPT-5.4 Support (#978): New default OpenAI flagship model added to the model registry
Subagent Backend Inheritance (#978): New inherit_spawning_agent_backend option — subagents automatically inherit the spawning agent’s backend
Subagent Final Answer Strategy (#978): New final_answer_strategy option for child orchestrator final-answer policy (winner_reuse, winner_present, synthesize)
Per-Agent Subagent Agents (#978): Per-agent subagent_agents override and orchestrator config file support with robust JSON parsing

Changed#

Decomp Mode Cooperates with Checklist (#978): Decomposition mode now cooperates with the checklist workflow for unified quality-gated subtask iteration
System Prompt Focus (#978): Refocused system prompt on evaluating entire output quality
Verification Prompts (#978): Improved verification_latest prompts for faster verification rounds

Fixed#

Checklist & Proposal Injections (#978): Fixed proposal injection improvements for more reliable checklist behavior
Task Plan Refresh (#978): Fixed task plan refresh during quality rounds
Codex Prompt Caching (#978): Fixed prompt caching calculation for pricing accuracy
Skill Prefix Handling (#978): Fixed skill prefix handling edge cases

Technical Details#

Major Focus: Multimodal tools, subagent enhancements, GPT-5.4, decomp+checklist cooperation
PRs Merged: #978 (improve_verification_time)
Contributors: @ncrispino (6 commits), @HenryQi (1 commit)

[0.1.59] - 2026-03-04#

Added#

Planning Improvements (#969): Smarter quality rounds with improved planning
- Auto-add improvements to task plan for better iteration tracking
- Plan review enhancements for more thorough quality evaluation
Checklist & Evaluation Enhancements (#969): More reliable evaluation pipeline
- Better eval gen config for more accurate quality assessments
- Checklist fixes for consistent behavior across rounds
- Gemini tool name normalization for MCP compatibility (ease for MCP)

Changed#

Subagent Behavior (#969): Adjusted subagent behavior and manager enhancements
- Improved subagent coordination and task delegation
- Docker skill write access fixes for containerized execution
Video Generation Skills (#969): Adjusted video gen skill behavior
- No fallback to animated on errors — fail cleanly instead
- Video understanding criticality improvements
- Impact metric restoration for quality assessment

Fixed#

Answer Anonymization (#969): Fixed answer anonymization during evaluation
Quickstart & Tests (#969): Updated quickstart flow and test suite
Plan & Docker Fixes (#969): Small fixes for plan mode and Docker execution

Technical Details#

Major Focus: Quality round improvements — planning, evaluation, subagents, media fixes
PRs Merged: #969 (improve_quality_rounds)
Contributors: @ncrispino (7 commits), @HenryQi (1 commit)

[0.1.58] - 2026-03-02#

Added#

Comprehensive Multimodal Revamp: Major expansion of multimodal generation and understanding capabilities
- ElevenLabs TTS & STT (#942): High-quality voice synthesis and transcription via generate_media and read_media tools
- Nano Banana 2 Image Generation (#951): New default image generation model with higher quality output
- Grok Image/Video Generation: Grok multimedia generation support via xAI API
- Media Generation Skills: New reusable skills for image, video, and audio generation workflows
- Multi-Turn Image Editing: Continuation IDs for iterative image editing sessions
Nvidia NIM Backend (#962): First-class provider integration for NVIDIA Inference Microservices
- Support for NVIDIA-hosted models via NIM API
- Full integration with MassGen’s multi-agent coordination
Quality Rethinking Subagent (#964): New quality_rethinking subagent type for targeted per-element craft improvements
- Explicit improve/preserve listings in checklists
- Better label refresh ordering for more coherent checklist updates
CLI Mode Flags: New command-line flags mirroring TUI toggles
- --quick, --single-agent, --coordination-mode, --personas flags
- Plan mode accessible from command line

Changed#

Logging Architecture Refactor: Fixed concurrent logging for parallel multi-agent execution with LoggingSession isolation
- Each agent gets isolated logging context preventing log interleaving
Evaluation Criteria Defaults: Sensible defaults for evaluation criteria when not explicitly specified
Checklist Label Refresh Ordering: Improved ordering of checklist label refreshes for better coherence

Fixed#

Subagent Hardening (#964): Better ‘@’ parsing and error handling for multiple submit_checklist calls
- Clearer subagent context and improved error messages
Pre-Collaboration Checklist: Fixed checklist behavior before collaboration phase
Evaluation Criteria Defaults: Fixed default handling for evaluation criteria

Technical Details#

Major Focus: Multimodal revamp, Nvidia NIM backend, quality rethinking subagent, checklist improvements
PRs Merged: #962 (Nvidia NIM), #964 (Subagent hardening)
Contributors: @ncrispino (11 commits), @AbhimanyuAryan (1 commit)

[0.1.57] - 2026-02-27#

Added#

Subagent Delegation Protocol (#955, MAS-325): File-based delegation for container-to-host subagent spawning
- SubagentLaunchWatcher polls shared delegation directory for request files
- Atomic JSON-based DelegationRequest/DelegationResponse exchange protocol
- Workspace path validation against allowlist for security
- Cancel sentinel support for graceful subagent termination
Builder Subagent Type (#955): New subagent for executing substantial pre-specified work with fresh context
- Transformative redesigns, large artifact generation, complex multi-file rewrites
- Prescriptive spec input with positive goals AND forbidden patterns (negative constraints)
- Auto-triggered by checklist when transformative changes identified
Claude Code Reasoning Parameters (#955): Updated SDK integration with new unified reasoning config
- Migrated from deprecated max_thinking_tokens to reasoning config dict
- Supports type (adaptive/enabled/disabled), effort (low/medium/high/max), budget_tokens
- Backward compatible with legacy configurations
Substantiveness Tracking (#955): Checklist captures specific planned changes to prevent satisficing
- List format: transformative, structural, incremental items with descriptions
- decision_space_exhausted flag for convergence signaling
- Builder subagent suggestion when transformative changes identified
- Novelty subagent injection when transformation count = 0 (plateau detection)
Diagnostic Report Gating (#955): Optional quality gate requiring structured diagnostic reports
- Validates report file existence, minimum length, and markdown format
- Required sections: Failure Patterns, Root Causes, Goal Alignment
Verification Subdirectory for Scratch (#955): Organized scratch work with verification subdirectory support

Changed#

Subagent Workspace Management (#955): Auto-mounted parent workspace (read-only) by default via include_parent_workspace
- Eliminates need for context_paths: ["./"] — subagents get parent workspace automatically
- context_paths now for additional paths only (peer workspaces, external resources)
Evaluation Criteria (#955): Cleaned up subagent paths and eval criteria organization
Memory Config Simplified (#955): Simplified memory config option to only final presentation
Per-Agent Checklist Scoring (#955): Support for evaluating multiple agents separately with format detection

Fixed#

Subagent Launch for Codex (#955): Fixed codex backend subagent spawning
Subagent Timing (#955): Improved synchronization and timeout handling
Subagent Temp Dir (#955): Fixed temporary workspace directory support
Subagent Type Initialization (#955): Fixed type definitions and initialization
Test Fixes (#955): Various test updates for new features

Documentation, Configurations and Resources#

New massgen/subagent_types/builder/SUBAGENT.md - Builder subagent type definition
Updated massgen/subagent_types/evaluator/SUBAGENT.md - Enhanced evaluator guidance
New docs/modules/coordination_workflow.md - End-to-end coordination lifecycle documentation
Updated docs/modules/subagents.md - Delegation protocol and workspace management
Updated massgen/configs/BACKEND_CONFIGURATION.md - Reasoning parameter documentation
New ROADMAP_v0.1.58.md - Next release roadmap

Technical Details#

Major Focus: Subagent delegation protocol, builder subagent, convergence improvements
PRs Merged: #955 (Delegation protocol, builder subagent, reasoning params, eval improvements)
Files Changed: 68 files, +7348/-503 lines
New Tests: test_launch_watcher.py, test_launch_watcher_e2e.py, test_subagent_delegated_mode.py, test_round_resume.py, test_checklist_tools_server.py (substantiveness), test_write_mode_scratch.py, test_claude_code_skills_config.py, test_gepa_evaluation_flow.py, test_novelty_injection.py
Contributors: @ncrispino (8 commits), @HenryQi (2 commits)

[0.1.56] - 2026-02-25#

Added#

Critic Subagent (#945): New subagent type for honest, unbiased quality assessment
- Detects genuine vs incremental improvement across refinement rounds
- First impression, quality ceiling assessment, incrementalism verdict, independent E-criterion scoring
- Describes the 10/10 vision and distance to excellence
- Complements existing subagent types (evaluator, explorer, researcher, novelty)
Spec Plan Mode (#945): Formal requirements specification before execution
- plan_mode="spec" for structured requirements gathering
- Spec creation, approval modal, and execution pipeline
- TUI spec mode state with dedicated mode bar support
- Spec storage and changedoc integration
read_media Conversation Continuity (#945): Follow-up conversations on supported media (image) via continue_from conversation_id
- Multi-turn image analysis with severity parsing
ask_others Targeted Messaging (#937): target_agents parameter for focused agent-to-agent communication
- Validation and per-target response counting
- Shadow-agent prompt improvements for prior work separation
Codex OAuth Login Fix (#937, MAS-322): Codex backend always available in WebUI regardless of OPENAI_API_KEY
- OAuth authentication fix via codex login
Background Subagent Continuation (#945): Non-blocking subagent task execution
- Enhanced subagent state tracking and graceful cancellation
Docker Configuration Mounting (#945): Claude and Codex configuration mounting options for Docker containers

Changed#

Evaluation Criteria Taxonomy (#945): Updated from core/stretch to must/should/could tiers
Novelty Subagent Enhancement (#945): Updated guidance for growth-oriented refinement
Multimodal Tool Configs (#945): Updated text-to-image, text-to-speech, and text-to-video generation configs

Fixed#

Test and spec reading fixes (#945)
Audio cleanup for future release stability (#945)

Documentation, Configurations and Resources#

New massgen/subagent_types/critic/SUBAGENT.md - Critic subagent type definition
Updated massgen/subagent_types/novelty/SUBAGENT.md - Enhanced novelty guidance
Updated massgen/tool/_multimodal_tools/TOOL.md - Audio multimodal documentation
Updated massgen/configs/features/background_subagent_example.yaml
Updated multimodal tool configs (text-to-image, text-to-speech, text-to-video)
New ROADMAP_v0.1.57.md - Next release roadmap

Technical Details#

Major Focus: Spec plan mode, targeted messaging, critic subagent
PRs Merged: #945 (Spec mode, critic subagent, audio multimodal), #937 (Codex OAuth, ask_others targeting)
Files Changed: 89 files, +8684/-1089 lines
New Tests: 16 new test files covering spec execution, spec storage, spec approval modal, audio multimodal, read_media analysis/followup, refinement quality, and more
Contributors: @HenryQi (3 commits), @MuL1ian (3 commits), and the MassGen team (4 commits)

[0.1.55] - 2026-02-23#

Added#

Specialized Subagent Types (#938): Discovery-based system for specialized subagent roles via SUBAGENT.md frontmatter
- Built-in types: evaluator (programmatic verification), explorer (investigation), researcher (deep analysis), novelty (breaks refinement plateaus)
- TUI visualization for subagent roles
Dynamic Evaluation Criteria (#938): GEPA-inspired task-specific evaluation criteria generation replacing static E1-E4 items
- Domain-specific presets (persona, decomposition, evaluation, prompt, analysis)
- Core/stretch categorization for smarter convergence off-ramps
- Score scale 0-10
- Config: evaluation_criteria_generator
Native Backend Image Routing (#938, MAS-300): understand_image routes to agent’s own backend (Claude, Gemini, Grok, Claude Code, Codex) instead of always using OpenAI
- Fallback to OpenAI for backends without image_understanding capability
Configurable Video Frame Extraction (#938): Scene-based (PySceneDetect) or uniform extraction modes
- max_frames cost guardrail (default 30, max 60)
- Config: multimodal_config.video
Remotion Skill in Quickstart (#938): Video generation/editing skill installed when selected during quickstart

Changed#

Checklist System Update (#938): T-prefix to E-prefix naming, 0-100 to 0-10 score scale, item_categories for core/stretch, convergence off-ramp when all core items pass
Unified Pre-Collaboration (#938): Persona generation, decomposition, and eval criteria generation unified as composable primitives

Fixed#

Background subagent cancel name fix (#938)
Initial TUI sizing fix (#938)

Documentation#

New docs/modules/composition.md - Composable primitives, phase architecture, domain-specific checklist gates

Technical Details#

Major Focus: Specialized subagent types, dynamic evaluation criteria, native image routing, video frame extraction
PRs Merged: #938 (Subagent roles / specialized types)
Contributors: @ncrispino and the MassGen team

[0.1.54] - 2026-02-20#

Added#

Copilot SDK Backend (#862): New copilot backend using github-copilot-sdk
- Native MCP server integration and custom tool handling
- Session management with cache invalidation
- Auth via GitHub subscription
Subagent Runtime Messaging (#926): New send_message_to_subagent tool to steer running background subagents mid-execution
- Supports per-agent targeting within subagent orchestrators
Gemini 3.1 Pro Support (#926, MAS-312): gemini-3.1-pro-preview model added to capabilities registry
Per-Agent Injection Targeting (#926): Injections can target specific agents or broadcast to all

Changed#

MCP Hooks Improvements (#926): Hook middleware for subagent MCP servers, InjectionDeliveryStatus enum, hook-dir argument for PostToolUse injection
Type Annotation Modernization (#926): Codebase-wide migration from typing.Dict/List/Optional/Union to modern dict/list/X | None syntax

Fixed#

MCP hooks issue fix (#926)
Subagent message sending fix (#926)
fstmcp version fix (#920)

Technical Details#

Major Focus: Subagent runtime messaging, Copilot SDK backend, Gemini 3.1 Pro support
PRs Merged: #862 (Copilot SDK backend), #926 (Subagent messaging), #921 (Cloud infra research), #920 (Minor fixes)
Contributors: @ncrispino and the MassGen team

[0.1.53] - 2026-02-18#

Added#

Background Tool Execution (#917): Non-blocking lifecycle tools for long-running work
- start_background_tool, get_background_tool_status, get_background_tool_result, wait_for_background_tool, cancel_background_tool, list_background_tools
- Compatible with custom tools and MCP server tools
Planning Task Verification (#917): Tasks now require verification and verification_method fields by default
- --no-require-verification flag to opt out
- Framework-injected tasks exempt from verification requirements
TUI Background Job Indicators (#917): Agent status ribbon with background job indicators
- Background tasks modal with lifecycle controls
Subagent Infrastructure (#917): Groundwork for specialized subagent types
- Evaluator and Explorer type definitions via SUBAGENT.md frontmatter

Changed#

Tool Argument Normalization (#917): Consistent argument handling across backends

Fixed#

Task plan verification improvements
Codex reasoning config alignment

Technical Details#

Major Focus: Background tool execution, planning verification, TUI background indicators
PRs Merged: #917 (Background tools & subagent infrastructure)
Contributors: @ncrispino and the MassGen team

[0.1.52] - 2026-02-16#

Added#

Dedicated Final Answer Modal (#901): Tabbed modal with Answer tab (markdown content, post-evaluation, and file list) and Workspace/Review Changes tab (diff review)
- Trophy header with agent identity and model name
- Approve/Reject/Cancel action bar with rework controls for iteration
Substantive Gate (#901): Quality gate preventing coordination from continuing with only incremental changes
- Tracks transformative/structural/incremental classification
- Detects decision_space_exhausted for convergence
- Config: require_substantiveness: true (mandatory in checklist)
Novelty Injection (#901): Creative pressure injection when agents converge or stall
- Levels: none (default), gentle, moderate, aggressive
- Intensifies after restarts
- Config: novelty_injection in coordination section
Agent Identity & Versioning (#901): Unique agent identity with versioned answer labels (e.g., agent1.2)
- answer_label_mapping for provenance tracking
Subagent Evaluation Infrastructure (#901): Foundation for delegating evaluation to spawned subagent instances

Changed#

First Answer Non-Restart (#901): First answer from each agent no longer triggers automatic restarts even if quality checks fail, enabling more natural coordination flow

Fixed#

Approved/rejected state display in final answer card
Auto-open workspace behavior
Final answer view in main timeline
Tool spacing in final card

Documentation, Configurations and Resources#

Substantive Gate Config: New require_substantiveness YAML parameter (mandatory in checklist)
Novelty Injection Config: New novelty_injection parameter in coordination section (none/gentle/moderate/aggressive)

Technical Details#

Major Focus: Final answer modal redesign, substantive gate, novelty injection, agent identity versioning
PRs Merged: #901 (Final answer improvements)
Contributors: @ncrispino and the MassGen team

[0.1.51] - 2026-02-13#

Added#

Change Documents (Changedoc) (#896): Decision journals agents write in tasks/changedoc.md during coordination, capturing decision provenance, rationale, and code traceability
- Observation context: changedocs passed to other agents in <changedoc> tags for shared decision awareness
- Config: enable_changedoc: true (default on)
Changedoc-Anchored Evaluation Checklist (#896): 5 changedoc-specific checklist items for structured quality evaluation
- Decision Completeness, Rationale Quality, Traceability, Output Quality, Novel Elements
Checklist Gap Report (#896): Mandatory structured gap analysis before verdict
- Config: checklist_require_gap_report: true (default on)
Drift Conflict Policy: Configurable handling of target-file drift when applying isolated changes
- drift_conflict_policy: skip|prefer_presenter|fail
Scratch Directory in Worktrees: .massgen_scratch/ for agent temporary files, git-excluded
CLI --cwd-context Flag: Inject CWD into context paths — ro/read for read-only, rw/write for write access
- Equivalent to Ctrl+P in TUI
Final Presentation Matrix: Deterministic decision matrix for final presentation path selection

Changed#

Review Modal Improvements: Multi-context, multi-file diff visualization with critique capabilities
Mode Bar Responsive Labels: Compact labels adapting to terminal width

Fixed#

Final presentation fallback for empty presentations
Task execution timing fixes

Documentation, Configurations and Resources#

Changedoc System Prompt Sections: New <changedoc> observation context blocks in agent system prompts
Checklist Gap Report Config: New checklist_require_gap_report YAML parameter (default: true)
Drift Conflict Policy Config: New drift_conflict_policy YAML parameter (skip/prefer_presenter/fail)
Scratch Directory Convention: .massgen_scratch/ added to .gitignore in worktrees

Technical Details#

Major Focus: Change documents for multi-agent coordination traceability, changedoc-anchored evaluation checklists
PRs Merged: #896 (Changedoc system), even_execute_time branch
Contributors: @ncrispino and the MassGen team

[0.1.50] - 2026-02-11#

Added#

Chunked Plan Execution (#877): Plans now divided into chunks (e.g., C01_foundation) and executed one chunk at a time with progress checkpoints
- Chunk browsing in TUI with chunk-level progress tracking
- Frozen plan snapshots preserve original plan state during execution
- target_steps and target_chunks parameters for plan sizing
- Dynamic mode for adaptive plan depth controls
Iterative Planning Review Modal (#877): New modal with Continue Planning / Quick Edit / Finalize Plan options
- Allows plan iteration before execution begins
- Quick edit for inline plan adjustments
Skill Lifecycle Management (#878): New lifecycle modes (create_or_update, create_new, consolidate) for evolving skills
- Skill organizer for merging overlapping skills into consolidated workflows
- SKILL_REGISTRY.md routing guide for skill discovery and selection
- Lifecycle mode selection during skill creation
Previous-Session Skills (#878): Load evolving skills from past run logs with load_previous_session_skills config
- Automatic skill discovery from previous session log directories
Local Skills MCP (#878): New MCP tool for skill list/read access in Docker/local execution contexts
- Enables skill access without filesystem tools

Changed#

Worktree Improvements (#877): Branch accumulation across rounds, cross-agent diff visibility via generate_branch_summaries(), orphan cleanup
- Branches accumulate across coordination rounds instead of being recreated
- Other agents can see diffs from worktree branches via branch summaries
Responsive TUI Mode Bar (#877): Vertical/horizontal adaptive layout with compact labels on narrow terminals
TUI Homescreen & Theming (#877): Improved welcome screen layout, CSS refinements, palette updates for light/dark themes
Skills Modal (#878): Source grouping (builtin/project/user/previous_session), quick actions (Enable All/Disable All)
Plan Depth Controls (#877): Dynamic mode, target_steps/target_chunks parameters for plan sizing

Fixed#

Test Fixes (#877): Fixed hooks, Docker mounts, and snapshot tests across the test suite

Technical Details#

Major Focus: Chunked plan execution for safer long-form task completion, skill lifecycle management with consolidation
PRs Merged: #877 (Chunk planning mode), #878 (Improve skill handling)
Contributors: @ncrispino and the MassGen team

[0.1.49] - 2026-02-09#

Added#

Log Analysis Mode in TUI (#869): New “Analyzing” state in the TUI mode bar for in-app run analysis
- Mode bar cycle: Normal → Planning → Executing → Analyzing
- Browse and select log directories and turns directly in the TUI
- Configurable analysis profiles for different analysis depths
- Empty submit in analysis mode runs default analysis on selected target
Fairness Gate for Coordination (#869): Prevents fast agents from dominating coordination rounds
- Configurable fairness_lead_cap_answers to limit how far ahead one agent can get
- max_midstream_injections_per_round to control injection frequency
- Ensures balanced participation across agents of different speeds
Checklist Voting Tool (#869): New checklist_tools_server.py MCP server for structured quality evaluation
- Binary pass/fail scoring for objective quality assessment
- Structured checklist-based evaluation replacing subjective voting
Automated Testing Infrastructure (#869): CI/CD workflow (tests.yml), SVG snapshot baselines, testing strategy spec, 16+ new test files
- GitHub Actions CI pipeline for automated test execution
- SVG snapshot baseline testing for TUI visual regression
- Comprehensive testing strategy specification
Skills Modal in TUI (#869): New modal for discovering and toggling skills in interactive mode
- skills_modals.py for skill discovery and management in TUI
Docker Overlay Images (#869): Dockerfile.overlay and build script for Agent Browser and OpenSkills integration

Changed#

Persona Easing in TUI Mode Bar (#869): Persona easing toggle now accessible from the TUI mode bar
Improved Decomposition Prompts (#869): Better hook injection for non-hook backends
Enhanced System Prompt Sections (#869): Project instructions discovery and checklist evaluation blocks
Expanded Skills Installer (#869): Playwright, Agent Browser, and OpenSkills support
Native Codex & Claude Code Skills (#869): Direct skill integration for both backends

Fixed#

Shadow Agent Chunk Type Comparison (#861): Fixed “[No response generated]” errors caused by incorrect chunk type comparison
Round Banner Timing (#869): Round banner no longer appears before final answer is locked
Hook Injection for Non-Hook Backends (#869): Corrected decomposition prompt injection for backends without native hook support
Final Answer Lock Responsiveness (#869): Improved lock timing and reduced hover lag
Multiple Test Failures (#869): Fixed hooks, persona easing, Docker mounts, and snapshot tests

Documentation, Configurations and Resources#

Testing Strategy: New docs/modules/testing.md with testing architecture and CI gates
SVG Snapshots: Baseline snapshots in massgen/tests/snapshot_tests/
CI/CD Pipeline: .github/workflows/tests.yml for automated testing

Technical Details#

Major Focus: Coordination quality improvements (log analysis TUI, fairness gate, checklist voting), automated testing infrastructure
PRs Merged: #869 (Automate testing), #861 (Shadow agent fix)
Files Modified:
- New: massgen/mcp_tools/servers/checklist_tools_server.py, massgen/frontend/displays/textual/widgets/modals/skills_modals.py
- Modified: massgen/orchestrator.py (fairness gate), massgen/persona_generator.py (easing), massgen/frontend/displays/textual_widgets/mode_bar.py (analysis mode)
- Infrastructure: .github/workflows/tests.yml, Dockerfile.overlay, massgen/tests/ (16+ new test files)
Contributors: @ncrispino, @MuL1ian, and the MassGen team

[0.1.48] - 2026-02-06#

Added#

Decomposition Coordination Mode (#858): New coordination mode that decomposes tasks into subtasks assigned to individual agents
- Task decomposer with presenter agent role for final synthesis
- TUI mode bar toggle, subtask assignment display, and generation modals
- Quickstart wizard integration for decomposition mode selection
Worktree Isolation (#857): Git worktree-based isolation for agent file writes with review workflow
- New write_mode config parameter (auto/worktree/isolated/legacy)
- IsolationContextManager for per-round worktree creation with .massgen_scratch/ directories
- ChangeApplier and review modal for approving/rejecting changes before applying to original paths
- WorktreeManager and ShadowRepo infrastructure for git and non-git directories
- Deprecation of use_two_tier_workspace in favor of write_mode
Stop Tool (#858): New tool enabling agents to signal completion and exit workflows
Global Answer Limits (#858): Orchestrator-level max_answers config alongside existing per-agent controls

Changed#

Quickstart Wizard Docker Setup (#857): Docker setup step integrated into quickstart wizard when Docker mode is selected, with animated pull progress and real-time stdout streaming
Codex Backend (#858): Default model updated from gpt-5.2-codex to gpt-5.3-codex

Fixed#

Light Theme Visibility (#857): Fixed invisible mode bar underlines, separator lines, and toast notifications in light theme with new semantic CSS variables
Subagent Timeout (#857): Added timeout exemption for subagent-related MCP tools (spawn_subagents, get_subagent_status, cancel_subagents) that manage their own timeouts
Post-evaluation Restarts (#857): Disabled max_orchestration_restarts in quickstart defaults to prevent TUI crash on restart

Documentation, Configurations and Resources#

Agent Workspaces Guide: New docs/source/user_guide/agent_workspaces.rst for worktree isolation workflow
Worktrees Module: New docs/modules/worktrees.md with integration examples
Decomposition Configuration: Updated docs/source/reference/yaml_schema.rst, configuration.rst, and running-massgen.rst with decomposition mode examples
Backends Guide: Updated docs/source/user_guide/backends.rst with Codex model update
Capabilities Registry: Updated massgen/backend/capabilities.py with gpt-5.3-codex

Technical Details#

Major Focus: Decomposition coordination mode, worktree isolation for file writes, quickstart improvements
Files Modified:
- Orchestrator: massgen/orchestrator.py (decomposition + worktree isolation logic)
- New: massgen/task_decomposer.py, massgen/infrastructure/worktree_manager.py, massgen/infrastructure/shadow_repo.py
- New: massgen/filesystem_manager/_isolation_context_manager.py, massgen/filesystem_manager/_change_applier.py
- New: massgen/frontend/displays/textual/widgets/modals/review_modal.py, massgen/frontend/displays/textual/widgets/modals/input_modals.py
- TUI: Mode bar decomposition toggle, subagent decomposition display, quickstart wizard Docker step
- Docs: docs/source/user_guide/agent_workspaces.rst, docs/modules/worktrees.md
Dependencies: Added gitpython
Contributors: @ncrispino and the MassGen team

[0.1.47] - 2026-02-04#

Added#

Codex Backend (#843): New codex backend type for OpenAI Codex CLI
- Local and Docker execution modes with workspace mounting
- OAuth and API key authentication
- NativeToolMixin abstract mixin for shared native tool handling between Codex and Claude Code
- Custom and workflow MCP servers (custom_tools_server.py, workflow_tools_server.py) for exposing MassGen tools to CLI-based backends

Changed#

TUI Theme System (#842): Refactored to palette-based architecture with unified base.tcss replacing per-widget inline CSS
- Semantic CSS variables for consistent cross-component theming
- Theme palette files for dark and light variants
- Removed legacy transparent.tcss
Per-agent Voting Sensitivity (#842): Voting sensitivity (strict/balanced/lenient) now configurable per-agent, overriding orchestrator-level defaults with rewritten evaluation criteria
Claude Code Backend (#843): Refactored to use NativeToolMixin with native filesystem support and OS-level sandbox, extracting shared tool handling logic
Round Display Tracking (#842): Vote and answer submissions now track and display submission round numbers in TUI timeline and coordination UI
Gemini Backend (#842): Globally unique tool call ID generation and configuration improvements

Fixed#

Final Presentation Display (#842): Fixed rendering issues with final presentation box
MCP Tool Call Error Handling (#842): Enhanced error handling for invalid MCP tool calls with clearer user guidance

Documentation, Configurations and Resources#

Backends User Guide: Updated docs/source/user_guide/backends.rst with Codex backend documentation
Interactive Mode Design: New docs/modules/interactive_mode.md architecture document
Capabilities Registry: Updated massgen/backend/capabilities.py with Codex models (gpt-5.2-codex, gpt-5.1-codex, gpt-5-codex, gpt-4.1)
Backend Integrator Skill: New massgen/skills/backend-integrator/SKILL.md for guided backend integration workflows
OpenSpec Documents: Interactive mode proposal, design, vision, and spec documents

Technical Details#

Major Focus: Codex backend integration, TUI theme refactoring, per-agent voting sensitivity
Files Modified:
- Backend: massgen/backend/codex.py (new), massgen/backend/native_tool_mixin.py (new), massgen/backend/claude_code.py (refactored)
- TUI: massgen/frontend/displays/textual_themes/base.tcss (new), palette files (new/moved), widget CSS extraction
- MCP: massgen/mcp_tools/custom_tools_server.py (new), massgen/mcp_tools/workflow_tools_server.py (new)
- Docs: docs/source/user_guide/backends.rst, docs/modules/interactive_mode.md
Contributors: @ncrispino and the MassGen team

[0.1.46] - 2026-02-02#

Added#

Subagent TUI Streaming (#821): Stream and display subagents almost identically to main process in TUI
- Clickable subagent preview cards that expand to full timeline views
- Real-time event streaming from subprocess logs via symlinks
- Unified display components reused for both main agents and subagents
- Subagent rounds tracking and status visualization
Enhanced Final Presentation Display:
- Final presentation now includes workspace visualization
- Winning agent highlighted with clear visual indicator
- Workspace symlinks (curr_workspace) for easy access to final agent’s workspace
- Improved final answer formatting with better separation from reasoning

Changed#

TUI Event Architecture Refactor: Major refactor to structured event emission pipeline
- Single source of truth for TUI display creation shared between main and subagent views
- Unified event parsing for consistent tool displays across agent types
- Stream chunk handling removed in favor of direct event emission (phase 4 refactor)
- Improved event streaming architecture for better maintainability
Subagent Display Improvements:
- Refactored subagent rendering to remove older streams and prevent clutter
- Better debugging support with enhanced logging
- Tool numbering fixes for consistent display

Fixed#

Banner Display Issues: Fixed banners not showing up for first coordination round
Tool Call ID Handling: Fixed issue when tool call IDs are not alphanumeric (e.g., kimi2.5 models)
Round Tracking: Improved round tracking logic for more accurate status display

Documentation, Configurations and Resources#

Tutorial Video GIFs: New docs/source/_static/images/tutorial-*.gif files for visual documentation
Module Documentation: New docs/modules/subagents.md comprehensive guide for subagent architecture
Updated Documentation: docs/source/index.rst with tutorial GIF previews and updated video links
OpenSpec Design Docs: Multiple design documents for TUI refactoring and event pipeline architecture

Technical Details#

Major Focus: Subagent TUI streaming, event architecture refactor, final presentation improvements
Files Modified:
- TUI: massgen/frontend/displays/textual_widgets/subagent_screen.py, subagent_card.py, event handling modules
- Subagent: massgen/subagent/manager.py with improved logging directory structure
- Final presentation: Enhanced workspace handling and visual indicators
- Docs: docs/modules/subagents.md, docs/source/index.rst
Contributors: @ncrispino (23 commits), @HenryQi, @franklinnwren, and the MassGen team

[0.1.45] - 2026-01-31#

Changed#

BREAKING (Soft): Default display changed from rich_terminal to textual_terminal
- All users now get the superior TUI experience by default
- Existing configs with display_type: "rich_terminal" will show deprecation warning and use TUI
- Use --display rich flag to force legacy Rich display
- Updated ALL 160+ example configs to use textual_terminal

Improved#

Setup Wizard: --setup and --quickstart now generate configs with TUI display by default
Documentation: Enhanced with prominent TUI feature descriptions and benefits
First-Run Experience: Clear explanation of TUI benefits for new users

Deprecated#

Rich Terminal Display: rich_terminal display type is now deprecated in favor of textual_terminal
- Configs using rich_terminal will show warning and auto-convert to TUI
- Use --display rich to explicitly request legacy Rich display

Fixed#

Documentation Paths: Fixed case study page paths for proper rendering
PyPI Packaging: Added missing files to MANIFEST.in for complete package distribution
ReadTheDocs Config: Updated Python version to 3.12 for documentation builds

Documentation, Configurations and Resources#

Updated Documentation: docs/quickstart/installation.rst and docs/quickstart/running-massgen.rst with TUI as default
Config Migration: Example configs in massgen/configs/ updated to use textual_terminal
ReadTheDocs: Updated .readthedocs.yaml with Python 3.12

Technical Details#

Major Focus: TUI default transition, config migration, documentation improvements
Files Modified:
- Configs: All YAML files in massgen/configs/
- Docs: docs/source/quickstart/*.rst, .readthedocs.yaml
- Packaging: MANIFEST.in, pyproject.toml
Contributors: @ncrispino, @HenryQi, and the MassGen team

[0.1.44] - 2026-01-28#

Added#

Execute Mode: Independent mode for browsing and executing existing plans (#819)
- Cycle through modes: Normal → Planning → Execute via Shift+Tab or mode bar click
- Plan selector popover shows up to 10 recent plans with timestamps and prompts
- “View Full Plan” button opens modal with all plan tasks
- Empty submission (just pressing Enter) executes selected plan
- Context paths preserved from planning phase to execution phase
- Warning shown if no plans exist when trying to enter Execute mode
Case Studies Setup Guide: Interactive setup instructions on case studies page (#818)
- “Try it yourself” collapsible sections with setup guide
- Quick start command: uv run massgen --web
- Model selection guidance (Claude 4.5 Opus, Gemini 3 Pro, GPT 5.2)
- Terminal config file example for CLI users
- Helper text prompting users to compare MassGen with single-agent baselines

Fixed#

Plan Mode Separation: Fixed bug where planning instructions were injected during execute mode
- Planning prompt prepending now only occurs for plan_mode == "plan"
- Execute mode uses build_execution_prompt() without planning overhead
Tool Call Spacing: Fixed spacing issues in tool card display
Timeline Performance: Improved scrolling performance with viewport optimization and reduced timeline size limits

Changed#

Context Paths Storage: PlanMetadata now includes context_paths field in massgen/plan_storage.py
- Context paths stored during finalize_planning_phase
- Restored automatically in prepare_plan_execution_config during execution
- Enables consistent file/directory access between planning and execution
Empty Submission Support: Input widget now allows empty submission in execute mode
- Placeholder text: “Press Enter to execute selected plan - or type instructions”
- Removed input text guard to enable plan execution without additional input
Plan Options Widget: Enhanced PlanOptionsPopover with “View Full Plan” functionality
- New ViewPlanRequested message for modal communication
- Better plan browsing experience

Documentation, Configurations and Resources#

Case Studies Enhancement: docs/source/case_studies/index.html with setup guide
- New docs/source/case_studies/terminal_config.txt with example YAML configuration
- Video tutorial links moved higher for better discoverability
- Added contextual notes for baseline comparisons
Shortcuts Documentation: Updated shortcuts_modal.py with Shift+Tab mode cycling description

Technical Details#

Major Focus: Execute mode for independent plan selection, TUI performance improvements, case studies UX
Files Modified:
- TUI: textual_terminal_display.py, mode_bar.py, plan_options.py, multi_line_input.py, content_sections.py
- Plan system: plan_storage.py, plan_execution.py, tui_modes.py
- Backend: claude_code.py (tool tracking improvements)
- Docs: index.rst, case_studies/index.html
Contributors: @ncrispino and the MassGen team

[0.1.43] - 2026-01-26#

Added#

Tool Call Batching: Consecutive MCP tool calls are now grouped into collapsible tree views (#815)
- Shows 3 items by default, collapses rest with “+N more” indicator
- Click to expand full list
- Respects Timeline Chronology Rule: tools only batch when consecutive (no intervening content)
- New ToolBatchCard widget and ToolBatchTracker state machine
Interactive Case Studies: New documentation page with visual comparisons (#812)
- Side-by-side SVG comparisons between MassGen and single-agent solutions
- Iterative refinement examples showing multi-round improvements
- Collapsible sections with baseline visualizations
Video Tutorials Section: New documentation with Getting Started and Development videos
- Prominent CTAs linking to YouTube tutorials
- Descriptive text for each video category
Plan Mode Enhancements: New PlanOptionsPopover widget for plan management
- Browse recent plans with quick access
- Plan depth selector (thorough/balanced/quick)
- Broadcast mode toggle (human/agents/none)
- Plan validation before execution
Quoted Path Support: Paths with spaces now work correctly using quotes
- @"/path/with spaces/file.txt" syntax for context injection
- Tab completion support for quoted paths
- Write permission suffix works with quotes: @"/path/file.txt":w

Fixed#

Final Presentation Display: Fixed critical bug where final answers weren’t displayed properly
- Reasoning text now separated from actual answer content
- Visual distinction: reasoning collapsed/smaller, answer prominent
- Fixed content filtering in ContentNormalizer.should_display logic
Bottom Status Bar: Fixed status bar not showing in certain scenarios
Scrolling Bar: Fixed scrolling bar on right side display issues
Mode Buttons: Fixed mode button interaction and alignment
Task Highlighting: Fixed task highlighting in task plan cards
Toast Location: Fixed toast notification positioning

Changed#

Reasoning/Content Display: Enhanced formatting with vertical line indicators for thinking blocks
Tool Presentation: Improved tool card visual presentation
Demo GIF: Updated docs/source/_static/images/readme.gif with higher resolution

Documentation, Configurations and Resources#

Interactive Case Studies: New docs/source/case_studies/index.html with SVG comparisons
- Example SVGs for Claude, GPT, Gemini, and MassGen outputs
- docs/source/case_studies/example_svgs/ directory with visualization assets
Homepage Updates: Updated docs/source/index.rst with case studies CTA and video tutorials section
OpenSpec Proposals: Multiple TUI improvement specifications in openspec/changes/:
- add-tui-tool-call-batching/ - Tool batching design and implementation
- improve-tui-final-presentation-display/ - Final presentation fix specs
- fix-tui-mode-bar-alignment/ - Mode bar alignment fix
- fix-tui-tool-card-spacing/ - Tool card spacing improvements
- add-tui-workflow-comprehension/ - Workflow comprehension enhancements

Technical Details#

Major Focus: TUI UX polish, tool call batching, documentation enhancements
Contributors: @ncrispino (22 commits), @franklinnwren (8 commits), @HenryQi (3 commits) and the MassGen team

[0.1.42] - 2026-01-23#

Added#

TUI Visual Redesign: Comprehensive visual overhaul with modern “Conversational AI” aesthetic (#806)
- Phase 1: Unified input card with integrated mode toggles, rounded corners (╭╮╰╯), simplified radio-style indicators
- Phase 2: Agent tabs redesign with dot indicators (◉ active, ○ waiting, ✓ done), two-line display (name + model)
- Phase 3: Tool cards with adaptive density - collapsed by default, click to expand parameters/results
- Phase 4: Welcome screen improvements with centered input and muted help hints
- Phase 5: Task lists with visual progress bars, “X of Y” counts, and “← current” markers
- Phase 6: Modal polish with rounded containers, consistent headers, softer borders, unified button styling
- Phase 7: Header polish with bullet separators, desaturated color palette, warmer tones
- Phase 8: Professional visual polish throughout
- Phase 9: Edge-to-edge borderless container layout
- Phase 11: UX polish with collapsible reasoning blocks, scroll indicators
- Phase 12: CSS-based round navigation (partial)
- Phase 13: Backend integration with token usage updates for TUI status ribbon
Human Input Queue: Inject messages to agents mid-stream during execution
- HumanInputHook for queuing and injecting human input during agent execution
- Thread-safe queue with per-agent tracking (each message delivered once per agent)
- Callback support for TUI visual indicator updates
- Messages persist until turn ends, allowing injection to multiple agents

Fixed#

AG2 Single-Agent Coordination: Fixed coordination issues for single-agent AG2 setups (#804)
- Single agent can now vote for itself after producing its first answer
- Properly clears restart_pending flag for single-agent scenarios
- Fixes stuck coordination when using AG2 adapter with single agent
Plan Execution in TUI: Fixed plan-then-execute workflow in Textual TUI
Planning Prompt Improvements: Better subagent clarity and planning guidance

Changed#

Token Usage Updates: Orchestrator now emits token_usage_update stream chunks for real-time TUI status updates
Plan Session ID: Orchestrator accepts optional plan_session_id to prevent workspace contamination during plan execution

Documentation, Configurations and Resources#

TUI Redesign Handoffs: Design handoff documents for implementation phases
- New docs/dev_notes/tui_redesign_phase6_handoff.md for modal improvements
- New docs/dev_notes/tui_redesign_phase9_11_13_handoff.md for layout and UX polish
OpenSpec Proposals: Complete TUI redesign specification in openspec/changes/update-tui-conversational-design/
- proposal.md - Full 13-phase redesign proposal
- design.md - Visual design decisions and rationale
- specs/tui/spec.md - Detailed component specifications
- tasks.md - Implementation task breakdown
- HANDOFF_PHASE12.md - Phase 12 handoff for CSS round navigation

Technical Details#

Major Focus: TUI visual redesign, human input injection, AG2 single-agent fixes
Contributors: @ncrispino, @HenryQi, @db-ol and the MassGen team

[0.1.41] - 2026-01-21#

Added#

Async Subagent Execution: Background subagent execution with async_=True parameter (MAS-214)
- Parent agents continue working while subagents run in background
- Non-blocking spawn_subagents returns immediately with running status
- Parent can poll for subagent completion and retrieve results
- Configurable injection strategies: tool_result (default) or user_message
- Batch injection when multiple subagents complete simultaneously
Result Polling: Check subagent completion status and retrieve results
- Poll for completed background subagents when ready
- Results returned in structured XML format with metadata
- Includes execution time, token usage, and workspace paths
Subagent Round Timeouts: Per-round timeout control for subagents
- New subagent_round_timeouts configuration section
- Supports initial_round_timeout_seconds, subsequent_round_timeout_seconds, round_timeout_grace_seconds
- Inherits from parent timeout_settings if omitted

Configuration#

New Subagent Parameters: Extended YAML configuration options
- enable_subagents: Enable subagent tools for parallel task execution
- subagent_default_timeout: Default timeout in seconds (default: 300)
- subagent_min_timeout: Minimum allowed timeout (default: 60)
- subagent_max_timeout: Maximum allowed timeout (default: 600)
- subagent_max_concurrent: Maximum concurrent subagents (default: 3)
- subagent_round_timeouts: Per-round timeout settings for subagents
- async_subagents: Async execution settings (enabled, injection_strategy)

Documentation, Configurations and Resources#

Subagents Guide: Updated docs/source/user_guide/advanced/subagents.rst with async execution section
Async Example Config: New massgen/configs/features/async_subagent_example.yaml
OpenSpec Proposals: Design documents in openspec/changes/add-async-subagent-execution/
- proposal.md - Feature proposal and impact analysis
- design.md - Architecture decisions and implementation details
- specs/subagent/spec.md - Detailed specification

Technical Details#

Major Focus: Async subagent execution, subagent round timeouts, subagent configuration parameters
Contributors: @ncrispino, @HenryQi and the MassGen team

[0.1.40] - 2026-01-19#

Added#

Textual TUI Interactive Mode: Interactive terminal UI with --display textual for interactive MassGen sessions
- Real-time agent output streaming with syntax highlighting
- Agent tab bar for switching between agents and post-evaluation views
- Keyboard-driven navigation with extensive keyboard shortcuts
- Keyboard navigation with j/k scrolling and :q to quit
- Comprehensive modals:
  - ? or h: Keyboard shortcuts help
  - f: Full agent output
  - c: Cost breakdown (token usage and costs)
  - m: Tool metrics
  - v: Vote results
  - o: Orchestrator events
  - s: System status
  - p: MCP server status
  - b: Answer browser with side-by-side comparisons
  - t: Coordination timeline
  - w: Workspace file browser with tree navigation and file preview
- Context path injection UI with @ syntax support
- Human feedback integration with prompt modal
- Enhanced final answer presentation with formatting
- Plan execution mode selection UI
- Scrolling improvements with visual indicators
- Tool input/output display with color-coded formatting

Changed#

Final Answer View: Improved presentation and formatting in Textual TUI
Subagent Display: Fixed subagent rendering and progress bar updates
Context Path Handling: Enhanced context path validation and display
Broadcasting: Improved broadcasting behavior for questions similar to context injection

Fixed#

Tool Inputs Not Showing: Fixed issue where tool inputs were not displayed in later answers
Empty Space Issue: Resolved empty space rendering problem in agent answers
Scrolling: Fixed scrolling behavior and visual indicators
Cancellation: Improved Ctrl+C handling and graceful shutdown
Menu Display: Fixed issue with too many items being displayed in menus
Click Handling: Resolved click event issues in TUI
Path Permissions: Fixed workspace path permission handling
Task Plan Display: Fixed task plan rendering in TUI

Documentation, Configurations and Resources#

Textual TUI Architecture: New docs/dev_notes/textual_tui_architecture.md for TUI implementation details
Textual UI Developer Skill: New massgen/skills/textual-ui-developer/SKILL.md for TUI development workflows
OpenSpec Proposals: Multiple design documents in openspec/changes/:
- add-tui-modes/ - TUI modes design and specs
- tui-production-upgrade/ - Enhanced TUI widgets
- update-textual-tui-polish/ - TUI polish and refinements
Updated CLAUDE.md: Enhanced project instructions with TUI development guidance
Updated Config: Modified massgen/configs/basic/multi/three_agents_default.yaml for TUI testing

Technical Details#

Major Focus: Textual TUI interactive mode, keyboard navigation, workspace browser, performance optimization
Contributors: @ncrispino, @praneeth999, @HenryQi and the MassGen team

[0.1.39] - 2026-01-16#

Added#

Plan and Execute Workflow: Complete plan-then-execute workflow separating “what to build” from “how to build it”
- --plan-and-execute: Create plan then immediately execute it
- --execute-plan <id|path|latest>: Execute an existing plan without re-planning
- --broadcast <human|agents|false>: Control planning collaboration (auto-switches to false in automation mode)
Task Verification Workflow: New verified status for distinguishing implementation from validation
- Status flow: pending → in_progress → completed → verified
- verification_group labels for batch verification (e.g., “foundation”, “frontend_ui”)
- get_tasks_awaiting_verification() and get_verification_group_status() helpers
- Agents verify entire groups at logical checkpoints
Plan Storage System: Persistent plan management in .massgen/plans/
- Plan structure: plan_metadata.json, execution_log.jsonl, plan_diff.json
- frozen/ directory for immutable planning-phase snapshots
- workspace/ directory for modified plan after execution
- Plan IDs use timestamp format: YYYYMMDD_HHMMSS_microseconds

Changed#

Planning Prompt Improvements: Updated guidance to focus on outcomes over implementation
- “Describe WHAT the final product needs, not HOW to build it”
- Verification methods must be automated (not manual inspection)
- Quality focus: “If it’s visual, it should LOOK good”

Fixed#

Response API Function Call Messages: Sanitized function_call messages for OpenAI Response API compatibility (#792)
- Filter function_call messages to only include valid fields (type, name, arguments, call_id, id)
- Remove invalid fields like ‘content’ that cause Unknown parameter errors
- Ensure ‘arguments’ field is JSON-serialized string, not an object
- Fixes: Unknown parameter: 'input[N].content' and Invalid type for 'input[N].arguments'
Plan Execution Edge Cases: Various fixes for plan execution workflow
- Single-agent config handling for both agent: and agents: shapes
- Plan collection path fixed to look for tasks/plan.json (file) not plan/ (directory)
- Subprocess deadlock prevention by merging stderr into stdout
- Argparse handling for questions starting with - via -- end-of-options marker
- Progress calculation now counts verified tasks as completed

Documentations, Configurations and Resources#

Planning Mode Guide: Updated docs/source/user_guide/advanced/planning_mode.rst with plan-and-execute workflow
Roadmap: New ROADMAP_v0.1.40.md for next release planning

Technical Details#

Major Focus: Plan-and-execute workflow, task verification, plan storage system
Contributors: @ncrispino, @HenryQi, @db-ol and the MassGen team

[0.1.38] - 2026-01-15#

Added#

Task Planning Mode: Create structured plans for future workflows with --plan flag (plan-only, no auto-execution)
- --plan: Enable task planning mode for structured work breakdown
- --plan-depth: Control planning granularity (shallow/medium/deep)
- Planning prompt prefix for configurable depth
- Outputs feature_list.json with task dependencies and priorities
Two-Tier Workspace: Git-backed scratch/deliverable separation
- use_two_tier_workspace: true config option
- scratch/ directory for work-in-progress
- deliverable/ directory for complete, self-contained outputs
- Automatic [INIT], [SNAPSHOT], [TASK] git commits
- Task completion triggers git commit with completion notes
- Agents can use git log to review work history
Project Instructions Auto-Discovery: CLAUDE.md/AGENTS.md support following agents.md standard
- Automatic discovery from context paths (via @path syntax)
- Hierarchical “closest wins” algorithm for monorepo support
- CLAUDE.md takes precedence over AGENTS.md at same level
- Contents injected into system prompts with softer framing
Batch Image Analysis: Multi-image support in media tools
- understand_image accepts images dict for named multi-image comparison
- read_media accepts inputs list for batch image processing
- Dict keys become reference names in prompts for image identification
- max_concurrent parameter for concurrency control
Docker Health Monitoring: Container diagnostics on MCP failures
- get_container_health() for health status checking
- get_container_logs() and save_container_logs() for log retrieval
- Automatic log capture when MCP disconnections occur
- Health info tracked in enforcement events
Enhanced Enforcement Tracking: Improved status.json visibility
- finish_reason: "timeout", "completed", "error", or "in_progress"
- finish_reason_details: Human-readable explanation
- is_complete: Boolean completion status
- Fields appear at top of status.json for immediate visibility

Changed#

Improved Deliverable Guidance: System prompts emphasize self-contained packages
- Checklist: all required files, dependencies, assets, README
- Explicit examples for different artifact types
- Soft timeout message reinforces complete deliverables
Git History in System Prompt: Agents aware of version control
- Commit prefix documentation: [INIT], [SNAPSHOT], [TASK]
- Guidance to use git log for reviewing work history

Fixed#

Vote Tracking Bug: Ignored votes no longer leak into final results
- Clear agent_states[agent_id].votes when vote ignored due to restart
- Sync between agent_states and coordination_tracker.votes
Soft→Hard Timeout Race Condition: Guaranteed progression
- Hard timeout now calculated from soft timeout injection time
- Soft timeout must fire before hard timeout can trigger
- RoundTimeoutState class for shared state between hooks
MCP Reset on Restart: Full tools restored after hard timeout restart
- Reset _mcp_initialized = False in handle_restart()
- Forces MCP re-initialization (17 tools vs 2)
Circuit Breaker for Hard Timeout: Prevents infinite denial loops
- Tracks consecutive denied tool calls
- Warning after 3+ consecutive denials
- Force terminate after 10 blocked tool calls
use_two_tier_workspace Config Pass-Through: Flag now reaches orchestrator
- Added to CoordinationConfig creation in cli.py
- Planning MCP server receives --use-two-tier-workspace flag

Documentations, Configurations and Resources#

Project Integration Guide: New docs/source/user_guide/files/project_integration.rst
Debugging Assumptions: Added guidance to CLAUDE.md for log analysis
OpenSpec Proposals: New openspec/changes/add-enforcement-observability/ and openspec/changes/add-task-planning-mode/
Skills: New massgen/skills/massgen-log-analyzer/SKILL.md
Roadmap: Renamed ROADMAP_v0.1.38.md to ROADMAP_v0.1.39.md

Technical Details#

Major Focus: Task planning, two-tier workspaces, project instructions, timeout reliability
Contributors: @ncrispino, @chiwang, @HenryQi and the MassGen team

[0.1.37] - 2026-01-12#

Added#

Execution Traces: Full execution history preserved as searchable markdown files (MAS-226)
- Trace file format: Human-readable execution_trace.md saved alongside snapshots
- Compression recovery: Agents can read trace files to recover detailed history after context compression
- Cross-agent access: Other agents can access execution traces in temp workspaces to understand approaches
- Full content preservation: Tool calls, results, and reasoning blocks saved without truncation
- Grep-friendly: Searchable format for debugging and analysis
Claude Code Thinking Mode: Streaming buffer support for Claude Code reasoning
- Thinking content captured in streaming buffer for trace files
- Integration with execution trace system
Voting Execution Traces: Vote reasoning captured in execution trace files
- Full vote context preserved for analysis

Changed#

Standardized Agent Labeling: Consistent agent identification across backends
- Unified labeling format for multi-agent coordination
- Improved workspace anonymization for cross-agent sharing
Gemini Thinking Mode: Fixed thinking/reasoning content handling
- Proper streaming buffer integration for Gemini reasoning blocks
Streaming Buffer Improvements: Enhanced reasoning content capture
- Better handling of thinking blocks across providers
- Improved trace file generation

Fixed#

Claude Code Backend: Fixed skills and tool handling issues
Config Builder: Fixed configuration generation edge cases
Round Timeout Handling: Improved timeout behavior during coordination

Documentations, Configurations and Resources#

Timeouts Guide: Updated docs/source/reference/timeouts.rst with comprehensive timeout documentation
Backends Guide: Updated docs/source/user_guide/backends.rst with OpenRouter support
Logging Guide: Updated docs/source/user_guide/logging.rst with execution trace information
Debug Config: New massgen/configs/debug/round_timeout_test.yaml for timeout testing
OpenSpec: New openspec/changes/add-execution-traces/ with proposal and specs

Technical Details#

Major Focus: Execution traces for context recovery, thinking mode improvements, standardized agent labeling
Contributors: @ncrispino, @chiwang, @HenryQi and the MassGen team

[0.1.36] - 2026-01-09#

Added#

Hook Framework: General hook framework for extending agent behavior at key execution points (MAS-215)
- PreToolUse hooks: Execute before tool invocation for permission validation and argument modification
- PostToolUse hooks: Execute after tool results for content injection and processing
- Injection strategies: tool_result (append to output) and user_message (separate message)
- Built-in hooks: MidStreamInjectionHook for cross-agent updates, HighPriorityTaskReminderHook for task completion
- Custom hooks: Python callable hooks with glob-style pattern matching (*, Write|Edit, mcp__*)
- Error handling: Configurable fail-open (default) or fail-closed behavior for security-critical hooks
- Debug support: debug_delay_seconds and debug_delay_after_n_tools for testing mid-stream injection
Unified @path Context Handling: Inline context path references in prompts
- Inline file picker: Type @ in CLI to trigger autocomplete popup (like Claude Code)
- Syntax support: @path (read), @path:w (write), @dir/ (directory)
- Context accumulation: Paths from earlier turns remain accessible in later turns
- Permission upgrade: @file in turn 1, @file:w in turn 2 grants write permission
- Deferred agent creation: Docker containers launch once with all paths from first prompt
Claude Code Native Hooks: Integration with Claude Code’s hook system
- Support for Claude Code temp filesystem tools permission handling

Changed#

Docker Resource Management: Clean up Docker resources when recreating agents for new @path references
- Prevents resource leaks during interactive sessions with path changes
Installation Instructions: Revised README with clearer uv installation steps
- Streamlined quickstart guide for faster onboarding

Fixed#

Path Handling: Fixed path reference handling for Web UI and Rich CLI
- Consistent behavior across CLI interactive mode, automation mode, and Web UI

Documentations, Configurations and Resources#

Hook Framework Guide: New docs/source/user_guide/advanced/hooks.rst with comprehensive hook documentation
File Operations Guide: Updated docs/source/user_guide/files/file_operations.rst with @path syntax
Installation Guide: Updated docs/source/quickstart/installation.rst with uv instructions
Hook Config Example: New massgen/configs/hooks/example_hooks.yaml for hook configuration
Debug Config: New massgen/configs/debug/injection_delay_test.yaml for testing mid-stream injection
OpenSpec: New openspec/changes/add-hook-framework/ and openspec/changes/unify-context-path-handling/ proposals

Technical Details#

Major Focus: Hook framework for agent lifecycle events, unified @path syntax, Claude Code integration
Contributors: @ncrispino, @franklinnwren, @HenryQi and the MassGen team

[0.1.35] - 2026-01-07#

Added#

Log Analysis CLI Command: New massgen logs analyze for AI-assisted log analysis (MAS-227)
- Prompt mode (default): Generates analysis prompt referencing massgen-log-analyzer skill for coding CLIs
- Self-analysis mode (--mode self): Runs 3-agent MassGen team for multi-perspective analysis
- Per-turn analysis reports: Reports placed at turn_N/ANALYSIS_REPORT.md instead of per-attempt
- Supports --turn/-t for specific turn, --force/-f for overwrite, --ui for UI mode selection
- Enhanced massgen logs list with “Analyzed” column and --analyzed/--unanalyzed filters
Logfire Workflow Analysis Attributes: Comprehensive observability for understanding agent behavior (MAS-199)
- Round context: massgen.round.intent, available_answers, answer_previews for workflow explanation
- Vote context: Extended massgen.vote.reason (500 chars), answer_label_mapping for vote analysis
- Agent work products: massgen.agent.files_created, file_count for detecting repeated work
- Restart context: massgen.restart.reason, trigger, triggered_by_agent
- Local file references: massgen.log_path, agent.log_path, answer_path for hybrid access
direct_mcp_servers Config Option: Keep specific MCP servers as direct protocol tools
- When enable_code_based_tools: true, exempts specified servers from code-only filtering
- Useful for debugging/monitoring tools (e.g., Logfire) that need immediate access
- Subagents automatically inherit direct_mcp_servers from parent
- Logs warning if server not found in mcp_servers
Task Context Module: New massgen/context/ package for unified context management
- TaskContext class for managing agent task state and context

Changed#

Skill & Voting Improvements: Enhanced skill execution and voting coordination
- MCPs can now run directly in certain scenarios
- Improved skill parameter handling
Analysis Per-Turn: Log analysis now operates at turn level rather than attempt level
- More intuitive organization of analysis reports

Fixed#

Unknown Tool Handling: Unknown/malformed tool names (e.g., Gemini’s default_api: prefix) no longer cause agent termination (MAS-225)
- Only client-provided external tools trigger external tool call path
- Unknown tools logged and skipped gracefully
Vote-Only Mode: Fixed agents wasting rounds when reaching max_new_answers_per_agent
- System message now correctly omits new_answer tool
- Internal tool filtering uses agent-specific tools
- Prevents hallucinated new_answer calls from passing validation
Grok Backend: Fixed tool handling issues
Gemini Backend: Fixed tool-related problems and parameter handling
Metadata Saving: Config loader now returns raw/unexpanded config to avoid logging secrets

Documentations, Configurations and Resources#

Logging Guide: Updated docs/source/user_guide/logging.rst with CLI quick reference and analysis workflow
Code-Based Tools Guide: New “Direct MCP Servers” section in docs/source/user_guide/tools/code_based_tools.rst
CLI Reference: Updated docs/source/reference/cli.rst with logs analyze command documentation
YAML Schema: Added direct_mcp_servers parameter in docs/source/reference/yaml_schema.rst
Analysis Configs: New massgen/configs/analysis/log_analysis.yaml and log_analysis_cli.yaml
Skill Update: Comprehensive update to massgen/skills/massgen-log-analyzer/SKILL.md
OpenSpec: New openspec/changes/add-logfire-workflow-analysis/ with proposal and specs

Technical Details#

Major Focus: Log analysis CLI, Logfire workflow attributes, direct MCP servers, tool handling fixes
Contributors: @ncrispino, @chiwang, @HenryQi and the MassGen team

[0.1.34] - 2026-01-05#

Added#

OpenAI-Compatible Server: Local HTTP server exposing MassGen as an OpenAI-compatible API
- Run with massgen server or python -m massgen.openai_server
- Compatible with any OpenAI SDK client for easy integration
- Aggregates usage statistics in server responses
- Uses massgen run backend for feature parity with CLI
Dynamic Model Discovery: Authenticated model listing for Groq and Together backends
- Fetches available models via API instead of hardcoded lists
- Supports OpenAI-compatible model discovery endpoints
- Design documentation in docs/dev_notes/discovery/
Review Skill: New skill for code review workflows

Changed#

WebUI Improvements: Enhanced frontend experience
- File diff display for workspace changes
- Answer refresh polling for real-time updates
- Optimized workspace browser timing and performance
- Better caching for office documents and scanning
- Removed unnecessary workspace browser elements
Subagent System Reliability: Improved multi-agent coordination
- Better status tracking and error handling
- Cancellation recovery improvements
- Context and media handling fixes
- Warning improvements for subagent operations
Pre-commit Workflow: Added convenience scripts for pre-commit hooks

Fixed#

OpenAI Server: Fixed null args handling in server responses
WebUI Status Tracking: Fixed “Done” status tracking error
Responses Compression: Fixed compression input issue
Superseded Vote Tracking: Fixed vote tracking for superseded responses
Historical Workspace: Fixed workspace history retrieval problems
Logfire Optional: Made Logfire truly optional in base_with_custom_tool_and_mcp.py
Persona Handling: Use persona JSONs even if generation not finished

Documentations, Configurations and Resources#

HTTP Server Integration Guide: New docs/source/user_guide/integration/http_server.rst for OpenAI-compatible server usage
Model Discovery Design: New docs/dev_notes/backend_model_listing.md design document for backend model listing (MAS-163)
Subagent Documentation: Updated docs/source/user_guide/advanced/subagents.rst with status tracking and recovery details
CLI Reference: Updated docs/source/reference/cli.rst with server command documentation
Skills: New massgen/skills/release-prep/SKILL.md for release automation, new massgen/skills/pr-checks/SKILL.md for code review

Technical Details#

Major Focus: OpenAI-compatible server, dynamic model discovery, WebUI improvements, subagent reliability
Contributors: @ncrispino, @Angela, @maxim-saplin, @chiwang, @randombet, @HenryQi and the MassGen team

[0.1.33] - 2026-01-02#

Added#

Reactive Context Compression: Automatic conversation compression when context length errors are detected
- Summarizes older messages while preserving recent context
- Supports all major backends: OpenAI, Claude, Gemini, OpenRouter, Grok
- Includes message truncation fallback when compression alone is insufficient
Streaming Buffer System: Tracks accumulated streaming content for compression recovery
- Captures text deltas, tool calls, tool results, and reasoning/thinking content
- New --save-streaming-buffers CLI flag to save buffers for debugging
- New persist_conversation_buffers config option for cross-agent buffer inspection

Changed#

File Overwrite Protection: write_file tool now refuses to overwrite existing files (use edit_file instead)
Task Plan Duplicate Protection: create_task_plan MCP tool prevents re-creating plans after recovery, avoiding duplicate work
Grok Backend MCP Tools: Fixed MCP tools visibility by removing incorrect stream method override
Circuit Breaker Debugging: Added agent_id, error_type, and error_message parameters for better failure diagnostics
Voting Prompts: Improved agent coordination prompts to encourage answer synthesis before voting
Subagent Failure Handling: Results now include both workspace and log_path for debugging failed/timed-out subagents

Fixed#

GPT-5 Model Behavior: System prompt adjustments ensure MassGen task planning is used over native model planning
Gemini Vote-Only Mode: Fixed vote_only parameter handling in Gemini backend streaming
Subagent Failed Paths: Fixed subagent MCP server handling of failed subagent results
Incomplete Response Recovery: Added recovery mechanism when API streams end early, preserving partial content

Documentations, Configurations and Resources#

Context Compression Design Doc: New docs/dev_notes/context_compression_design.md with architecture, testing, and backend-specific notes
Test Configurations: New test_reactive_compression.yaml for compression testing

Technical Details#

Major Focus: Reactive context compression, streaming buffer system, MCP tool protections
Contributors: @ncrispino and the MassGen team

[0.1.32] - 2025-12-31#

Changed#

Session Export Multi-Turn Support: Enhanced massgen export command with multi-turn session handling
- New --turns flag for turn range selection (all, N, N-M, latest)
- Workspace options: --no-workspace, --workspace-limit (default 500KB per agent)
- Export controls: --yes (skip prompts), --dry-run, --verbose, --json
- Multi-turn file collection preserves turn/attempt structure in exported gists
Logfire Optional Dependency: Moved Logfire from required to optional [observability] dependency
- Install with pip install massgen[observability] to enable Logfire tracing
- Helpful error message when --logfire flag used without Logfire installed
- Reduces default installation size for users who don’t need observability
Per-Attempt Logging: Each orchestration restart attempt now has isolated log files
- Separate massgen.log and execution_metadata.yaml per attempt directory
- Log handlers reconfigured on restart via set_log_attempt() function
- Viewer adjusted to handle multiple attempt directories
Office Document PDF Conversion: Automatic PDF conversion for DOCX/PPTX/XLSX when sharing sessions
- Uses Docker + LibreOffice for headless conversion
- Includes both original file (for download) and PDF (for preview) in gists
- Tries sudo image first (mcp-runtime-sudo), falls back to standard image

Documentations, Configurations and Resources#

Installation Documentation: Clarified uv run commands for tests and examples in README and quickstart docs
Logfire Documentation: Updated installation instructions for observability optional extra

Technical Details#

Major Focus: Multi-turn session export, Logfire optional dependency, per-attempt logging
Contributors: @ncrispino @AbhimanyuAryan and the MassGen team

[0.1.31] - 2025-12-29#

Added#

Logfire Observability Integration: Comprehensive structured logging and tracing via Logfire
- Automatic LLM instrumentation for OpenAI, Anthropic Claude, and Google Gemini backends
- Tool execution tracing for MCP and custom tools with timing metrics
- Agent coordination observability with per-round spans and token usage logging
- Enable via --logfire CLI flag or MASSGEN_LOGFIRE_ENABLED=true environment variable
- Graceful degradation to loguru when Logfire is disabled
- New massgen-log-analyzer skill for AI-assisted log analysis

Fixed#

Azure OpenAI Native Tool Call Streaming: Tool calls now accumulated and yielded as structured tool_calls chunks instead of plain content
OpenRouter Web Search Logging: Fixed logging output for web search operations

Documentations, Configurations and Resources#

Logfire Documentation: New docs/source/user_guide/logging.rst with usage guide and SQL query examples
Python Installation Guide: Added link to Python installation guide in quickstart docs

Technical Details#

Major Focus: Logfire observability integration, Azure OpenAI tool call streaming
Contributors: @ncrispino @AbhimanyuAryan @shubham2345 @franklinnwren and the MassGen team

[0.1.30] - 2025-12-26#

Added#

OpenRouter Web Search Plugin: Native web search integration via OpenRouter’s plugins array
- Maps enable_web_search to {"id": "web"} plugin format
- Configurable search engine (exa/native) and max_results parameters
- Added to research preset’s auto-enabled web search backends

Changed#

Persona Generator Diversity Modes: Enhanced persona generation with two diversity modes and phase-based adaptation
- New diversity_mode: perspective (different values/priorities) or implementation (different solution types)
- Phase-based adaptation: strong personas for exploration, softened for convergence
- Multi-turn persistence via persist_across_turns option
- Web UI integration with toggle in coordination settings
Azure OpenAI Multi-Endpoint Support: Support both Azure-specific and OpenAI-compatible endpoints
- Auto-detect endpoint format and use appropriate client (AsyncAzureOpenAI vs AsyncOpenAI)
- Conditionally disable stream_options for Ministral/Mistral models
Environment Variable Expansion in Configs: Use ${VAR} syntax in YAML/JSON config files for flexible configuration

Fixed#

Azure OpenAI Workflow Tool Extraction: Improved JSON parsing with fallback patterns for models outputting tool arguments without tool_name wrapper
Persistent Memory Retrieval: Fixed regression by enabling retrieval on first turn
Backend Tool Registration: Fixed tool registration and updated binary file extensions list

Documentations, Configurations and Resources#

OpenRouter Web Search Configs: New single_openrouter_web_search.yaml and openrouter_web_search.yaml
Azure Multi-Endpoint Config: Updated azure_openai_multi.yaml with env var examples
Diversity Documentation: Updated docs/source/user_guide/advanced/diversity.rst with new diversity modes

Technical Details#

Major Focus: OpenRouter web search, persona diversity modes, Azure OpenAI compatibility
Contributors: @ncrispino @shubham2345 @AbhimanyuAryan @maxim-saplin and the MassGen team

[0.1.29] - 2025-12-24#

Added#

Subagent System: Spawn parallel child MassGen processes for independent task execution
- New spawn_subagents tool for agents to delegate parallelizable work
- Process isolation with independent workspaces per subagent
- Automatic inheritance of parent agent’s backend configuration
- Result aggregation with workspace paths and token usage tracking
- Configurable via enable_subagents, subagent_default_timeout, and subagent_max_concurrent

Changed#

Tool Metrics with Distribution Statistics: Enhanced get_tool_metrics_summary() with per-call averages and output distribution stats (min/max/median)
CLI Config Builder Per-Agent System Messages: New mode in massgen --quickstart for assigning different system messages per agent (“Skip”, “Same for all”, “Different per agent”)

Fixed#

OpenAI Responses API Duplicate Items: Fixed duplicate item errors when using previous_response_id by skipping manual item addition when response ID is passed
Response Formatter Function Call ID Preservation: Preserved ‘id’ field in function_call messages for proper pairing with reasoning items (required by OpenAI Responses API)

Documentations, Configurations and Resources#

Subagent Documentation: New docs/source/user_guide/advanced/subagents.rst with usage guide, configuration examples, and best practices
Subagent Example Configs: New massgen/configs/features/test_subagent_orchestrator.yaml and test_subagent_orchestrator_code_mode.yaml

Technical Details#

Major Focus: Subagent parallel execution system, OpenAI Responses API compatibility
Contributors: @ncrispino and the MassGen team

[0.1.28] - 2025-12-22#

Added#

Web UI Artifact Previewer: Preview workspace artifacts directly in the web interface
- Support for multiple formats: PDF, DOCX, PPTX, XLSX, images, HTML, SVG, Markdown, Mermaid diagrams
- New ArtifactPreviewModal and InlineArtifactPreview components with Sandpack code preview

Changed#

Unified Multimodal Tools: Consolidated read_media for understanding and generate_media for generation
- Understanding: Image, audio, and video analysis with backend selector routing to Gemini, OpenAI, or OpenRouter
- Generation: Create images (gpt-image-1, Imagen), videos (Sora, Veo), and audio (TTS) with provider selection
- New generation/ module with modular _image.py, _video.py, _audio.py implementations
OpenRouter Tool-Capable Model Filtering: Model list now filters to only show models supporting tool calling
- Checks supported_parameters for “tools” capability before including models

Fixed#

Azure OpenAI Tool Calls and Workflow Integration: Comprehensive fixes for Azure OpenAI backend
- Parameter filtering to exclude unsupported Azure parameters (api_version, azure_endpoint, enable_rate_limit)
- Fixed tool_choice parameter handling (only set when tools are provided)
- Message filtering for Azure’s tool message validation requirements
- Fallback extraction for Azure’s {"content":"..."} response format
Web UI Display and Cancellation: Fixed display issues and proper cancellation handling
- Coordination tracker display fixes
- Proper cancellation propagation in web server
Docker Background Shell: Fixed background shell execution in Docker environments
Docker Sudo Configuration: Fixed Dockerfile.sudo configuration

Documentations, Configurations and Resources#

Multimodal Tools Documentation: Updated massgen/tool/_multimodal_tools/TOOL.md with generation capabilities
Web UI Components: New artifact renderer components in webui/src/components/artifactRenderers/

Technical Details#

Major Focus: Multimodal backend integration, artifact preview system, Azure OpenAI compatibility
Contributors: @ncrispino @shubham2345 @AbhimanyuAryan and the MassGen team

[0.1.27] - 2025-12-19#

Added#

Session Sharing via GitHub Gist: Share MassGen sessions with collaborators using massgen export (MAS-16)
- Uploads session logs to GitHub Gist (requires gh CLI authenticated)
- Returns shareable URL to MassGen Viewer (https://massgen.github.io/MassGen-Viewer/?gist=...)
- Manage shares with massgen shares list and massgen shares delete <gist_id>
- Auto-excludes large files, debug logs, and redacts API keys
- New massgen/share.py module (373 lines)
- New massgen/session_exporter.py for session export logic
Log Analysis CLI Command: New massgen logs command for analyzing run logs with metrics visualization, tool breakdown, and export to JSON/CSV formats
- New massgen/logs_analyzer.py with LogAnalyzer class (433 lines)
- Enhanced massgen/cli.py with logs subcommand integration
Per-LLM Call Time Tracking: Detailed timing metrics for individual LLM API calls
- Track time spent on each API call across all backends (Claude, Gemini, OpenAI, Grok)
- Aggregate timing statistics in metrics summary
- Enhanced massgen/backend/base.py with timing instrumentation
- New timing fields in massgen/backend/response.py
Gemini 3 Flash Model Support: Added gemini-3-flash-preview model
- Enhanced massgen/backend/capabilities.py with new models and release dates
- New config: massgen/configs/providers/gemini/gemini_3_flash.yaml
Web UI Context Paths Wizard: New ContextPathsStep component in quickstart wizard for configuring file context paths
Web UI “Open in Browser” Button: Added button to open workspaces directly in browser from answer views
- Enhanced massgen/frontend/web/server.py with browser open endpoint

Changed#

CLI Config Builder Enhancements: Per-agent web search toggles, system message configuration, and improved default model selection
- Enhanced massgen/config_builder.py with _get_provider_capabilities() helper (+234 lines)
- Added per-agent enable_web_search toggle and system message prompts during quickstart
Logging System Improvements: Enhanced logger configuration with better formatting and file output (logger_config.py)

Fixed#

Web Search Call Message Preservation: Fixed response formatter to preserve web_search_call messages like reasoning messages (_response_formatter.py)
Claude Code Tool Permissions: Fixed tool allow issue for Claude Code backend
- Fixed massgen/backend/claude_code.py
- Fixed massgen/filesystem_manager/_filesystem_manager.py
Orchestrator Workflow Timeout: Fixed timeout handling in orchestrator error respawn logic (massgen/orchestrator.py)
Workflow Restart Loop: Fixed issue where workflow would search first then keep running into workflow restarted errors (massgen/backend/response.py)

Documentations, Configurations and Resources#

Session Sharing Documentation:
- Updated docs/source/user_guide/logging.rst: Sharing sessions guide
- Updated docs/source/reference/cli.rst: Export and shares CLI reference
- Updated docs/source/quickstart/running-massgen.rst: Quickstart sharing guide
Log Analysis Documentation:
- Updated docs/source/user_guide/logging.rst: massgen logs command guide
Configuration Examples:
- massgen/configs/providers/gemini/gemini_3_flash.yaml: Gemini 3 Flash configuration
- massgen/configs/debug/error_respawn_test.yaml: Orchestrator error respawn testing
Web UI Components:
- New webui/src/components/wizard/ContextPathsStep.tsx (234 lines): Context paths wizard step
- Enhanced webui/src/stores/wizardStore.ts: Context path state management
- Enhanced webui/src/components/FinalAnswerView.tsx: Share and open in browser buttons

Technical Details#

Major Focus: Session sharing, log analysis tooling, per-LLM timing, CLI config builder UX, Web UI enhancements
Contributors: @ncrispino @praneeth999 and the MassGen team

[0.1.26] - 2025-12-17#

Added#

Docker Diagnostics Module: Comprehensive error detection with platform-specific resolution steps for Docker issues (binary not installed, daemon not running, permission denied, images missing)
Web UI Setup & Configuration System: Guided first-run experience with new SetupPage, ConfigEditorModal, CoordinationStep components, enhanced wizard flow, and backend API endpoints for API key management and environment checks
Shadow Agent Response Depth: Test-time compute scaling via response_depth parameter (low/medium/high) controlling solution complexity in broadcast responses

Changed#

Model Registry Updates: Added GPT-5.1-Codex family (gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1-codex-mini), updated Claude model naming to alias notation (claude-sonnet-4-5), changed defaults to gpt-5.1-codex and claude-opus-4-5
Shadow Agent Claude Code Compatibility: Special handling for Claude Code backend conversation history in shadow agent spawning

Fixed#

Claude Code API Key Handling: Fixed API key configuration and environment variable handling
Web UI Asset Loading: Fixed configuration and static asset paths (MAS-160)
Package Dependencies: Fixed pyproject.toml dependency specification (MAS-161)

Documentations, Configurations and Resources#

Updated agent communication docs with response depth and Claude Code limitation notice; added Claude Code API key examples to backend docs; updated broadcast config examples with response_depth

Technical Details#

Major Focus: Web UI setup experience, Docker diagnostics, shadow agent test-time compute scaling
Contributors: @ncrispino and the MassGen team

[0.1.25] - 2025-12-15#

Added#

UI-TARS Custom Tool: New custom tool for ByteDance’s UI-TARS-1.5-7B model for GUI automation with vision and reasoning
- Connects to UI-TARS via HuggingFace Inference Endpoints
- Image understanding capabilities for browser and desktop automation workflows
GPT-5.2 Model Support: Added OpenAI’s latest GPT-5.2 model as new default (replacing gpt-5.1)
Evolving Skill Creator System: Framework for creating and iterating on reusable workflow plans
- Skills capture steps, Python scripts, and learnings that improve through iteration
- Support for loading skills from previous sessions
- Enhanced system message builder (+67 lines) and system prompt sections (+130 lines)

Changed#

Textual Terminal Display Enhancement: Improved terminal UI with adaptive layouts and dark/light theming
- Adaptive layout management for different terminal sizes and agent states
- Enhanced modal and panel components for better agent coordination visualization

Fixed#

OpenRouter Gemini Reasoning Details: Preserved reasoning_details in streaming responses for complete reasoning chain
LiteLLM Provider Context Paths: Fixed file path handling for configuration and documentation references

Documentations, Configurations and Resources#

UI-TARS Configuration Examples:
- massgen/configs/tools/custom_tools/ui_tars_browser_example.yaml: Browser automation example
- massgen/configs/tools/custom_tools/ui_tars_docker_example.yaml: Docker automation example
Evolving Skills Documentation:
- massgen/configs/skills/skills_with_previous_sessions.yaml: Previous session skills configuration
- massgen/skills/evolving-skill-creator/SKILL.md (209 lines): Skill creator guide
- Updated docs/source/user_guide/tools/skills.rst (+112 lines): Code mode guide
Textual Terminal Themes:
- massgen/frontend/displays/textual_terminal/dark.tcss (+164 lines)
- massgen/frontend/displays/textual_terminal/light.tcss (+180 lines)
Documentation Updates:
- Updated docs/source/reference/python_api.rst (+158 lines): LiteLLM provider guide
- Updated docs/source/reference/supported_models.rst: GPT-5.2 model entry
- Updated docs/source/user_guide/backends.rst (+11 lines): Backend updates

Technical Details#

Major Focus: UI-TARS computer use backend, evolving skills framework, Textual terminal UI improvements
Contributors: @ncrispino @praneeth999 @franklinnwren and the MassGen team

[0.1.24] - 2025-12-12#

Changed#

Enhanced Cost Tracking Across Multiple Backends: Expanded token counting and cost calculation to support additional providers
- Added real-time token usage tracking for OpenRouter, xAI/Grok, Gemini, and Claude Code backends
- New /inspect option c displays detailed cost breakdown with per-agent token usage (input, output, reasoning, cached)
- Per-round token history tracking via get_round_token_history() method
- Aggregated cost totals and tool metrics across all agents in coordination status
- Improved cost ordering and formatting in display tables

Technical Details#

Major Focus: Multi-backend cost tracking with real-time visibility
Contributors: @ncrispino and the MassGen team

[0.1.23] - 2025-12-10#

Added#

Turn History Inspection System: New /inspect command for reviewing agent outputs and coordination data from any turn
- /inspect or /inspect <N> to view specific turn details with interactive menu
- /inspect all to list all turns in the session with task summaries and winning agents
- Menu options for viewing individual agent outputs, final answers, system logs, and coordination tables
Web UI Automation Mode: Streamlined interface for programmatic and monitoring workflows
- New AutomationView component with phase/elapsed time status header and session polling
- --automation flag enables timeline-focused view with LOG_DIR and STATUS path output
- Session persistence API (mark_session_completed) preserves completed sessions in session list

Changed#

Docker Container Persistence for Multi-Turn: Containers now persist across turns for faster transitions
- New SessionMountManager class pre-mounts session directory to Docker containers
- Eliminates container recreation between turns (sub-second vs 2-5 second transitions)
- Automatic visibility of new turn workspace directories without remounting
Multi-Turn Cancellation Handling: Improved Ctrl+C behavior in multi-turn mode
- Flag-based cancellation instead of raising exceptions from signal handlers
- Coordination loop detects cancellation flag and stops Rich display before printing messages
- Terminal state restoration via _restore_terminal_for_input() after display cancellation
- Cancelled turns now build proper history entries with partial results
Async Execution Consistency: New utilities for safe async-from-sync execution
- New run_async_safely() helper for nested event loop handling
- ThreadPoolExecutor pattern prevents async generator ignored GeneratorExit errors
- Fixed mem0 adapter async lifecycle issues

Documentations, Configurations and Resources#

Multi-Turn Mode Documentation: Updated docs/source/user_guide/sessions/multi_turn_mode.rst with /inspect command documentation, turn history inspection examples, and updated slash command reference

Technical Details#

Major Focus: Async consistency, Web UI automation mode, Docker persistence for multi-turn, turn history inspection
Contributors: @ncrispino and the MassGen team

[0.1.22] - 2025-12-08#

Added#

Shadow Agent System: Lightweight agent clones that respond to broadcast questions without interrupting parent agents
- New massgen/shadow_agent.py with ShadowAgentSpawner class (482 lines)
- Shadow agents share parent’s backend (stateless) and copy full conversation history
- Includes parent’s current turn context: text content, tool calls, MCP calls, and reasoning
- Uses simplified system prompt (preserves identity, removes workflow tools)
- Generates tool-free text responses with debug file saving support (--debug flag)

Changed#

Broadcast Channel Architecture: Replaced inject-then-continue pattern with parallel shadow agent spawning
- New _spawn_shadow_agents() method using asyncio.gather() for true parallelization
- Parent agents continue working uninterrupted while shadows respond
- Informational messages injected to parent agents after shadow responds (“FYI, you were asked X…”)
- Deprecated respond_to_broadcast tool (responses now automatic)
Agent Context Tracking: Enhanced SingleAgent to track current turn state for shadow agent access
- New attributes: _current_turn_content, _current_turn_tool_calls, _current_turn_reasoning, _current_turn_mcp_calls
- Context cleared at start of each turn and populated during stream processing
- Enables shadow agents to see parent’s work-in-progress

Documentations, Configurations and Resources#

Agent Communication Documentation: Updated docs/source/user_guide/advanced/agent_communication.rst with shadow agent architecture details, full context responses explanation, and deprecated respond_to_broadcast notice

Technical Details#

Major Focus: Shadow agent architecture for non-blocking, context-aware broadcast responses
Contributors: @ncrispino and the MassGen team

[0.1.21] - 2025-12-05#

Added#

Graceful Cancellation System: Ctrl+C during coordination saves partial progress instead of losing work
- New massgen/cancellation.py with CancellationManager class (177 lines)
- First Ctrl+C saves and exits gracefully; second Ctrl+C forces immediate exit
- In multi-turn mode, first Ctrl+C returns to prompt instead of exiting

Changed#

Session Restoration for Incomplete Turns: Cancelled sessions can be resumed with --continue
- Partial answers combined into conversation history with agent attribution
- All agent workspaces preserved and provided as read-only context on resume
- New get_partial_result() method in Orchestrator for mid-coordination state capture

Documentations, Configurations and Resources#

Graceful Cancellation Guide: New docs/source/user_guide/sessions/graceful_cancellation.rst (196 lines)

Technical Details#

Major Focus: Graceful cancellation with partial progress preservation for multi-turn sessions
Contributors: @ncrispino and the MassGen team

[0.1.20] - 2025-12-03#

Added#

Web UI System: Browser-based real-time visualization for multi-agent coordination
- New massgen/frontend/web/server.py FastAPI server with WebSocket endpoints (1808 lines)
- New massgen/frontend/displays/web_display.py display adapter for web streaming (730 lines)
- React frontend with 18+ components: AgentCarousel, AnswerBrowser, Timeline, VoteVisualization
- CLI flags: --web, --web-port, --web-host for launching web server
- Quickstart wizard, real-time streaming with syntax highlighting, and multi-turn session support

Changed#

Automatic Computer Use Docker Setup: Auto-creates Ubuntu 22.04 container with Xfce desktop for GUI automation
- New setup_computer_use_docker() function with auto-detection of computer_use_docker_example configs
- Container includes X11 virtual display (:99), xdotool, Firefox, Chromium, and scrot
Response API Formatter Enhancement: Improved function call handling for multi-turn contexts
- Preserves function_call entries and generates stub outputs for calls without recorded responses

Fixed#

Web UI Multi-turn Support: Fixed frontend session continuation and follow-up question handling
Timeline Tracking: Fixed timeline arrows and backend event sequencing

Documentations, Configurations and Resources#

Web UI Guide: New docs/source/user_guide/webui.rst (250 lines) covering display modes, timeline visualization, and workspace browsing
Computer Use Documentation: Enhanced docs/source/user_guide/advanced/computer_use.rst (+66 lines) with environment naming conventions and automatic setup instructions
Filesystem-First Mode Documentation: New docs/source/user_guide/filesystem_first.rst (872 lines, experimental v0.2.0+) documenting 98% context reduction via on-demand tool discovery
LLM Council Comparison: New docs/source/reference/comparisons.rst (155 lines) comparing MassGen vs LLM Council with feature tables, UI differences, and architectural comparisons

Technical Details#

Major Focus: Web UI for real-time coordination visualization, automatic Docker setup for computer use agents
Contributors: @voidcenter @ncrispino @praneeth999 and the MassGen team

[0.1.19] - 2025-12-01#

Added#

LiteLLM Integration & Programmatic API: MassGen as a LiteLLM custom provider with direct Python interface
- New massgen/litellm_provider.py with MassGenLLM class and register_with_litellm() (452 lines)
- New run() and build_config() functions for programmatic execution without CLI
- Model string formats: massgen/<example>, massgen/model:<model>, massgen/path:<config>, massgen/build
- New NoneDisplay silent display class for suppressing output in programmatic/LiteLLM use
- Auto-detection of backends from model names (e.g., gpt-5 → openai, claude-sonnet-4-5 → claude)

Changed#

Claude Strict Tool Use & Structured Outputs: Enhanced Claude backend with schema validation and improved defaults
- New enable_strict_tool_use config flag with recursive additionalProperties: false patching
- New output_schema parameter for structured JSON outputs (requires Sonnet 4.5 or Opus 4.1)
- Per-tool opt-out via strict: false on individual tools
- Increased default max_tokens and improved tool_result handling
- ConfigValidator validation for enable_strict_tool_use and output_schema fields
Gemini Exponential Backoff: Automatic retry mechanism for rate limit errors
- New BackoffConfig dataclass with configurable retry parameters
- Handles HTTP 429 (rate limit) and 503 (service unavailable) with jittered backoff
- Retry-After header support and Gemini-specific error pattern matching

Documentations, Configurations and Resources#

Documentation Reorganization: Major restructure into files/, tools/, integration/, sessions/, and advanced/ sections with streamlined quickstart guides
Configuration Examples: massgen/configs/providers/claude/strict_tool_use_example.yaml for strict tool use with custom and MCP tools

Technical Details#

Major Focus: LiteLLM provider integration, Claude strict tool use with structured outputs, Gemini rate limit resilience
Contributors: @ncrispino @praneeth999 and the MassGen team

[0.1.18] - 2025-11-28#

Added#

Agent Communication System: Agents can now ask questions to other agents and optionally humans via the ask_others() tool
- Three modes: disabled (default), agent-to-agent only (broadcast: "agents"), or human-only (broadcast: "human")
- Blocking execution with inline response delivery into agent context
- Human interaction UI with timeout, skip options, and session-persistent Q&A history
- Rate limiting and serialized calls to prevent spam and duplicate prompts
- Comprehensive event tracking in coordination logs
Claude Programmatic Tool Calling: Code execution can now invoke custom and MCP tools programmatically
- New enable_programmatic_flow backend flag that automatically enables code execution sandbox
- Custom and MCP tools callable from Claude’s code sandbox via allowed_callers marking
- Requires claude-opus-4-5 or claude-sonnet-4-5 models with streaming indicators for invocations
Claude Tool Search (Deferred Loading): Server-side tool discovery for large tool sets
- New enable_tool_search flag with tool_search_variant option ("regex" or "bm25")
- Tools with defer_loading: true discovered on-demand, reducing initial context size
- Per-tool and per-MCP-server override support with streaming indicators

Changed#

Backend Capabilities Enhancement: Added tool search and programmatic flow capability flags to massgen/backend/capabilities.py (+17 lines)
ConfigValidator Enhancement: Added enable_programmatic_flow and enable_tool_search boolean field validation (+2 lines)

Documentations, Configurations and Resources#

Claude Advanced Tooling Guide: New docs/claude-advanced-tooling.md covering model requirements, API betas, configuration examples, and streaming cues
Agent Communication Documentation: New docs/source/user_guide/agent_communication.rst with broadcast modes, serialization, Q&A history, and examples
Configuration Examples:
- massgen/configs/providers/claude/programmatic_with_two_tools.yaml - Programmatic tool calling with custom and MCP tools
- massgen/configs/providers/claude/tool_search_example.yaml - Tool search with visible and deferred tools
- massgen/configs/broadcast/test_broadcast_agents.yaml - Agent-to-agent broadcast communication
- massgen/configs/broadcast/test_broadcast_human.yaml - Human broadcast communication with Q&A prompts

Technical Details#

Major Focus: Agent communication system with human broadcast support, Claude programmatic tool calling from code execution, Claude tool search for deferred tool discovery
Contributors: @ncrispino @praneeth999 and the MassGen team

[0.1.17] - 2025-11-26#

Added#

Textual Terminal Display System: Interactive terminal UI using the Textual library for enhanced agent coordination visualization
- New massgen/frontend/displays/textual_terminal_display.py (1673 lines)
- Multi-panel layout with dedicated views for each agent and orchestrator status
- Real-time streaming content display with syntax highlighting support
- Emoji fallback mapping for terminals without Unicode support
- Content filtering for critical patterns (votes, status changes, tools, presentations)
- Keyboard shortcuts for display interaction and safe keyboard mode
- Automatic file output with session logging to agent-specific files
- Thread-safe display updates with buffered content batching
Dark and Light Themes: TCSS stylesheets for customizable terminal appearance
- New massgen/frontend/displays/textual_themes/dark.tcss (322 lines)
- New massgen/frontend/displays/textual_themes/light.tcss (322 lines)
- VS Code-inspired color schemes with styled containers for post-evaluation and final stream panels

Changed#

CoordinationUI Enhancement: Extended display coordination with Textual Terminal support
- Enhanced massgen/frontend/coordination_ui.py with Textual display integration (+348 lines)
- New textual_terminal display type option alongside existing rich_terminal and simple displays
- Automatic fallback when Textual library is not available
- Unified reasoning content processing across all display types
Display Module Restructuring: Improved display initialization and base class architecture
- Enhanced massgen/frontend/displays/__init__.py with Textual display exports (+30 lines)
- Enhanced massgen/frontend/displays/terminal_display.py with shared base functionality (+45 lines)
- Better separation of concerns between display implementations

Documentations, Configurations and Resources#

Textual Configuration Example: Reference configuration for Textual terminal display
- New massgen/configs/basic/single_agent_textual.yaml (17 lines)
Dependencies: Added Textual library for modern terminal UI
- Updated pyproject.toml and requirements.txt with textual>=0.47.0

Technical Details#

Major Focus: Textual Terminal Display for enhanced agent coordination visualization with theme support
Contributors: @praneeth999 and the MassGen team

[0.1.16] - 2025-11-24#

Added#

Terminal Evaluation System: Automated terminal session recording and AI-powered evaluation using VHS
- New docs/source/user_guide/terminal_evaluation.rst comprehensive evaluation guide (450 lines)
- New massgen/tests/test_terminal_evaluation.py with test suite (336 lines)
- New massgen/tests/demo_terminal_evaluation.py demonstration script (210 lines)
- Records terminal sessions as GIFs using VHS (Video Home System)
- Analyzes session recordings with multimodal models (GPT-4.1, Claude)
- Evaluates agent performance, UI quality, and interaction patterns
- Automated testing workflows for continuous quality monitoring
LiteLLM Cost Tracking Integration: Accurate cost calculation using LiteLLM’s pricing database
- New calculate_cost_with_usage_object() in massgen/token_manager/token_manager.py (+178 lines)
- New docs/dev_notes/litellm_cost_tracking_integration.md design documentation (581 lines)
- New massgen/tests/test_litellm_integration.py comprehensive test suite (331 lines)
- New massgen/tests/test_backend_cost_tracking.py integration tests (183 lines)
- Integrates LiteLLM pricing database covering 500+ models with auto-updates
- Handles reasoning tokens for o1/o3 models with separate pricing
- Handles cached tokens for Claude and OpenAI prompt caching
- Fallback to legacy calculation when LiteLLM unavailable
- More accurate cost estimates than manual price tables
Memory Archiving System: Persistent memory with multi-turn session support
- Enhanced massgen/orchestrator.py with memory archiving capabilities (+51 lines)
- Enhanced massgen/system_message_builder.py with archive management (+170 lines)
- Enhanced massgen/system_prompt_sections.py with archiving instructions (+201 lines)
- Enhanced massgen/cli.py with session continuation support (+15 lines)
- Enables archiving long-term memory for session persistence
- Supports multi-turn conversations with memory continuity
- Improved memory retrieval and context management
MassGen Self-Evolution Skills: Skills for MassGen to develop and maintain itself
- New massgen/skills/massgen-config-creator/SKILL.md for creating valid YAML configurations (183 lines)
- New massgen/skills/massgen-develops-massgen/SKILL.md for self-improvement and feature development (490 lines)
- New massgen/skills/massgen-release-documenter/SKILL.md for changelog and documentation updates (252 lines)
- New massgen/skills/model-registry-maintainer/SKILL.md for maintaining model registry (483 lines)
- Enables MassGen to maintain its own codebase and documentation
- Self-documenting release workflows
- Automated configuration validation and generation
- Model registry updates with pricing and capability tracking

Changed#

Docker Infrastructure Enhancement: Parallel image pulling, VHS recording support, and improved container management
- Enhanced massgen/cli.py with parallel Docker image pulling (+242 lines)
- Enhanced massgen/docker/Dockerfile with VHS installation and improved build process (+44 lines total)
- Enhanced massgen/docker/Dockerfile.sudo with VHS support and enhanced permissions (+47 lines total)
- Enhanced massgen/filesystem_manager/_filesystem_manager.py with VHS utilities and better Docker integration (+50 lines)
- Parallel pulling of multiple Docker images for faster setup
- VHS (Video Home System) integration for terminal session recording in Docker containers
- Better error handling and progress reporting
- Improved Docker container lifecycle management
Model Registry Updates: Expanded model support with accurate pricing and metadata
- Enhanced massgen/backend/capabilities.py with new models and release dates (+45 lines)
- Added Grok 4.1 family models (grok-4.1, grok-4.1-mini) with pricing
- Added GPT-4.1 family models for terminal evaluation
- Added release dates to all models in BACKEND_CAPABILITIES
- Removed o4 models (don’t exist in production)
- Removed unsupported Gemini experimental models
- Improved model metadata for better cost tracking
Configuration Builder Enhancement: Improved model selection and configuration workflow
- Enhanced massgen/config_builder.py with better model defaults (+73 lines)
- Enhanced massgen/cli.py with improved config selection interface (+65 lines)
- Better model recommendations based on use case
- Improved validation and error messages

Fixed#

Status Mode Log Directory: Fixed missing log directory creation in status mode
- Fixed massgen/cli.py to create log directories before writing
- Prevents errors when running in status/automation mode
Filesystem Docker Zod Schema: Resolved MCP tool argument parsing in Docker
- Enhanced massgen/backend/chat_completions.py with schema validation (+16 lines)
- Enhanced massgen/backend/claude_code.py with improved MCP handling (+13 lines)
- Enhanced massgen/mcp_tools/security.py with schema fixes (+2 lines)
- Fixed Zod schema errors preventing proper tool call execution
- MCP tools now correctly parse arguments in Docker filesystem mode

Documentations, Configurations and Resources#

Terminal Evaluation Documentation: Complete guide for automated terminal testing
- New docs/source/user_guide/terminal_evaluation.rst with setup and usage (450 lines)
- Covers VHS configuration, recording workflows, evaluation strategies
- Best practices for multimodal session analysis
Memory Filesystem Mode Enhancement: Expanded documentation for memory integration
- Updated docs/source/user_guide/memory_filesystem_mode.rst with archiving workflows (+172 lines)
- Documents memory persistence across sessions
- Multi-turn conversation patterns with memory continuity
- Best practices for long-running agent interactions
Skills Documentation Updates: Enhanced skills guide with self-evolution examples
- Updated docs/source/user_guide/skills.rst with MassGen self-evolution skills (+178 lines)
- Documents the four new MassGen-specific skills
- Examples of self-maintaining systems
- Guidelines for creating meta-skills
Custom Tools Documentation: Improved custom tools integration guide
- Updated docs/source/user_guide/custom_tools.rst with terminal evaluation examples (+103 lines)
- Documents VHS integration patterns
- Best practices for recording and evaluation tools
Configuration Examples: New YAML configurations for v0.1.16 features
- New massgen/configs/meta/massgen_evaluates_terminal.yaml for terminal evaluation (72 lines)
- New massgen/configs/tools/custom_tools/terminal_evaluation.yaml example config (88 lines)
- Updated massgen/configs/skills/test_memory.yaml with memory archiving examples
- Updated massgen/configs/tools/filesystem/code_based/example_code_based_tools.yaml with Docker improvements

Technical Details#

Major Focus: Terminal evaluation infrastructure, LiteLLM cost tracking integration, memory archiving system, MassGen self-evolution capabilities
Contributors: @ncrispino and the MassGen team

[0.1.15] - 2025-11-21#

Added#

Persona Generation System: Automatic generation of diverse system messages for multi-agent configurations
- New massgen/persona_generator.py for LLM-powered persona creation (365 lines)
- Enhanced massgen/orchestrator.py with persona generation orchestration (+122 lines)
- Enhanced massgen/agent_config.py with persona configuration support (+5 lines)
- Enhanced massgen/cli.py with --generate-personas flag (+54 lines)
- Multiple generation strategies: complementary, diverse, specialized, adversarial
- Configurable backend for persona generation (defaults to gpt-4o-mini)
- Custom persona guidelines support for domain-specific generation
- Increases response diversity without manual system message crafting

Changed#

Docker Distribution & Custom Tools Enhancement: GitHub Container Registry integration with custom tools support
- Enhanced .github/workflows/docker-publish.yml with comprehensive CI/CD pipeline (+96 lines)
- Enhanced massgen/docker/Dockerfile and Dockerfile.sudo with MassGen pre-installation (+13 lines each)
- Enhanced massgen/filesystem_manager/_docker_manager.py with improved container management (+37 lines)
- Enhanced massgen/cli.py with Docker-related commands and improvements (+104 lines)
- Custom tools can now run in isolated Docker containers for security and portability (Issue #510)
- ARM architecture support for Apple Silicon and ARM-based cloud instances
- Automated Docker image pruning during CI builds
Config Builder Enhancement: Improved interactive configuration experience
- Enhanced massgen/config_builder.py with better model selection and defaults (+17 lines)

Documentations, Configurations and Resources#

Installation Documentation Overhaul: Comprehensive Docker and setup guides
- Updated docs/source/quickstart/installation.rst with Docker installation instructions (+150 lines)
- Updated docs/source/index.rst with improved getting started guide (+66 lines)
- Detailed GitHub Container Registry pull instructions
- Platform-specific Docker setup guidance
Persona Generation Configuration Example: Reference configuration for persona diversity
- New massgen/configs/basic/multi/persona_diversity_example.yaml with strategy and backend configuration (123 lines)
Pre-commit Hooks Enhancement: Additional code quality checks
- New scripts/precommit_check_package_name.py for package name validation (39 lines)
- Updated .pre-commit-config.yaml with package name check (+6 lines)

Technical Details#

Major Focus: Persona generation for agent diversity, Docker distribution improvements, GitHub Container Registry integration
Contributors: @ncrispino and the MassGen team

[0.1.14] - 2025-11-19#

Added#

Parallel Tool Execution System: Configurable concurrent tool execution across all backends with asyncio-based scheduling
- New concurrent_tool_execution configuration parameter for local parallel execution control
- New parallel_tool_calls parameter support for OpenAI Response API (controls model behavior)
- New disable_parallel_tool_use parameter for Claude backend (inverse toggle for tool parallelism)
- New max_concurrent_tools semaphore limit for execution speed control (default: 10)
- Enhanced massgen/backend/response.py with parallel execution infrastructure (+239 lines)
- Enhanced massgen/backend/base_with_custom_tool_and_mcp.py with _execute_tool_calls method (+186 lines)
- Enhanced massgen/api_params_handler/_response_api_params_handler.py with parameter handling (+20 lines)
- Unified handling of custom and MCP tool calls with optional concurrent execution
- Works with Response, ChatCompletions, Gemini, and Claude backends
- Model-level controls (parallel_tool_calls) separate from local execution controls (concurrent_tool_execution)
Gemini 3 Pro Model Support: Full integration for Google’s Gemini 3 Pro model with function calling
- Enhanced massgen/backend/gemini.py with Gemini 3 Pro compatibility (60 lines modified)
- Fixed function calling behavior specific to Gemini 3 Pro model
- Native support for Gemini’s parallel function calling capabilities

Changed#

Config Builder Enhancement: Interactive quickstart workflow with guided configuration creation
- Enhanced massgen/config_builder.py with interactive prompts and improved UX (+394 lines)
- Enhanced massgen/cli.py with quickstart command integration and improved interface (+214 lines)
- Enhanced massgen/backend/capabilities.py with model metadata (+3 lines)
- Streamlined onboarding experience from setup to first run
- Improved provider selection and configuration validation
- Better integration with config selection workflow
- Better error messages and user guidance
- Previously introduced in v0.1.9, now significantly enhanced for user experience
MCP Registry Client: Enhanced MCP server metadata fetching with official registry integration
- New massgen/mcp_tools/registry_client.py for fetching server descriptions from official MCP registry (358 lines)
- New massgen/tests/test_mcp_registry_client.py comprehensive test suite (184 lines)
- Enhanced massgen/mcp_tools/security.py with registry integration (+49 lines)
- Fetches metadata from https://registry.modelcontextprotocol.io/v0/servers
- Enhances system prompts with server descriptions for better agent understanding
- Builds upon v0.1.13’s MCP server registry (server_registry.py) with external registry support
Planning System Enhancements: Improved skill and tool search capabilities in planning mode
- Enhanced massgen/mcp_tools/planning/_planning_mcp_server.py with better search logic (+44 lines)
- Enhanced massgen/system_prompt_sections.py with refined planning prompts (+34 lines)
- Enhanced massgen/orchestrator.py with planning coordination (+21 lines)
- Enhanced massgen/system_message_builder.py with planning context (+12 lines)
- PR #534: Commit 98b1ec6f
- Better discovery of available skills and tools during planning phase
- Improved agent decision-making for tool selection
- More accurate task decomposition with tool awareness
NLIP Routing Streamlining: Simplified and unified NLIP execution flow across backends
- Refactored massgen/backend/response.py with streamlined routing (net -209 lines)
- Refactored massgen/backend/claude.py with unified handling (+98 lines modified)
- Refactored massgen/backend/gemini.py with consistent patterns (+178 lines modified)
- Unified custom and MCP tool call handling with improved NLIP routing
- Reduced code complexity while maintaining full NLIP functionality
- Better error handling and async management in NLIP message routing
- Builds upon v0.1.13’s NLIP integration with cleaner implementation
Coordination Tracking Enhancement: Improved status monitoring for automation workflows
- Enhanced massgen/coordination_tracker.py with parallel tool execution tracking (+23 lines)
- Better visibility into concurrent tool execution status for automation mode

Documentations, Configurations and Resources#

Parallel Tool Execution Configuration Guide: Comprehensive documentation for tool execution parallelism
- New docs/parallel-tool-execution.md complete configuration reference (179 lines)
- Explains model-level vs. local execution controls
- Backend-specific configuration examples for OpenAI, Claude, Gemini
- Quick reference for all parallelism-related parameters
- Execution flow diagrams and best practices
Configuration Examples: New YAML configurations demonstrating v0.1.14 features
- massgen/configs/tools/custom_tools/gpt5_nano_custom_tool_with_mcp_parallel.yaml: Parallel tool execution example with configurable concurrency
- massgen/configs/tools/filesystem/code_based/example_code_based_tools.yaml: Updated with enhanced instructions for code-based tools (+52 lines)
- massgen/configs/providers/gemini/gemini_3_pro.yaml: Configuration template for Gemini 3 Pro model (30 lines)
CI/CD Workflow Configuration: Docker image publishing automation
- .github/workflows/docker-publish.yml: Automated Docker build and publish workflow for releases (60 lines)
- Integration with GitHub Container Registry for automated container deployment
Docker Configuration Updates: Enhanced Docker setup for development and deployment
- massgen/docker/Dockerfile: Improvements for standard Docker builds (+7 lines)
- massgen/docker/Dockerfile.sudo: Enhanced sudo mode support (+7 lines)

Technical Details#

Major Focus: Parallel tool execution infrastructure, interactive quickstart experience, MCP registry client integration, Gemini 3 Pro support, NLIP routing optimization
Contributors: @praneeth999 @ncrispino and the MassGen team

[0.1.13] - 2025-11-17#

Added#

Code-Based Tools System (CodeAct Paradigm): Tool integration via importable Python code instead of schema-based tools
- New massgen/filesystem_manager/_tool_code_writer.py for writing MCP tool wrappers to workspace (450 lines)
- New massgen/mcp_tools/code_generator.py for generating Python wrapper code from MCP schemas (507 lines)
- New massgen/mcp_tools/server_registry.py for MCP server catalog with auto-discovery (205 lines)
- Enhanced massgen/filesystem_manager/_filesystem_manager.py with code-based tools setup (+562 lines)
- Agents import and use tools as native Python functions with type hints and docstrings
- Reduces token usage by 98% through on-demand tool loading (Anthropic research)
- Pre-configured registry with popular MCP servers (Playwright, GitHub, Context7, Memory)
- Auto-discovery eliminates manual MCP server configuration
NLIP (Natural Language Interface Protocol) Integration: Advanced tool routing with natural language interface
- Enhanced massgen/backend/response.py with NLIP routing infrastructure (+134 lines)
- Enhanced massgen/backend/claude.py, gemini.py, chat_completions.py with NLIP support (+255 lines total)
- Enhanced massgen/orchestrator.py with orchestrator-level NLIP configuration (+48 lines)
- Routes tool execution requests through natural language interface
- Multi-backend support across Claude, Gemini, and OpenAI
- Per-agent or orchestrator-level configuration with fallback to direct execution
- Enables natural language task decomposition and intelligent tool selection
Skills Installation System: Cross-platform automated skills installer
- New massgen/utils/skills_installer.py for automated skills installation (350 lines)
- New scripts/init_skills.sh and scripts/init.sh for shell-based setup (650 lines total)
- massgen --setup-skills command for one-command installation
- Installs openskills CLI, Anthropic skills collection, and Crawl4AI skill
- Cross-platform support: Windows, macOS, Linux with idempotent installation
- Comprehensive progress indicators and error handling

Changed#

Tool Size & Command-Line Enhancements: Increased tool capacity and improved CLI execution
- Updated massgen/backend/utils.py tool truncation threshold from 10,000 to 15,000 characters
- Enhanced massgen/backend/bash_cli.py with command-line-only mode improvements
- Commit: b51067b8 “Command line only mode; increase tool size from 10k to 15k”
- Allows more comprehensive tool documentation and examples
- Improved command parsing and error handling
- Better integration with code-based tools workflow
Exclude File Operation MCPs: Removed filesystem MCP tools in favor of native file operations
- Updated massgen/mcp_tools/mcp_manager.py to exclude @modelcontextprotocol/server-filesystem (+204 lines)
- Commit: 5bdf46bf “Adjusted prompts and added TOOL.md for custom tools”
- Prevents redundancy with MassGen’s built-in filesystem operations
- Reduces token usage from duplicate tool definitions
- Clearer tool usage patterns for agents

Documentations, Configurations and Resources#

TOOL.md Documentation System: Standardized documentation format for custom tools
- New massgen/tool/_video_tools/TOOL.md for video tools documentation (161 lines)
- New massgen/tool/_web_tools/TOOL.md for web scraping tools documentation (161 lines)
- New massgen/tool/_playwright_mcp/TOOL.md for Playwright MCP documentation (201 lines)
- Standardized structure: name, description, category, tasks, keywords, usage examples
- Frontmatter metadata in YAML format for tool discovery
- Clear “When to Use This Tool” and “When NOT to Use” sections
- Function signatures with parameter descriptions and return types
- Configuration prerequisites and setup instructions
- Common use cases and limitations documentation
- Enables agents to understand tool capabilities and make informed decisions
- Total: 12 new TOOL.md files across custom tools directory (~3,800 lines)
Configuration Examples: New YAML configurations for v0.1.13 features
- massgen/configs/tools/filesystem/code_based/example_code_based_tools.yaml: Code-based tools with auto-discovery and shared tools directory (153 lines)
- massgen/configs/tools/filesystem/exclude_mcps/test_minimal_mcps.yaml: Minimal MCPs with command-line file operations and memory filesystem mode (37 lines)
- massgen/configs/examples/nlip_basic.yaml: Basic NLIP protocol support with router and translation settings (54 lines)
- massgen/configs/examples/nlip_openai_weather_test.yaml: OpenAI with NLIP integration for custom tools and MCP servers (36 lines)
- massgen/configs/examples/nlip_orchestrator_test.yaml: Orchestrator-level NLIP configuration for multi-agent coordination (47 lines)
Skills Installation Documentation: Comprehensive guides for skills setup
- Updated scripts/init.sh with detailed help text and options (438 lines)
- Updated scripts/init_skills.sh with skip flags for selective installation (212 lines)
- Examples: ./init.sh --skip-docker, ./init_skills.sh --skip-anthropic
Code-Based Tools User Guide: Complete documentation for CodeAct paradigm implementation
- New docs/source/user_guide/code_based_tools.rst (726 lines)
- Quick start examples and configuration
- Explains 98% context reduction benefit (Anthropic research)
- Covers workspace structure, Python wrapper generation, async workflows
- Real-world examples: weather forecasting, GitHub integration, multi-tool composition
MCP Server Registry Reference: Documentation for built-in MCP server catalog
- New docs/source/reference/mcp_server_registry.rst (219 lines)
- Documents all pre-configured MCP servers (Context7, GitHub, Filesystem, Memory, etc.)
- Connection examples and tool listings
- API key requirements and configuration
- Auto-discovery setup instructions
Installation Guide Updates: Enhanced setup documentation with automation scripts
- Updated docs/source/quickstart/installation.rst (+115 lines)
- Automated development setup using scripts/init.sh
- Script options and flags documentation
- System requirements and verification steps
- Windows support roadmap notes
Documentation Updates: Enhanced existing guides with v0.1.13 features
- Updated docs/source/user_guide/file_operations.rst (+44 lines) - Code-based tools integration
- Updated docs/source/user_guide/mcp_integration.rst (+71 lines) - Registry and auto-discovery
- Updated docs/source/reference/yaml_schema.rst (+5 lines) - Code-based tools configuration options

Technical Details#

Major Focus: CodeAct paradigm implementation, MCP registry infrastructure, skills installation automation, TOOL.md documentation standard, self-evolution capabilities, NLIP integration
Contributors: @qidanrui @ncrispino @franklinnwren @praneeth999 and the MassGen team

[0.1.12] - 2025-11-14#

Added#

Semtools Skill: Semantic search capabilities using embedding-based similarity matching
- New massgen/skills/semtools/SKILL.md for meaning-based code and document search (606 lines)
- Rust-based CLI for high-performance semantic search beyond keyword matching
- Workspace management for indexing large codebases with fast repeated searches
- Document parsing support for PDFs, DOCX, PPTX with optional API integration
- Discovery-focused search finding relevant code without knowing exact keywords
- Complements traditional ripgrep (keyword) and ast-grep (syntax) search tools
Serena Skill: Symbol-level code understanding via Language Server Protocol (LSP)
- New massgen/skills/serena/SKILL.md for IDE-like semantic code analysis (499 lines)
- Symbol discovery across 30+ programming languages (classes, functions, variables, types)
- Reference tracking to find all usage locations of symbols
- Precise code editing with surgical symbol-level insertions
- LSP-powered understanding of code structure, scope, and relationships
- Enables symbol-aware refactoring and navigation capabilities
System Message Builder: New modular system for constructing agent prompts
- New massgen/system_message_builder.py for flexible prompt composition (488 lines)
- Separates prompt construction logic from orchestrator
- Enables better organization and reusability of system prompt components
- Foundation for improved prompt engineering and customization

Changed#

System Prompt Architecture: Complete refactoring for improved LLM attention and effectiveness
- Enhanced massgen/system_prompt_sections.py with hierarchical prompt structure (1286 lines)
- Reorganized prompt ordering to place critical instructions (skills, memory) at optimal positions
- Reduced message template redundancy in message_templates.py (-682 lines)
- Simplified orchestrator prompt assembly in orchestrator.py (-428 lines)
- Applied 2025 prompt engineering best practices: XML structure, attention management, priority signaling
- Improved skills and memory system visibility to agents through better positioning
Skills System Refactoring: Enhanced architecture with local execution support
- Local Mode: Skills can now execute directly without Docker containers
- Directory Reorganization: Moved file-search from skills/always/file_search/ to skills/file-search/
- Semantic Search Skills: Promoted semtools and serena from optional to core skills directory
- Enhanced massgen/filesystem_manager/skills_manager.py for local execution support
- Enhanced massgen/filesystem_manager/_code_execution_server.py for local skill commands (+71 lines)
- Enhanced massgen/filesystem_manager/_filesystem_manager.py with local mode capabilities (+173 lines)
- Enhanced massgen/filesystem_manager/_docker_manager.py for skills integration (+59 lines)
- Updated massgen/backend/claude_code.py for local skill execution (+26 lines)
Gemini Computer Use Tool: Multi-agent support with Docker integration
- Enhanced massgen/tool/_gemini_computer_use/gemini_computer_use_tool.py (949 lines total, +446 lines)
- Added Docker container support for browser and desktop automation
- New screenshot capture functions for Docker environments (take_screenshot_docker)
- New action execution system for Docker (execute_docker_action)
- X11 display integration with xdotool for precise control
- VNC compatibility for remote visualization and debugging
- Multi-agent coordination capabilities for collaborative computer use
Browser Automation Tool: Enhanced screenshot management
- Updated massgen/tool/_browser_automation/browser_automation_tool.py to save screenshots as files (+39 lines)
- New output_filename parameter to save screenshots directly to agent workspace
- Automatic workspace path resolution with agent_cwd parameter
- Reduces token usage by avoiding base64-encoded screenshot returns
- Better integration with file-based workflows and serena skill

Documentations, Configurations and Resources#

System Prompt Architecture Documentation: Comprehensive design document for prompt refactoring
- New docs/dev_notes/system_prompt_architecture_redesign.md (593 lines)
- Documents LLM attention management and hierarchical structure principles
- Explains XML-based prompt engineering for Claude models
- Covers priority signaling and position-based emphasis strategies
- Implementation roadmap for future prompt improvements
Computer Use Visualization Guide: Multi-agent computer use documentation
- New docs/backend/docs/COMPUTER_USE_VISUALIZATION.md (455 lines)
- Covers VNC setup and remote visualization workflows
- Documents multi-agent coordination patterns for computer use
- Troubleshooting guide for Docker-based automation
- Architecture diagrams for computer use tool integration
Skills Documentation Update: Enhanced skills system guide
- Updated docs/source/user_guide/skills.rst with local mode documentation (+222 lines)
- Covers new semantic search skills (semtools/serena)
- Documents skill directory reorganization
- Local vs Docker execution trade-offs and best practices
YAML Schema Documentation: Configuration reference updates
- Updated docs/source/reference/yaml_schema.rst with skills configuration options (+36 lines)
- Documents local mode parameters and skill settings
Computer Use Tools Guide: Enhanced documentation
- Updated docs/backend/docs/COMPUTER_USE_TOOLS_GUIDE.md with Gemini Docker support (+94 lines)
- Multi-agent computer use configuration examples
- VNC viewer setup instructions
Configuration Examples: New YAML configurations for v0.1.12 features
- massgen/configs/tools/custom_tools/multi_agent_computer_use_example.yaml: Multi-agent coordination for computer use (194 lines)
- massgen/configs/tools/custom_tools/gemini_computer_use_docker_example.yaml: Gemini with Docker automation (84 lines)
- Updated massgen/configs/tools/custom_tools/simple_browser_automation_example.yaml: File-based screenshot workflow
VNC Viewer Script: Automated VNC setup for computer use visualization
- New scripts/enable_vnc_viewer.sh for quick VNC configuration (40 lines)
- Streamlines Docker-based computer use debugging and monitoring

Technical Details#

Major Focus: System prompt architecture refactoring, semantic search skills (semtools/serena), local skill execution, multi-agent computer use with Docker
Contributors: @ncrispino @franklinnwren @Henry-811 and the MassGen team

[0.1.11] - 2025-11-12#

Added#

Skills System: Modular prompting framework for enhancing agent capabilities
- New SkillsManager class in massgen/filesystem_manager/skills_manager.py for dynamic skill loading and injection (158 lines)
- File Search Skill: Always-available skill for searching files and code across workspace (massgen/skills/always/file_search/SKILL.md, 280 lines)
- Automatic skill discovery and loading from massgen/skills/ directory structure
- Docker-compatible skill mounting and environment setup
- Skills organized into always/ (auto-included) and optional/ categories
- Flexible skill injection into agent system prompts via orchestrator
- Configuration examples in massgen/configs/skills/ (skills_basic.yaml, skills_existing_filesystem.yaml, skills_with_memory.yaml)
Memory MCP Tool & Filesystem Integration: MCP server for agent memory management with filesystem persistence and combined workflows
- New massgen/mcp_tools/memory/ module with memory MCP server implementation (513 lines total)
- MemoryMCPServer in _memory_mcp_server.py (352 lines) for memory CRUD operations with automatic filesystem sync
- Memory data models in _memory_models.py (161 lines) with short-term and long-term memory tiers
- Memory persistence to workspace under memory/short_term/ and memory/long_term/ directories
- Markdown-based memory storage format for human readability
- Integration with orchestrator for cross-agent memory sharing (+218 lines in orchestrator.py)
- Memory-specific message templates for memory operations (+95 lines in message_templates.py)
- Combined workflows: Simultaneous use of memory MCP tools and filesystem operations for advanced workflows
- Enables agents to maintain persistent memory while manipulating files
- Configuration examples demonstrating integrated workflows for long-running projects requiring both code changes and learned context
- Inspired by Letta’s context hierarchy design pattern
Rate Limiting System (Gemini): Multi-dimensional rate limiting for Gemini API calls and agent startup
- New massgen/backend/rate_limiter.py (321 lines) with comprehensive rate limiting infrastructure
- Support for multiple limit types: requests per minute (RPM), tokens per minute (TPM), requests per day (RPD)
- Model-specific rate limits with configurable thresholds for Gemini models
- Graceful cooldown periods with exponential backoff
- Agent startup rate limiting to prevent API quota exhaustion
- Test suite in massgen/tests/test_rate_limiter.py (122 lines)
- Configuration system in massgen/configs/rate_limits/ with rate_limits.yaml and rate_limit_config.py (180 lines)
- CLI flag --enable-rate-limiting for opt-in rate limiting

Changed#

Claude Code Backend: Improved Windows support for long system prompts
- Enhanced handling of long system prompts on Windows platforms
- Resolved command-line length limitations and encoding issues
- Updated massgen/backend/claude_code.py with more robust Windows compatibility (27 lines changed)
Planning MCP Server: Added filesystem task persistence within workspace
- Tasks now saved to agent workspace instead of separate tasks/ directory
- Improved task organization and workspace management
- Enhanced massgen/mcp_tools/planning/_planning_mcp_server.py (+84 lines)
- Removed standalone tasks/ skill in favor of integrated planning

Fixed#

Rate Limiter Asyncio Lock: Resolved asyncio lock event loop error
- Fixed asyncio lock reuse across different event loops causing errors
- Improved rate limiter thread safety and event loop handling
- Updated massgen/backend/rate_limiter.py and added comprehensive tests

Documentations, Configurations and Resources#

Skills System Documentation: Comprehensive guide for using and creating skills
- New docs/source/user_guide/skills.rst (473 lines)
- Covers skill structure, loading mechanisms, and best practices
- Examples of creating custom skills for specific agent capabilities
Memory-Filesystem Mode Documentation: Guide for integrated memory and filesystem workflows
- New docs/source/user_guide/memory_filesystem_mode.rst (883 lines)
- Demonstrates combining memory MCP tools with filesystem operations
- Configuration examples and use case scenarios
Rate Limiting Documentation: Complete rate limiting configuration guide
- New docs/rate_limiting.md (254 lines)
- Model-specific rate limits and configuration examples
- Best practices for managing API quotas
- New massgen/configs/rate_limits/README.md (108 lines)
Skills Configuration Examples: Three YAML configurations for skills usage
- massgen/configs/skills/skills_basic.yaml: Basic skills setup
- massgen/configs/skills/skills_existing_filesystem.yaml: Skills with filesystem integration
- massgen/configs/skills/skills_with_memory.yaml: Skills with memory MCP integration
Filesystem Tool Discovery Design: Comprehensive design document for new tool paradigm
- New docs/dev_notes/filesystem_tool_discovery_design.md (1,582 lines)
- Proposes shift from context-based to filesystem-based tool discovery
- Enables attaching 100+ MCP servers without context pollution
- Details progressive disclosure and code-based tool composition
- Includes implementation proposals and technical architecture

Technical Details#

Major Focus: Skills system for modular agent prompting, memory MCP tool with filesystem persistence, multi-dimensional rate limiting, memory-filesystem integration mode
Contributors: @ncrispino @abhimanyuaryan @qidanrui @sonichi @Henry-811 and the MassGen team

[0.1.10] - 2025-11-10#

Added#

Docker Custom Image Support: Example Dockerfile for extending MassGen base image with custom packages
- New massgen/docker/Dockerfile.custom-example demonstrating how to add ML/data science packages, development tools, and system utilities
- Template for creating specialized Docker images for specific project needs

Changed#

Docker Authentication Configuration: Restructured to nested dictionary format for better organization
- New command_line_docker_credentials structure consolidating all credential-related settings
- Nested mount array for credential file mounting (ssh_keys, git_config, gh_config, npm_config, pypi_config)
- Nested env_file, env_vars, and pass_all_env for environment variable management
- Nested additional_mounts for custom volume mounting
- Migration from flat parameters (command_line_docker_mount_ssh_keys, command_line_docker_pass_env_vars, etc.) to organized nested structure
- Enhanced massgen/filesystem_manager/_docker_manager.py and _filesystem_manager.py with new configuration parsing
Docker Package Management: New nested configuration structure for dependency installation
- New command_line_docker_packages structure with auto_install_deps, auto_install_on_clone, and preinstall settings
- Support for pre-installing Python, npm, and system packages before agent execution
- Improved dependency detection and installation workflow
Framework Interoperability Streaming: Real-time intermediate step streaming for external framework agents
- LangGraph Streaming: Updated massgen/tool/_extraframework_agents/langgraph_lesson_planner_tool.py (78 lines changed)
  - Now yields intermediate updates from each workflow node (standards, lesson_plan, reviewed_plan)
  - Distinguishes between logs (is_log=True) and final output using result type
  - Enables real-time progress tracking during LangGraph workflow execution
- SmoLAgent Streaming: Updated massgen/tool/_extraframework_agents/smolagent_lesson_planner_tool.py (60 lines changed)
  - Streams ActionStep and PlanningStep outputs as logs during agent execution
  - FinalAnswerStep yielded as final output
  - Set verbosity_level=0 to prevent duplicate console output
- Both frameworks now provide visibility into multi-step reasoning processes
Parallel Execution Safety: Extended automatic workspace isolation to all execution modes
- Parallel execution safety now works in both --automation and normal modes (previously automation-only)
- Automatic Docker container naming with unique instance ID suffixes (e.g., massgen-agent_a-a1b2c3d4)
- Enhanced massgen/filesystem_manager/_filesystem_manager.py with instance ID generation for all modes

Fixed#

Session Management: Resolved CLI session handling issues
- Fixed session restoration edge cases in massgen/cli.py
- Improved error handling for session state loading

Documentations, Configurations and Resources#

MassGen Contributor Handbook: Comprehensive contributor guide addressing issue #387
- New handbook website at https://massgen.github.io/Handbook/
- Eight major sections: Case Studies, Issues, Development, Documentation, Release, Announcements, Marketing, and Resources
- Workflow diagrams illustrating contribution pipeline from research to release
- Seven contribution tracks with assigned track owners
- Communication channels and meeting schedules (daily sync 5:30pm PST, research 6:00pm PST)
- Getting started guide for new contributors
Docker Configuration Examples: Three new YAML configurations for advanced Docker workflows
- massgen/configs/tools/code-execution/docker_custom_image.yaml: Using custom Docker images
- massgen/configs/tools/code-execution/docker_full_dev_setup.yaml: Complete development environment setup
- massgen/configs/tools/code-execution/docker_github_readonly.yaml: Read-only GitHub access configuration
Automation Documentation: Enhanced parallel execution section
- Updated docs/source/user_guide/automation.rst clarifying automatic isolation works in all modes
- Added Docker container isolation examples with unique container naming
- Clarified that --automation flag is for output control, not parallel safety
Code Execution Design Documentation: Updated Docker configuration architecture
- Enhanced docs/dev_notes/CODE_EXECUTION_DESIGN.md (90 lines revised)
- New credential and package management configuration examples
- Architecture diagrams for nested configuration structures
Computer Use Tools Documentation: Clarified Docker usage requirements
- Updated massgen/tool/_computer_use/README.md and QUICKSTART.md
- Specified Docker requirements for Claude computer use
- Added troubleshooting guide for computer use setup

Technical Details#

Major Focus: Docker configuration improvements with nested structures for credentials and packages, framework interoperability streaming enhancements, parallel execution safety across all modes, contributor handbook
Contributors: @ncrispino @Eric-Shang @franklinnwren and the MassGen team

[0.1.9] - 2025-11-07#

Added#

Session Management System: Comprehensive session state tracking and restoration for multi-turn conversations
- New massgen/session/ module with session state and registry management (530 lines total)
- SessionState dataclass for complete session state including conversation history, workspace paths, and turn metadata (_state.py, 219 lines)
- SessionRegistry for listing, managing, and restoring previous sessions (_registry.py, 311 lines)
- restore_session() function for seamless session continuation across CLI invocations
- Session metadata tracking including winning agents history and orchestrator turn data
- Automatic session storage with unique identifiers and timestamps
- Test suite in test_session_registry.py (201 lines)
Computer Use Tools: Browser and desktop automation capabilities for multi-agent workflows
- General Computer Use Tool: OpenAI computer-use-preview integration for automated browser/computer control (massgen/tool/_computer_use/computer_use_tool.py, 741 lines)
  - Support for browser environment (Playwright) and Docker container execution
  - Action execution: click, type, scroll, navigate, screenshot analysis
  - Configurable max iterations and safety controls
- Claude Computer Use Tool: Anthropic Claude Computer Use API integration (massgen/tool/_claude_computer_use/claude_computer_use_tool.py, 473 lines)
  - Native Claude Computer Use beta API support
  - Browser and desktop control with safety confirmations
  - Async execution with Playwright integration
- Gemini Computer Use Tool: Google Gemini-based computer control (massgen/tool/_gemini_computer_use/gemini_computer_use_tool.py, 503 lines)
  - Gemini model integration for computer use workflows
  - Screenshot analysis and action generation
- Browser Automation Tool: Lightweight browser automation for specific tasks (massgen/tool/_browser_automation/browser_automation_tool.py, 176 lines)
  - Focused browser automation without full computer use overhead
- Comprehensive test suite in test_computer_use.py (629 lines)
OpenAI Operator API Handler: Support for OpenAI’s computer-use-preview model
- New massgen/api_params_handler/_openai_operator_api_params_handler.py (72 lines)
- Specialized parameter handling for computer use actions
- Integration with computer use tool execution flow

Changed#

Config Builder Enhancement: Intelligent model matching and discovery
- Fuzzy Model Name Matching: New massgen/utils/model_matcher.py (214 lines) allowing approximate model name input
- Model Catalog System: New massgen/utils/model_catalog.py (218 lines) with curated lists of common models across providers
- Enhanced massgen/config_builder.py with automatic model search and suggestions
- Support for partial model names with intelligent completion (e.g., “sonnet” → “claude-sonnet-4-5-20250929”)
- Contribution from acrobat3 (K. from JP)
Backend Capabilities Enhancement: Expanded provider support with six new backend registrations
- Added Cerebras AI backend capabilities (llama models with WSE hardware acceleration)
- Added Together AI backend capabilities (Meta-Llama, Mixtral models)
- Added Fireworks AI backend capabilities (Llama, Qwen models with fast inference)
- Added Groq backend capabilities (Llama, Mixtral with LPU hardware)
- Added OpenRouter backend capabilities (unified access to 200+ models with audio/video support)
- Added Moonshot (Kimi) backend capabilities (Chinese-optimized models with long context)
- Updated massgen/backend/capabilities.py with comprehensive backend specifications
Memory System Improvement: Enhanced memory update logic for multi-agent coordination
- New massgen/memory/_update_prompts.py (276 lines) with specialized update prompts for mem0
- MASSGEN_UNIVERSAL_UPDATE_MEMORY_PROMPT: Philosophy for accumulating qualitative patterns vs statistics
- Improved fact merging logic focusing on actionable tool usage patterns and technical insights
Chat Agent Enhancement: Session restoration and improved orchestrator restart handling
- Session state restoration in massgen/chat_agent.py
- Enhanced turn tracking and workspace persistence
- Improved logging and coordination with orchestrator restarts
CLI Enhancement: Extended command-line interface for session management
- Session listing and restoration commands in massgen/cli.py
- Enhanced display selection and output formatting
- Support for continuing previous sessions with automatic state restoration

Documentations, Configurations and Resources#

Diversity System Documentation: Comprehensive guide for increasing agent diversity
- New docs/source/user_guide/diversity.rst (388 lines)
- Covers answer novelty requirements (lenient/balanced/strict)
- Documents DSPy question paraphrasing integration (from v0.1.8)
- Best practices for multi-agent diversity strategies
- Configuration examples and recommendations
Memory System Documentation: Updated memory user guide
- Updated docs/source/user_guide/memory.rst with enhanced memory update logic and configuration
Computer Use Configuration Examples: Five YAML configurations demonstrating computer use capabilities
- massgen/configs/tools/custom_tools/claude_computer_use_example.yaml: Claude-specific computer use
- massgen/configs/tools/custom_tools/gemini_computer_use_example.yaml: Gemini-specific computer use
- massgen/configs/tools/custom_tools/computer_use_example.yaml: General computer use with OpenAI
- massgen/configs/tools/custom_tools/computer_use_docker_example.yaml: Docker-based computer use
- massgen/configs/tools/custom_tools/computer_use_browser_example.yaml: Browser automation focus
Session Management Configuration: Example demonstrating session continuation
- massgen/configs/memory/grok4_gpt5_gemini_mcp_filesystem_test_with_claude_code.yaml: Multi-turn session with MCP filesystem
Computer Use Documentation:
- New massgen/backend/docs/COMPUTER_USE_TOOLS_GUIDE.md: Comprehensive guide for computer use tools (494 lines)
- New scripts/computer_use_setup.md: Setup instructions for computer use tools
- New scripts/setup_docker_cua.sh: Automated Docker setup script for computer use

Technical Details#

Major Focus: Session management with conversation restoration, computer use automation tools, intelligent config builder with fuzzy matching, expanded backend support, memory system enhancements
Contributors: @franklinnwren @ncrispino @Henry-811 and the MassGen team

[0.1.8] - 2025-11-05#

Added#

Automation Mode for LLM Agents: Complete infrastructure for running MassGen via LLM agents and programmatic workflows
- New --automation CLI flag for silent execution with minimal output (~10 lines vs 250-3,000+)
- New SilentDisplay class in massgen/frontend/displays/silent_display.py for automation-friendly output
- Real-time status.json monitoring file updated every 2 seconds via enhanced CoordinationTracker
- Meaningful exit codes: 0 (success), 1 (config error), 2 (execution error), 3 (timeout), 4 (interrupted)
- Automatic workspace isolation for parallel execution with unique suffixes
- Meta-coordination capabilities: MassGen running MassGen configurations
- Automatic log directory creation and management for automation sessions
DSPy Question Paraphrasing Integration: Intelligent question diversity for multi-agent coordination
- New massgen/dspy_paraphraser.py module with semantic-preserving paraphrasing (557 lines)
- Three paraphrasing strategies: “diverse”, “balanced” (default), “conservative”
- Configurable number of variants per orchestrator session
- Automatic semantic validation using SemanticValidationSignature to ensure meaning preservation
- Thread-safe caching system with SHA-256 hashing for performance
- Support for all backends (Gemini, OpenAI, Claude, etc.) as paraphrasing engines
Case Study Summary: Comprehensive documentation of MassGen capabilities
- New docs/CASE_STUDIES_SUMMARY.md providing centralized overview of 33 case studies (368 lines)
- Organized by category: Release Features, Research, Travel, Creative, In Development, Planned
- Covers versions v0.0.3 to v0.1.5 with status tracking and links to videos
- Statistics: 19 completed, 8 with video demonstrations, 6 categories

Changed#

Orchestrator Enhancement: Integration of DSPy paraphrasing and automation tracking
- Question variant distribution to different agents based on configured strategy
- Improved coordination event logging with structured status exports
CLI Enhancement: Extended command-line interface for automation workflows
- Enhanced display selection logic automatically choosing SilentDisplay in automation mode
- Improved output formatting optimized for LLM agent parsing and monitoring

Documentations, Configurations and Resources#

Case Study: Meta-level self-analysis demonstrating automation mode
- New docs/source/examples/case_studies/meta-self-analysis-automation-mode.md: Comprehensive case study showing MassGen analyzing its own v0.1.8 codebase using automation mode
Automation Documentation: Comprehensive guides for LLM agent integration
- New AI_USAGE.md: Complete guide for LLM agents running MassGen (319 lines)
- New docs/source/user_guide/automation.rst: Full automation guide with BackgroundShellManager patterns (890 lines)
- New docs/source/reference/status_file.rst: Complete status.json schema reference with field-by-field documentation (565 lines)
- Updated README.md and README_PYPI.md with automation mode sections (135 lines each)
DSPy Documentation: Complete implementation and usage guide
- New massgen/backend/docs/DSPY_IMPLEMENTATION_GUIDE.md: Comprehensive DSPy integration guide (653 lines)
- Covers quick start, configuration, strategies, troubleshooting, and semantic validation
- Includes paraphrasing examples and best practices
Meta-Coordination Configurations: MassGen running MassGen examples
- massgen/configs/meta/massgen_runs_massgen.yaml: Single agent autonomously running MassGen experiments
- massgen/configs/meta/massgen_suggests_to_improve_massgen.yaml: Self-improvement configuration
- Demonstrates automation mode usage for meta-coordination workflows
DSPy Configuration Example: New YAML configuration for DSPy-enabled coordination
- massgen/configs/basic/multi/three_agents_dspy_enabled.yaml: Three-agent setup with DSPy paraphrasing
Case Study Summary Documentation: Centralized case study reference
- New docs/CASE_STUDIES_SUMMARY.md: Comprehensive overview of all MassGen case studies with categorization and status tracking

Technical Details#

Major Focus: Automation infrastructure for LLM agents, DSPy-powered question paraphrasing, meta-coordination capabilities, comprehensive case study documentation
Contributors: @ncrispino @praneeth999 @franklinnwren @qidanrui @sonichi @Henry-811 and the MassGen team

[0.1.7] - 2025-11-03#

Added#

Agent Task Planning System: MCP-based task management with dependency tracking
- New massgen/mcp_tools/planning/ module with dedicated planning server (_planning_mcp_server.py)
- Task dataclasses with dependency validation and status management (planning_dataclasses.py)
- Support for task states (pending/in_progress/completed/blocked) with automatic transitions based on dependencies
- Orchestrator integration for plan-aware coordination
- Test suite in test_planning_integration.py and test_planning_tools.py
Background Shell Execution: Long-running command support with persistent sessions
- New BackgroundShell class in massgen/filesystem_manager/background_shell.py
- Shell lifecycle management with output streaming and real-time monitoring
- Automatic timeout handling for long-running processes
- Enhanced code execution server with background execution capabilities
- Test coverage in test_background_shell.py
Preemption Coordination: Multi-agent coordination with interruption support
- Agents can preempt ongoing coordination to submit better answers without full restart
- Enhanced coordination tracker with preemption event logging
- Improved orchestrator logic to preserve partial progress during preemption

Fixed#

System Message Handling: Resolved system message extraction in Claude Code backend for background shell execution
Case Study Documentation: Fixed broken links and outdated examples in older case studies

Documentations, Configurations and Resources#

Documentation Updates: New user guides and design documentation
- New docs/source/user_guide/agent_task_planning.rst: Task planning guide with usage patterns and API reference
- Updated docs/source/user_guide/code_execution.rst: Added 122 lines for background shell usage
- New docs/dev_notes/agent_planning_coordination_design.md: Comprehensive design document for agent planning and coordination system
- New docs/dev_notes/preempt_not_restart_design.md: 456-line design document with preemption algorithms
- Updated docs/source/development/architecture.rst: Added 61 lines for preemption coordination architecture
Configuration Examples: New YAML configurations demonstrating v0.1.7 features
- example_task_todo.yaml: Task planning configuration
- background_shell_demo.yaml: Background shell execution demonstration

Technical Details#

Major Focus: Agent task planning with dependencies, background command execution, preemption-based coordination
Contributors: @ncrispino @Henry-811 and the MassGen team

[0.1.6] - 2025-10-31#

Added#

Framework Interoperability: External agent framework integration as MassGen custom tools
- New massgen/tool/_extraframework_agents/ module with 5 framework integrations
- AG2 Lesson Planner Tool: Nested chat functionality wrapped as custom tool for multi-agent lesson planning (supports streaming)
- LangGraph Lesson Planner Tool: LangGraph graph-based workflows integrated as tool
- AgentScope Lesson Planner Tool: AgentScope agent system wrapped for lesson creation
- OpenAI Assistants Lesson Planner Tool: OpenAI Assistants API integrated as tool
- SmoLAgent Lesson Planner Tool: HuggingFace SmoLAgent integration for lesson planning
- Enables MassGen agents to delegate tasks to specialized external frameworks
- Each framework runs autonomously and returns results to MassGen orchestrator
- Note: Only AG2 currently supports streaming; other frameworks return complete results
Configuration Validator: Comprehensive YAML configuration validation system
- New ConfigValidator class in massgen/config_validator.py for pre-flight validation
- Memory configuration validation with detailed error messages
- Pre-commit hook integration for automatic config validation
- Comprehensive test suite in massgen/tests/test_config_validator.py
- Validates agent configurations, backend parameters, tool settings, and memory options
- Provides actionable error messages with suggestions for common mistakes

Changed#

Backend Architecture Refactoring: Unified tool execution with ToolExecutionConfig
- New ToolExecutionConfig dataclass in base_with_custom_tool_and_mcp.py for standardized tool handling
- Refactored ResponseBackend with unified tool execution flow
- Refactored ChatCompletionsBackend with unified tool execution flow
- Refactored ClaudeBackend with unified tool execution methods
- Eliminates duplicate code paths between custom tools and MCP tools
- Consistent error handling and status reporting across all tool types
- Improved maintainability and extensibility for future tool systems
Gemini Backend Simplification: Major architectural cleanup and consolidation
- Removed gemini_mcp_manager.py module
- Removed gemini_trackers.py module
- Refactored gemini.py to use manual tool execution via base class
- Streamlined tool handling and cleanup logic
- Removed continuation logic and duplicate code
- Updated _gemini_formatter.py for simplified tool conversion
- Net reduction of 1,598 lines through consolidation
- Improved maintainability and performance
Custom Tool System Enhancement: Improved tool management and execution
- Enhanced ToolManager with category management capabilities
- Improved tool registration and validation system
- Enhanced tool result handling and error reporting
- Better support for async tool execution
- Improved tool schema generation for LLM consumption

Documentations, Configurations and Resources#

Framework Interoperability Examples: 8 new configuration files demonstrating external framework integration
- AG2 Examples: ag2_lesson_planner_example.yaml, ag2_and_langgraph_lesson_planner.yaml, ag2_and_openai_assistant_lesson_planner.yaml
- LangGraph Examples: langgraph_lesson_planner_example.yaml
- AgentScope Examples: agentscope_lesson_planner_example.yaml
- OpenAI Assistants Examples: openai_assistant_lesson_planner_example.yaml
- SmoLAgent Examples: smolagent_lesson_planner_example.yaml
- Multi-Framework Examples: two_models_with_tools_example.yaml

Technical Details#

Major Focus: Framework interoperability for external agent integration, unified tool execution architecture, Gemini backend simplification, and configuration validation system
Contributors: @Eric-Shang @praneeth999 @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.1.5] - 2025-10-29#

Added#

Memory System: Complete long-term memory implementation with semantic retrieval
- New massgen/memory/ module with comprehensive memory management
- PersistentMemory via mem0 integration for semantic fact storage and retrieval
- ConversationMemory for short-term verbatim message tracking
- Automatic Context Compression when approaching token limits
- Memory Sharing for Multi-Turn Conversations with turn-aware filtering to prevent temporal leakage
- Session Management for memory isolation and continuation across runs
- Qdrant Vector Database Integration for efficient semantic search (server and local modes)
- Context Monitoring with real-time token usage tracking
- Fact extraction prompts with customizable LLM and embedding providers
- Supports OpenAI, Anthropic, Groq, and other mem0-compatible providers
Memory Configuration Support: New YAML configuration options
- Memory enable/disable toggle at global and per-agent levels
- Configurable compression thresholds (trigger_threshold, target_ratio)
- Retrieval settings (limit, exclude_recent for smart retrieval)
- Session naming for continuation and cross-session memory
- LLM and embedding provider configuration for mem0
- Qdrant connection settings (server/local mode, host, port, path)

Changed#

Chat Agent Enhancement: Memory integration for agent workflows
- Memory recording after agent responses (conversation and persistent)
- Memory retrieval on restart/reset for context restoration
- Integration with compression and context monitoring modules
Orchestrator Enhancement: Memory coordination for multi-agent workflows
- Memory initialization and management across agent lifecycles
- Memory cleanup on orchestrator shutdown

Documentations, Configurations and Resources#

Memory Documentation: Comprehensive memory system user guide
- New docs/source/user_guide/memory.rst
- Complete usage guide with quick start, configuration reference, and examples
- Design decisions documentation explaining architecture choices
- Troubleshooting guide for common memory issues
- Monitoring and debugging instructions with log examples
- API reference for PersistentMemory, ConversationMemory, and ContextMonitor
Configuration Examples: 5 new memory-focused YAML configurations
- gpt5mini_gemini_context_window_management.yaml: Multi-agent with context compression
- gpt5mini_gemini_research_to_implementation.yaml: Research to implementation workflow
- gpt5mini_high_reasoning_gemini.yaml: High reasoning agents with memory
- gpt5mini_gemini_baseline_research_to_implementation.yaml: Baseline research workflow
- single_agent_compression_test.yaml: Testing compression behavior
Infrastructure and Testing:
- Memory test suite with 4 test files in massgen/tests/memory/
- Additional memory tests: test_agent_memory.py, test_conversation_memory.py, test_orchestrator_memory.py, test_persistent_memory.py

Technical Details#

Major Focus: Long-term memory system with semantic retrieval and memory sharing for multi-turn conversations
Contributors: @ncrispino @qidanrui @kitrakrev @sonichi @Henry-811 and the MassGen team

[0.1.4] - 2025-10-27#

Added#

Multimodal Generation Tools: Comprehensive generation capabilities via OpenAI APIs
- New text_to_image_generation tool for generating images from text prompts using DALL-E models
- New text_to_video_generation tool for generating videos from text prompts
- New text_to_speech_continue_generation tool for text-to-speech with continuation support
- New text_to_speech_transcription_generation tool for audio transcription and generation
- New text_to_file_generation tool for generating documents (PDF, DOCX, XLSX, PPTX)
- New image_to_image_generation tool for image-to-image transformations
- Implemented in massgen/tool/_multimodal_tools/ with 6 new modules
Binary File Protection System: Enhanced security for file operations
- New binary file blocking in PathPermissionManager preventing text tools from reading binary files
- Added BINARY_FILE_EXTENSIONS set covering images, videos, audio, archives, executables, and Office documents
- New _validate_binary_file_access() method with intelligent tool suggestions
- Prevents context pollution by blocking Read, read_text_file, and read_file tools from binary files
- Comprehensive test suite in test_binary_file_blocking.py
Crawl4AI Web Scraping Integration: Advanced web content extraction tool
- New crawl4ai_tool for intelligent web scraping with LLM-powered extraction
- Implemented in massgen/tool/_web_tools/crawl4ai_tool.py

Changed#

Multimodal File Size Limits: Enhanced validation and automatic handling
- Automatic image resizing for files exceeding size limits
- Comprehensive size limit test suite in test_multimodal_size_limits.py
- Enhanced validation in understand_audio and understand_video tools

Documentations, Configurations and Resources#

PyPI Package Documentation: Standalone README for PyPI distribution
- New README_PYPI.md with comprehensive package documentation
- Improved package metadata and installation instructions
Release Management Documentation: Comprehensive release workflow guide
- New docs/dev_notes/release_checklist.md with step-by-step release procedures
- Detailed checklist for testing, documentation, and deployment
Binary File Protection Documentation: Enhanced protected paths user guide
- Updated docs/source/user_guide/protected_paths.rst with binary file protection section
- Documents 40+ protected binary file types and specialized tool suggestions
Configuration Examples: 9 new YAML configuration files
- Generation Tools: 8 multimodal generation configurations
  - text_to_image_generation_single.yaml and text_to_image_generation_multi.yaml
  - text_to_video_generation_single.yaml and text_to_video_generation_multi.yaml
  - text_to_speech_generation_single.yaml and text_to_speech_generation_multi.yaml
  - text_to_file_generation_single.yaml and text_to_file_generation_multi.yaml
- Web Scraping: crawl4ai_example.yaml for Crawl4AI integration

Technical Details#

Major Focus: Multimodal generation tools, binary file protection system, web scraping integration
Contributors: @qidanrui @ncrispino @sonichi @Henry-811 and the MassGen team

[0.1.3] - 2025-10-24#

Added#

Post-Evaluation Workflow Tools: Submit and restart capabilities for winning agents
- New PostEvaluationToolkit class in massgen/tool/workflow_toolkits/post_evaluation.py
- submit tool for confirming final answers
- restart_orchestration tool for restarting with improvements and feedback
- Post-evaluation phase where winning agent evaluates its own answer
- Support for all API formats (Claude, Response API, Chat Completions)
- Configuration parameter enable_post_evaluation_tools for opt-in/out
Custom Multimodal Understanding Tools: Active tools for analyzing workspace files using OpenAI’s GPT-4.1 API
- New understand_image tool for analyzing images (PNG, JPEG, JPG) with detailed metadata extraction
- New understand_audio tool for transcribing and analyzing audio files (WAV, MP3, FLAC, OGG)
- New understand_video tool for extracting frames and analyzing video content (MP4, AVI, MOV, WEBM)
- New understand_file tool for processing documents (PDF, DOCX, XLSX, PPTX) with text and metadata extraction
- Works with any backend (uses OpenAI for analysis)
- Returns structured JSON with comprehensive metadata
Docker Sudo Mode: Enhanced Docker execution with privileged command support
- New use_sudo parameter for Docker execution
- Sudo mode for commands requiring elevated privileges
- Enhanced security instructions and documentation
- Test coverage in test_code_execution.py

Changed#

Interactive Config Builder Enhancement: Improved workflow and provider handling
- Better flow from automatic setup to config builder
- Auto-detection of environment variables
- Improved provider-specific configuration handling
- Integrated multimodal tools selection in config wizard

Fixed#

System Message Warning: Resolved deprecated system message configuration warning
- Fixed system message handling in agent_config.py
- Updated chat agent to properly handle system messages
- Removed deprecated warning messages
Config Builder Issues: Multiple configuration builder improvements
- Fixed config display errors
- Improved config saving across different provider types
- Better error handling for missing configurations

Documentations, Configurations and Resources#

Multimodal Tools Documentation: Comprehensive documentation for new multimodal tools
- docs/source/user_guide/multimodal.rst: Updated with custom tools section
- massgen/tool/docs/multimodal_tools.md: Complete 779-line technical documentation
Docker Sudo Mode Documentation: Enhanced Docker execution documentation
- docs/source/user_guide/code_execution.rst: Added 98 lines documenting sudo mode
- massgen/docker/README.md: Updated with sudo mode instructions
Configuration Examples: New example configurations
- configs/tools/multimodal_tools/understand_image.yaml: Image analysis configuration
- configs/tools/multimodal_tools/understand_audio.yaml: Audio transcription configuration
- configs/tools/multimodal_tools/understand_video.yaml: Video analysis configuration
- configs/tools/multimodal_tools/understand_file.yaml: Document processing configuration
Example Resources: New test resources for v0.1.3 features
- massgen/configs/resources/v0.1.3-example/multimodality.jpg: Image example
- massgen/configs/resources/v0.1.3-example/Sherlock_Holmes.mp3: Audio example
- massgen/configs/resources/v0.1.3-example/oppenheimer_trailer_1920.mp4: Video example
- massgen/configs/resources/v0.1.3-example/TUMIX.pdf: PDF document example
Case Studies: New case study demonstrating v0.1.3 features
- docs/source/examples/case_studies/multimodal-case-study-video-analysis.md: Meta-level demonstration of multimodal video understanding with agents analyzing their own case study videos

Technical Details#

Major Focus: Post-evaluation workflow tools, custom multimodal understanding tools, Docker sudo mode
Contributors: @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.1.2] - 2025-10-22#

Added#

Claude 4.5 Haiku Support: Added latest Claude Haiku model
- New model: claude-haiku-4-5-20251001
- Updated model registry in backend/capabilities.py

Changed#

Planning Mode Enhancement: Intelligent automatic MCP tool blocking based on operation safety
- New _analyze_question_irreversibility() method in orchestrator analyzes questions to determine if MCP operations are reversible
- New set_planning_mode_blocked_tools(), get_planning_mode_blocked_tools(), and is_mcp_tool_blocked() methods in backend for selective tool control
- Dynamically enables/disables planning mode - read-only operations allowed during coordination, write operations blocked
- Planning mode supports different workspaces without conflicts
- Zero configuration required - works transparently
Claude Model Priority: Reorganized model list in capabilities registry
- Changed default model from claude-sonnet-4-20250514 to claude-sonnet-4-5-20250929
- Moved claude-opus-4-1-20250805 higher in priority order
- Updated in both Claude and Claude Code backends

Fixed#

Grok Web Search: Resolved web search functionality in Grok backend
- Fixed extra_body parameter handling for Grok’s Live Search API
- New _add_grok_search_params() method for proper search parameter injection
- Enhanced _stream_with_custom_and_mcp_tools() to support Grok-specific parameters
- Improved error handling for conflicting search configurations
- Better integration with Chat Completions API params handler

Documentations, Configurations and Resources#

Intelligent Planning Mode Case Study: Complete feature documentation
- docs/source/examples/case_studies/INTELLIGENT_PLANNING_MODE.md: Comprehensive guide for automatic planning mode
- Demonstrates automatic irreversibility detection
- Shows read/write operation classification
- Includes examples for Discord, filesystem, and Twitter operations
Configuration Updates: Enhanced YAML examples
- Updated 5 planning mode configurations in configs/tools/planning/ with selective blocking examples
- Updated three_agents_default.yaml with Grok-4-fast model
- Test coverage in test_intelligent_planning_mode.py

Technical Details#

Major Focus: Intelligent planning mode with selective tool blocking, model support enhancements
Contributors: @franklinnwren @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.1.1] - 2025-10-20#

Added#

Custom Tools System: Complete framework for registering and executing user-defined Python functions as tools
- New ToolManager class in massgen/tool/_manager.py for centralized tool registration and lifecycle management
- Support for custom tools alongside MCP servers across all backends (Claude, Gemini, OpenAI Response API, Chat Completions, Claude Code)
- Three tool categories: builtin, mcp, and custom tools
- Automatic tool discovery with name prefixing and conflict resolution
- Tool validation with parameter schema enforcement
- Comprehensive test coverage in test_custom_tools.py
Voting Sensitivity & Answer Novelty Controls: Three-tier system for multi-agent coordination
- New voting_sensitivity parameter with three levels: “lenient”, “balanced”, “strict”
- “Lenient”: Accepts any reasonable answer
- “Balanced”: Default middle ground
- “Strict”: High-quality requirement
- Answer novelty detection with _check_answer_novelty() method in orchestrator.py preventing duplicate answers
- Configurable max_new_answers_per_agent limiting submissions per agent
- Token-based similarity thresholds (50-70% overlap) for duplicate detection
Interactive Configuration Builder: Wizard for creating YAML configurations
- New config_builder.py module with step-by-step prompts
- Guided workflow for backend selection, model configuration, and API key setup
- Model-specific parameter handling (temperature, reasoning, verbosity)
- Tool enablement options (MCP servers, custom tools, builtin tools)
- Configuration validation and preview before saving
- Integration with massgen --config-builder command
Backend Capabilities Registry: Centralized feature support tracking
- New capabilities.py module in massgen/backend/ documenting backend capabilities
- Feature matrix showing MCP, custom tools, multimodal, and code execution support
- Runtime capability queries for backend selection

Changed#

Gemini Backend Architecture: Major refactoring for improved maintainability
- Extracted MCP management into gemini_mcp_manager.py
- Extracted tracking logic into gemini_trackers.py
- Extracted utilities into gemini_utils.py
- New API params handler _gemini_api_params_handler.py
- Improved session management and tool execution flow
Python Version Requirements: Updated minimum supported version
- Changed from Python 3.10+ to Python 3.11+ in pyproject.toml
- Ensures compatibility with modern type hints and async features
API Key Setup Command: Simplified command name
- Renamed massgen --setup-keys to massgen --setup for brevity
- Maintained all functionality for interactive API key configuration
Configuration Examples: Updated example commands
- Changed from python -m massgen.cli to simplified massgen command
- Updated 40+ configuration files for consistency

Fixed#

CLI Configuration Selection: Resolved error with large config lists
- Fixed crash when using massgen --select with many available configurations
- Improved pagination and display of configuration options
- Enhanced error handling for configuration discovery
CLI Help System: Improved documentation display
- Fixed help text formatting in massgen --help
- Better organization of command options and examples

Documentations, Configurations and Resources#

Case Study: Universal Code Execution via MCP: Comprehensive v0.0.31 feature documentation
- docs/source/examples/case_studies/universal-code-execution-mcp.md
- Demonstrates pytest test creation and execution across backends
- Shows command validation, security layers, and result interpretation
Documentation Updates: Enhanced existing documentation
- Added custom tools user guide and integration examples
- Reorganized case studies for improved navigation
- Updated configuration schema with new voting and tools parameters
Custom Tools Examples: 40+ example configurations
- Basic single-tool setups for each backend
- Multi-agent configurations with custom tools
- Integration examples combining MCP and custom tools
- Located in configs/tools/custom_tools/
Voting Sensitivity Examples: Configuration examples for voting controls
- configs/voting/gemini_gpt_voting_sensitivity.yaml
- Demonstrates lenient, balanced, and strict voting modes
- Shows answer novelty threshold configuration

Technical Details#

Major Focus: Custom tools system, voting sensitivity controls, interactive config builder, and comprehensive documentation
Contributors: @qidanrui @ncrispino @praneeth999 @sonichi @Eric-Shang @Henry-811 and the MassGen team

[0.1.0] - 2025-10-17 (PyPI Release)#

Added#

PyPI Package Release: Official MassGen package available on PyPI for easy installation via pip
Enhanced Documentation: Comprehensive Sphinx documentation with improved structure and clarity
- Rebuilt documentation with v0.1.0 version numbers
- Improved backend capabilities table with split multimodal columns
- Enhanced explanations for multimodal capabilities (Both, Understanding, Generation)
- Updated homepage with v0.1.0 features

Changed#

Documentation Updates: Major documentation improvements for PyPI release
- Updated version numbers across all documentation files
- Clarified multimodal capability terminology
- Enhanced backend configuration guides

Technical Details#

Major Focus: PyPI distribution and documentation improvements
Contributors: @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.32] - 2025-10-15#

Added#

Docker Execution Mode: Isolated command execution via Docker containers
- New DockerManager class for persistent container lifecycle management
- Container-based isolation with volume mounts for workspace and context paths
- Configurable resource limits (CPU, memory) and network isolation modes (none/bridge/host)
- Multi-agent support with dedicated containers per agent
- Build script and comprehensive Dockerfile for massgen/mcp-runtime image
- Enable via command_line_execution_mode: "docker" in agent configuration
- Test suite in test_code_execution.py covering Docker and local execution modes

Changed#

Code Execution via MCP: Extended v0.0.31’s execute_command tool with Docker execution mode
- Docker environment detection for automatic image verification
- Local command execution remains available via command_line_execution_mode: "local"
- Enhanced security layers for both local and Docker modes
Claude Code Backend: Docker mode integration and MCP tool handling improvements
- Automatic Bash tool disablement when Docker mode is enabled
- MCP tool auto-permission support via can_use_tool hook
- MCP server configuration format conversion (list to dict format)
- System message enhancements to prevent git repository confusion in Docker
MCP Tools Architecture: Major refactoring for simplicity and maintainability
- Renamed MultiMCPClient to MCPClient reflecting simplified architecture
- Removed deprecated converters.py module (275 lines removed)
- Streamlined client.py with 1,029 lines removed through consolidation
- Standardized type hints and module-level constants in backend_utils.py
- Simplified exception handling in exceptions.py and security validation in security.py

Fixed#

Configuration Examples: Improved configuration organization and usability
- Renamed configuration files for better discoverability
- Fixed CPU limits in example configurations to be runnable
- Reverted gemini_mcp_test.yaml for consistency
Orchestrator Timeout and Cleanup: Enhanced timeout handling and resource management
- Improved timeout mechanisms for better reliability
- Better cleanup of resources after orchestration sessions

Documentations, Configurations and Resources#

Docker Documentation: New comprehensive Docker mode guide in massgen/docker/README.md
- Complete Docker setup and usage documentation
- Build scripts and Dockerfile with detailed comments
- Security considerations for container-based execution
- Resource management and isolation strategies
Code Execution Design: Updated CODE_EXECUTION_DESIGN.md with Docker architecture details
New Configuration Files: Added 5 Docker-specific example configurations
- docker_simple.yaml: Basic single-agent Docker execution
- docker_multi_agent.yaml: Multi-agent Docker deployment
- docker_with_resource_limits.yaml: Resource-constrained Docker setup
- docker_claude_code.yaml: Claude Code with Docker execution
- docker_verification.yaml: Docker setup verification configuration

Technical Details#

Commits: 17 commits including Docker execution, MCP refactoring, and Claude Code enhancements
Files Modified: 32 files across backend, filesystem manager, MCP tools, and configurations
Major Features: Docker execution mode, MCP architecture simplification, Claude Code Docker integration
New Module: _docker_manager.py with DockerManager class (438 lines)
Dependencies Updated: docker>=7.0.0 added as optional dependency
Contributors: @ncrispino @praneeth999 @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.31] - 2025-10-14#

Added#

Code Execution via MCP: Universal command execution through MCP
- New execute_command MCP tool enabling bash/shell execution across Claude, Gemini, OpenAI (Response API), and Chat Completions providers (Grok, ZAI, etc.)
- AG2-inspired security with multi-layer protection: dangerous command sanitization, command filtering (whitelist/blacklist), PathPermissionManager hooks, path validation, timeout enforcement
- Command filtering with regex patterns for whitelist/blacklist control
- New MCP server _code_execution_server.py with subprocess-based local execution
- Test coverage in test_code_execution.py covering basics, path validation, command sanitization, output handling, and virtual environment detection
Audio Generation Tools: Text-to-speech and audio transcription capabilities via OpenAI APIs
- New generate_and_store_audio_no_input_audios tool for generating audio from text using gpt-4o-audio-preview model
- New generate_text_with_input_audio tool for transcribing audio files using OpenAI’s Transcription API
- New convert_text_to_speech tool for converting text to speech with gpt-4o-mini-tts model
- Support for multiple voices (alloy, echo, fable, onyx, nova, shimmer, coral, sage) and audio formats (wav, mp3, opus, aac, flac)
- Optional speaking instructions for tone and style control in TTS
- Automatic workspace organization with timestamp-based filenames
Video Generation Tools: Text-to-video generation via OpenAI’s Sora-2 API
- New generate_and_store_video_no_input_images tool for generating videos from text prompts
- Support for Sora-2 model with configurable video duration
- Asynchronous video generation with progress monitoring
- Automatic MP4 format with workspace storage and organization

Changed#

AG2 Group Chat Support: Enhanced AG2 adapter with native multi-agent group chat coordination
- New group chat manager integration with AG2’s GroupChat and GroupChatManager
- Configurable speaker selection modes: auto (LLM-based), round_robin, manual
- Support for nested conversations and workflow tools within group chat sessions
- Automatic tool registration/unregistration for clean group chat lifecycle
- Enhanced adapter architecture with group chat state management
- Better agent reinitialization and termination logic for multi-turn group conversations
- Test coverage in test_ag2_adapter.py and test_ag2_utils.py
File Operation Tracker: Enhanced with auto-generated file exemptions
- New _is_auto_generated() method to identify build artifacts and cache files
- Prevents permission errors when agents clean up after running tests or builds
Path Permission Manager: Added execute_command tool validation
- Added execute_command to command_tools set for bash-like security validation
- PreToolUse hooks now validate execute_command calls for dangerous patterns and path restrictions
- Enhanced test coverage with 93 new test lines for command tool validation
Message Templates: Added code execution result guidance
- New system message guidance when enable_command_execution=True instructing agents to explain test results and command outputs in their answers
- Better agent behavior for explaining what was tested and what results mean

Documentations, Configurations and Resources#

Code Execution Design Documentation: Comprehensive technical design document
- CODE_EXECUTION_DESIGN.md: Design doc covering architecture, security layers, implementation plan, virtual environment support, and future Docker enhancements
New Configuration Files: Added 8 new example configurations
- AG2 Group Chat: ag2_groupchat.yaml, ag2_groupchat_gpt.yaml
- Code Execution: basic_command_execution.yaml, code_execution_use_case_simple.yaml, command_filtering_whitelist.yaml, command_filtering_blacklist.yaml,
- Audio Generation: single_gpt4o_audio_generation.yaml, gpt4o_audio_generation.yaml
- Video Generation: single_gpt4o_video_generation.yaml

Technical Details#

Commits: 29 commits including AG2 group chat, code execution, audio/video generation, and enhancements
Files Modified: 39 files with 3,649 insertions and 154 deletions
Major Features: AG2 group chat, universal code execution via MCP, audio/video generation tools
New Tests: test_ag2_adapter.py, test_ag2_utils.py, test_code_execution.py
Contributors: @Eric-Shang @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.30] - 2025-10-10#

Changed#

Multimodal Support - Audio and Video Processing: Extended v0.0.27’s image-only multimodal foundation
- Audio file support with WAV and MP3 formats for Chat Completions and Claude backends
- Video file support with MP4, AVI, MOV, WEBM formats for Chat Completions and Claude backends
- Audio/video path parameters (audio_path, video_path) for local files and HTTP/HTTPS URLs
- Base64 encoding for local audio/video files with automatic MIME type detection
- Configurable media file size limits (default 64MB, configurable via media_max_file_size_mb)
- New audio/video content formatters in _chat_completions_formatter.py and _claude_formatter.py
- Enhanced base_with_mcp.py with 340+ lines of multimodal content processing
Claude Code Backend SDK Update: Updated to newer Agent SDK package
- Migrated from claude-code-sdk>=0.0.19 to claude-agent-sdk>=0.0.22
- Updated internal SDK classes: ClaudeCodeOptions → ClaudeAgentOptions
- Enhanced bash tool permission validation in PathPermissionManager
- Improved system message handling with SDK preset support
- New bash/shell/exec tool detection for dangerous operation prevention
Chat Completions Backend Enhancement: Qwen API provider integration
- Added Qwen API support to existing Chat Completions provider ecosystem
- New QWEN_API_KEY environment variable support
- Qwen-specific configuration examples for video understanding

Fixed#

Planning Mode Configuration: Fixed crash when configuration lacks coordination_config
- Added null check in orchestrator.py to prevent AttributeError
- Improved graceful handling of missing planning mode configuration
Claude Code System Message Handling: Resolved system message processing issues
- Fixed system message extraction and formatting in claude_code.py
- Better integration with Agent SDK for message handling
AG2 Adapter Import Ordering: Resolved import sequence issues
- Fixed import statements in adapters/utils/ag2_utils.py
- Pre-commit isort formatting corrections

Documentations, Configurations and Resources#

Case Studies: Comprehensive documentation for v0.0.28 and v0.0.29 features
- ag2-framework-integration.md: AG2 adapter system and external framework integration
- mcp-planning-mode.md: MCP Planning Mode design and implementation guide
New Configuration Files: Added 7 new example configurations
- ag2/ag2_case_study.yaml: AG2 framework integration case study configuration
- filesystem/cc_gpt5_gemini_filesystem.yaml: Claude Code, GPT-5, and Gemini filesystem collaboration
- basic/single/single_gemini2.5pro.yaml: Gemini 2.5 Pro single agent setup
- basic/single/single_openrouter_audio_understanding.yaml: Audio understanding with OpenRouter
- basic/single/single_qwen_video_understanding.yaml: Video understanding with Qwen API
- debug/test_sdk_migration.yaml: Claude Code SDK migration testing

Technical Details#

Commits: 20 commits including multimodal enhancements, Claude Code SDK migration, and documentation
Files Modified: 25 files with 2,501 insertions and 84 deletions
Major Features: Audio/video multimodal support, Claude Code Agent SDK migration, Qwen API integration
Dependencies Updated: anthropic>=0.61.0, claudecode>=0.0.12
Contributors: @ncrispino @praneeth999 @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.29] - 2025-10-08#

Added#

MCP Planning Mode: New coordination strategy for irreversible MCP actions
- New CoordinationConfig class with enable_planning_mode flag
- Agents plan without executing during coordination, winning agent executes during final presentation
- Orchestrator and frontend coordination UI support
- Support for multiple backends: Response API, Chat Completions, and Gemini
- Test suites in test_mcp_blocking.py and test_gemini_planning_mode.py
File Operation Tracker: Read-before-delete enforcement for safer file operations
- New FileOperationTracker class in filesystem_manager/_file_operation_tracker.py
- Prevents agents from deleting files they haven’t read first
- Tracks read files and agent-created files (created files exempt from read requirement)
- Directory deletion validation with comprehensive error messages
Path Permission Manager Enhancements: Integration with FileOperationTracker
- Added read/write/delete operation tracking methods to PathPermissionManager
- Integration with FileOperationTracker for read-before-delete enforcement
- Enhanced delete validation for files and batch operations
- Extended test coverage in test_path_permission_manager.py

Changed#

Message Templates: Improved multi-agent coordination guidance
- Added has_irreversible_actions support for context path write access
- Explicit temporary workspace path structure display for better agent understanding
- Task handling priority hierarchy and simplified new_answer requirements
- Unified evaluation guidance
MCP Tool Filtering: Enhanced multi-level filtering capabilities
- Combined backend-level and per-MCP-server tool filtering
- MCP-server-specific allowed_tools can override backend-level settings
- Merged exclude_tools from both backend and MCP server configurations
Backend Planning Mode Support: Extended planning mode to multiple backends
- Enhanced base.py, response.py, chat_completions.py, and gemini.py
- Gemini backend now supports planning mode with session-based tool execution
- Planning mode support across all major backend types

Fixed#

Circuit Breaker Logic: Enhanced MCP server initialization in base_with_mcp.py
Final Answer Context: Improved workspace copying when no new answer is provided
Multi-turn MCP Usage: Addressed non-use of MCP in certain scenarios and improved final answer autonomy
Configuration Issues: Updated Playwright automation configuration and fixed agent IDs

Documentations, Configurations and Resources#

MCP Planning Mode Examples: 5 new planning mode configurations in tools/planning/
- five_agents_discord_mcp_planning_mode.yaml: Discord MCP with planning mode (5 agents)
- five_agents_filesystem_mcp_planning_mode.yaml: Filesystem MCP with planning mode
- five_agents_notion_mcp_planning_mode.yaml: Notion MCP with planning mode (5 agents)
- five_agents_twitter_mcp_planning_mode.yaml: Twitter MCP with planning mode (5 agents)
- gpt5_mini_case_study_mcp_planning_mode.yaml: Case study configuration
MCP Example Configurations: New example configurations for MCP integration in tools/mcp/
- five_agents_travel_mcp_test.yaml: Travel planning MCP example (5 agents)
- five_agents_weather_mcp_test.yaml: Weather service MCP example (5 agents)
Debug Configurations: New debugging and testing utilities
- skip_coordination_test.yaml: Test configuration for skipping coordination rounds
Documentation Updates: Enhanced project documentation
- Updated permissions_and_context_files.md in backend/docs/ with file operation tracking details
- Updated README with AG2 as optional installation and uv tool instructions

Technical Details#

Commits: 23+ commits including planning mode, file operation tracking, and MCP enhancements
Files Modified: 43 files across agent config, backend, filesystem manager, MCP tools, and configurations
Major Features: MCP planning mode, FileOperationTracker, enhanced permissions, MCP tool filtering
New Tests: test_mcp_blocking.py, test_gemini_planning_mode.py for planning mode validation
Contributors: @ncrispino @franklinnwren @qidanrui @sonichi @praneeth999 and the MassGen team

[0.0.28] - 2025-10-06#

Added#

AG2 Framework Integration: Complete adapter system for external agent frameworks
- New massgen/adapters/ module with base adapter architecture (base.py, ag2_adapter.py)
- Support for AG2 ConversableAgent and AssistantAgent types
- Code execution capabilities with multiple executor types: LocalCommandLineCodeExecutor, DockerCommandLineCodeExecutor, JupyterCodeExecutor, YepCodeCodeExecutor
- Function/tool calling support for AG2 agents
- Async execution with a_generate_reply for autonomous operation
- AG2 utilities module for agent setup and API key management (adapters/utils/ag2_utils.py)
External Agent Backend: New backend type for integrating external frameworks
- New ExternalAgentBackend class supporting adapter registry pattern
- Bridge between MassGen orchestration and external agent frameworks via adapters
- Framework-specific configuration extraction and validation
- Currently supports AG2 with extensible architecture for future frameworks
AG2 Test Suite: Comprehensive test coverage for AG2 integration
- test_ag2_adapter.py: AG2 adapter functionality tests
- test_agent_adapter.py: Base adapter interface tests
- test_external_agent_backend.py: External backend integration tests

Fixed#

MCP Circuit Breaker Logic: Enhanced initialization for MCP servers
- Improved circuit breaker state management in base_with_mcp.py
- Better error handling during MCP server initialization

Documentations, Configurations and Resources#

AG2 Configuration Examples: New YAML configurations demonstrating AG2 integration
- ag2/ag2_single_agent.yaml: Basic single AG2 agent setup
- ag2/ag2_coder.yaml: AG2 agent with code execution
- ag2/ag2_coder_case_study.yaml: Multi-agent setup with AG2 and Gemini
- ag2/ag2_gemini.yaml: AG2-Gemini hybrid configuration
Design Documentation: Enhanced multi-source agent integration design
- Updated MULTI_SOURCE_AGENT_INTEGRATION_DESIGN.md with AG2 adapter architecture

Technical Details#

Commits: 12 commits including AG2 integration, testing, and configuration examples
Files Modified: 18 files with 1,423 insertions and 71 deletions
Major Features: AG2 framework integration, external agent backend, adapter architecture
New Module: massgen/adapters/ with AG2 support
Contributors: @Eric-Shang @praneeth999 @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.27] - 2025-10-03#

Added#

Multimodal Support - Image Processing: Foundation for multimodal content processing
- New stream_chunk module with base classes for multimodal content (base.py, text.py, multimodal.py)
- Support for image input and output in conversation messages
- Image generation and understanding capabilities for multi-agent workflows
- Multimodal content structure supporting images, audio, video, and documents (architecture ready)
File Upload and File Search: Extended backend capabilities for document operations
- File upload support integrated into Response backend via _response_api_params_handler.py
- File search functionality for enhanced context retrieval and Q&A
- Vector store management for file search operations
- Cleanup utilities for uploaded files and vector stores
Workspace Tools Enhancements: Extended MCP-based workspace management
- Added read_multimodal_files tool for reading images as base64 data with MIME type
Claude Sonnet 4.5 Support: Added latest Claude model to model mappings
- Support for Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
- Updated model registry in utils.py

Changed#

Message Architecture Refactoring: Extracted and refactored messaging system for multimodal support
- Extracted StreamChunk classes into dedicated module (massgen/stream_chunk/)
- Enhanced message templates for image generation workflows
- Improved orchestrator and chat agent for multimodal message handling
Backend Enhancements: Extended backends for multimodal and file operations
- Enhanced response.py with image generation, understanding, and saving capabilities
- Improved base_with_mcp.py with image handling for MCP-based workflows
- New api_params_handler module for centralized parameter management including file uploads
- Better streaming and error handling for multimodal content
Frontend Display Improvements: Enhanced terminal UI for multimodal content
- Refactored rich_terminal_display.py for rendering images in terminal
- Improved message formatting and visual presentation

Documentations, Configurations and Resources#

New Configuration Files: Added multimodal and enhanced filesystem examples
- gpt4o_image_generation.yaml: Multi-agent image generation setup
- gpt5nano_image_understanding.yaml: Multi-agent image understanding configuration
- single_gpt4o_image_generation.yaml: Single agent image generation
- single_gpt5nano_image_understanding.yaml: Single agent image understanding
- single_gpt5nano_file_search.yaml: Single agent file search example
- grok4_gpt5_gemini_filesystem.yaml: Enhanced filesystem configuration
- Updated claude_code_gpt5nano.yaml with improved filesystem settings
Case Study Documentation: New multi-turn-filesystem-support.md demonstrating v0.0.25 multi-turn capabilities with Bob Dylan website example
Presentation Materials: New applied-ai-summit.html presentation with updated build scripts and call-to-action slides
Example Resources: New multimodality.jpg for testing multimodal capabilities under massgen/configs/resources/v0.0.27-example/

Technical Details#

Major Features: Image processing foundation, StreamChunk architecture, file upload/search, workspace multimodal tools
New Module: massgen/stream_chunk/ with base, text, and multimodal classes
Contributors: @qidanrui @sonichi @praneeth999 @ncrispino @Henry-811 and the MassGen team

[0.0.26] - 2025-10-01#

Added#

File Deletion and Workspace Management: New MCP tools for workspace file operations
- New workspace deletion tools: delete_file, delete_files_batch for managing workspace files
- New comparison tools: compare_directories, compare_files for file diffing
- Consolidated _workspace_tools_server.py replacing previous _workspace_copy_server.py
- Improved workspace cleanup mechanisms for multi-turn sessions
- Proper permission checks for all file operations
File-Based Context Paths: Support for single file access without exposing entire directories
- Context paths can now be individual files, not just directories
- Better control over agent access to specific reference files
- Enhanced path validation distinguishing between file and directory contexts
Protected Paths Feature: Prevent agents from modifying specific reference files
- Protected paths within write-permitted context paths
- Agents can read but not modify protected files

Changed#

Code Refactoring: Improved module structure and import paths
- Moved utility modules from backend/utils/ to top-level massgen/ directory
- Relocated api_params_handler, formatter, and filesystem_manager modules
- Simplified import paths and improved code discoverability
- Better separation of concerns between backend-specific and shared utilities
Path Permission Manager: Major enhancements to permission system
- Enhanced will_be_writable logic for better permission state tracking
- Improved path validation distinguishing between context paths and workspace paths
- Comprehensive test coverage in test_path_permission_manager.py
- Better handling of edge cases and nested path scenarios

Fixed#

Path Permission Edge Cases: Resolved various permission checking issues
- Fixed file context path validation logic
- Corrected protected path matching behavior
- Improved handling of nested paths and symbolic links
- Better error handling for non-existent paths

Documentations, Configurations and Resources#

Example Resources: Added v0.0.26 example resources for testing new features
- Bob Dylan themed website with multiple pages and styles
- Additional HTML, CSS, and JavaScript examples
- Resources organized under massgen/configs/resources/v0.0.26-example/
Design Documentation: Added comprehensive design documentation
- New file_deletion_and_context_files.md documenting file deletion and context file features
- Updated permissions_and_context_files.md with v0.0.26 features
- Added detailed examples for protected paths and file context paths
Release Workflow Documentation: Added comprehensive release example checklist
- Step-by-step guide for release preparation in docs/workflows/release_example_checklist.md
- Best practices for testing new features
Configuration Examples: New configuration examples for v0.0.26 features
- gemini_gpt5nano_protected_paths.yaml: Protected paths example
- gemini_gpt5nano_file_context_path.yaml: File-based context paths example
- gemini_gemini_workspace_cleanup.yaml: Workspace cleanup example

Technical Details#

Commits: 20+ commits including file deletion tools, protected paths, and refactoring
Files Modified: 46 files with 4,343 insertions and 836 deletions
Major Features: File deletion tools, protected paths, file-based context paths, enhanced CLI prompts
New Tools: delete_file, delete_files_batch, compare_directories, compare_files MCP tools
Contributors: @praneeth999 @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.25] - 2025-09-29#

Added#

Multi-Turn Filesystem Support: Complete implementation for persistent filesystem context across conversation turns
- Automatic session management (no flag needed)
- Persistent workspace management across conversation turns with .massgen directory
- Workspace snapshot preservation and restoration between turns
- Support for maintaining file context and modifications throughout multi-turn sessions
- New configuration examples: two_gemini_flash_filesystem_multiturn.yaml, grok4_gpt5_gemini_filesystem_multiturn.yaml, grok4_gpt5_claude_code_filesystem_multiturn.yaml
- Design documentation in multi_turn_filesystem_design.md
SGLang Backend Integration: Added SGLang support to inference backend alongside existing vLLM
- New SGLang server support with default port 30000 and SGLANG_API_KEY environment variable
- SGLang-specific parameters support (e.g., separate_reasoning for guided generation)
- Auto-detection between vLLM and SGLang servers based on configuration
- New configuration two_qwen_vllm_sglang.yaml for mixed server deployments
- Unified InferenceBackend class replacing separate vllm.py implementation
- Updated documentation renamed from vllm_implementation.md to inference_backend.md
Enhanced Path Permission System: New exclusion patterns and validation improvements
- Added DEFAULT_EXCLUDED_PATTERNS for common directories (.git, node_modules, .venv, etc.)
- New will_be_writable flag for better permission state tracking
- Improved path validation with different handling for context vs workspace paths
- Enhanced test coverage in test_path_permission_manager.py

Changed#

CLI Enhancements: Major improvements to command-line interface
- Enhanced logging with configurable log levels and file output
- Improved error handling and user feedback
System Prompt Improvements: Refined agent system prompts for better performance
- Clearer instructions for file context handling
- Better guidance for multi-turn conversations
- Improved prompt templates for filesystem operations
Documentation Updates: Comprehensive documentation improvements
- Updated README with clearer installation instructions

Fixed#

Filesystem Manager: Resolved workspace and permission issues
- Fixed warnings for non-existent temporary workspaces
- Better cleanup of old workspaces
- Fixed relative path issues in workspace copy operations
Configuration Issues: Multiple configuration fixes
- Fixed multi-agent configuration templates
- Fixed code generation prompts for consistency

Technical Details#

Commits: 30+ commits including multi-turn filesystem, SGLang integration, and bug fixes
Files Modified: 33 files with 3,188 insertions and 642 deletions
Major Features: Multi-turn filesystem support, unified vLLM/SGLang backend, enhanced permissions
New Backend: SGLang integration alongside existing vLLM support
Contributors: @praneeth999 @ncrispino @qidanrui @sonichi @Henry-811 and the MassGen team

[0.0.24] - 2025-09-26#

Added#

vLLM Backend Support: Complete integration with vLLM for high-performance local model serving
- New vllm.py backend supporting VLLM’s OpenAI-compatible API
- Configuration examples in three_agents_vllm.yaml
- Comprehensive documentation in vllm_implementation.md
- Support for large-scale model inference with optimized performance
POE Provider Support: Extended ChatCompletions backend to support POE (Platform for Open Exploration)
- Added POE provider integration for accessing multiple AI models through a single platform
- Seamless integration with existing ChatCompletions infrastructure
GPT-5-Codex Model Recognition: Added GPT-5-Codex to model registry
- Extended model mappings in utils.py to recognize gpt-5-codex as a valid OpenAI model
Backend Utility Modules: Major refactoring for improved modularity
- New api_params_handler module for centralized API parameter management
- New formatter module for standardized message formatting across backends
- New token_manager module for unified token counting and management
- Extracted filesystem utilities into dedicated filesystem_manager module

Changed#

Backend Consolidation: Significant code refactoring and simplification
- Refactored chat_completions.py and response.py with cleaner API handler patterns
- Moved filesystem management from mcp_tools to backend/utils/filesystem_manager
- Improved separation of concerns with specialized handler modules
- Enhanced code reusability across different backend implementations
Documentation Updates: Improved documentation structure
- Moved permissions_and_context_files.md to backend docs
- Added multi-source agent integration design documentation
- Updated filesystem permissions case study for v0.0.21 and v0.0.22 features
CI/CD Pipeline: Enhanced automated release process
- Updated auto-release workflow for better reliability
- Improved GitHub Actions configuration
Pre-commit Configuration: Updated code quality tools
- Enhanced pre-commit hooks for better code consistency
- Updated linting rules for improved code standards

Fixed#

Streaming Chunk Processing: Resolved critical bugs in chunk handling
- Fixed chunk processing errors in response streaming
- Improved error handling for malformed chunks
- Better resilience in stream processing pipeline
Gemini Backend Session Management: Improved cleanup
- Implemented proper session closure for google-genai aiohttp client
- Added explicit cleanup of aiohttp sessions to prevent potential resource leaks

Technical Details#

Commits: 35 commits including backend refactoring, vLLM integration, and bug fixes
Files Modified: 50+ files across backend, utilities, configurations, and documentation
Major Refactor: Complete restructuring of backend utilities
New Backend: vLLM integration for high-performance local inference
Contributors: @qidanrui @sonichi @praneeth999 @ncrispino @Henry-811 and the MassGen team

[0.0.23] - 2025-09-24#

Added#

Backend Architecture Refactoring: Major consolidation of MCP functionality
- New base_with_mcp.py base class consolidating common MCP functionality (488 lines)
- Extracted shared MCP logic from individual backends into unified base class
- Standardized MCP client initialization and error handling across all backends
Formatter Module: Extracted message and tool formatting logic into dedicated module
- New massgen/formatter/ module with specialized formatters
- message_formatter.py: Handles message formatting across backends
- tool_formatter.py: Manages tool call formatting
- mcp_tool_formatter.py: Specialized MCP tool formatting

Changed#

Backend Consolidation: Massive code deduplication across backends
- Reduced chat_completions.py by 700+ lines
- Reduced claude.py by 700+ lines
- Simplified response.py by 468+ lines
- Total reduction: ~1,932 lines removed across core backend files

Fixed#

Coordination Table Display: Fixed escape key handling on macOS
- Updated create_coordination_table.py and rich_terminal_display.py

Technical Details#

Commits: 20+ commits focusing on backend refactoring and infrastructure improvements
Files Modified: 100+ files across backend, documentation, CI/CD, and presentation components
Lines Changed: Net reduction of ~1,932 lines through backend consolidation
Major Refactor: MCP functionality extracted into shared base_with_mcp.py base class
Contributors: @qidanrui @ncrispino @Henry-811 and the MassGen team

[0.0.22] - 2025-09-22#

Added#

Workspace Copy Tools via MCP: New file copying capabilities for efficient workspace operations
- Added workspace_copy_server.py with MCP-based file copying functionality (369 lines)
- Support for copying files and directories between workspaces
- Efficient handling of large files with streaming operations
- Testing infrastructure for copy operations
Configuration Organization: Major restructuring of configuration files for better usability
- New hierarchical structure: basic/, providers/, tools/, teams/ directories
- Added comprehensive README.md for configuration guide
- New BACKEND_CONFIGURATION.md with detailed backend setup
- Organized configs by use case and provider for easier navigation
- Added provider-specific examples (Claude, OpenAI, Gemini, Azure)
Enhanced File Operations: Improved file handling for large-scale operations
- Clear all temporary workspaces at startup for clean state
- Enhanced security validation in MCP tools

Changed#

Workspace Management: Optimized workspace operations and path handling
- Enhanced filesystem_manager.py with 193 additional lines
- Run MCP servers through FastMCP to avoid banner displays
Backend Enhancements: Improved backend capabilities
- Improved response.py with better error handling

Fixed#

Write Tool Call Issues: Resolved large character count problems
- Fixed write tool call issues when dealing with large character counts
Path Resolution Issues: Resolved various path-related bugs
- Fixed relative/absolute path workspace issues
- Improved path validation and normalization
Documentation Fixes: Corrected multiple documentation issues
- Fixed broken links in case studies
- Fixed config file paths in documentation and examples
- Corrected example commands with proper paths

Technical Details#

Commits: 50+ commits including workspace copy, configuration restructuring, and documentation improvements
Files Modified: 90+ files across configs, backend, mcp_tools, and documentation
Major Refactoring: Configuration file reorganization into logical categories
New Documentation: Added 762+ lines of documentation for configs and backends
Contributors: @ncrispino @qidanrui @Henry-811 and the MassGen team

[0.0.21] - 2025-09-19#

Added#

Advanced Filesystem Permissions System: Comprehensive permission management for agent file access
- New PathPermissionManager class for granular permission validation
- User context paths with configurable READ/WRITE permissions for multi-agent file sharing
- Test suite for permission validation in test_path_permission_manager.py
- Documentation in permissions_and_context_files.md for implementation guide
Function Hook Manager: Per-agent function call permission system
- Refactored FunctionHookManager to be per-agent rather than global
- Pre-tool-use hooks for validating file operations before execution
- Support for write permission enforcement during context agent operations
- Integration with all function-based backends (OpenAI, Claude, Chat Completions)
Grok MCP Integration: Extended MCP support to Grok backend
- Migrated Grok backend to inherit from Chat Completions backend
- Full MCP server support for Grok including stdio and HTTP transports
- Filesystem support through MCP servers
New Configuration Files: Added test and example configurations
- grok3_mini_mcp_test.yaml: Grok MCP testing configuration
- grok3_mini_mcp_example.yaml: Grok MCP usage example
- grok3_mini_streamable_http_test.yaml: Grok HTTP streaming test
- grok_single_agent.yaml: Single Grok agent configuration
- fs_permissions_test.yaml: Filesystem permissions testing configuration

Changed#

Backend Architecture: Unified backend implementations and permission support
- Grok backend refactored to use Chat Completions backend
- All backends now support per-agent permission management
- Enhanced context file support across Claude, Gemini, and OpenAI backends

Technical Details#

Commits: 20+ commits including permission system, Grok MCP, and terminal improvements
Files Modified: 40+ files across backends, MCP tools, permissions, and display modules
New Features: Filesystem permissions, per-agent hooks, Grok MCP via Chat Completions
Contributors: @Eric-Shang @ncrispino @qidanrui @Henry-811 and the MassGen team

[0.0.20] - 2025-09-17#

Added#

Claude Backend MCP Support: Extended MCP (Model Context Protocol) integration to Claude backend
- Filesystem support through MCP servers (FilesystemSupport.MCP) for Claude backend
- Support for both stdio and HTTP-based MCP servers with Claude Messages API
- Seamless integration with existing Claude function calling and tool use
- Recursive execution model allowing Claude to autonomously chain multiple tool calls in sequence without user intervention
- Enhanced error handling and retry mechanisms for Claude MCP operations
MCP Configuration Examples: New YAML configurations for Claude MCP usage
- claude_mcp_test.yaml: Basic Claude MCP testing with test server
- claude_mcp_example.yaml: Claude MCP integration example
- claude_streamable_http_test.yaml: HTTP transport testing for Claude MCP
Documentation: Enhanced MCP technical documentation
- MCP_IMPLEMENTATION_CLAUDE_BACKEND.md: Complete technical documentation for Claude MCP integration
- Detailed architecture diagrams and implementation guides

Changed#

Backend Enhancements: Improved MCP support across backends
- Extended MCP integration from Gemini and Chat Completions to include Claude backend
- Enhanced error reporting and debugging for MCP operations
- Added Kimi/Moonshot API key support in Chat Completions backend

Technical Details#

New Features: Claude backend MCP integration with recursive execution model
Files Modified: Claude backend modules (claude.py), MCP tools, configuration examples
MCP Coverage: Major backends now support MCP (Claude, Gemini, Chat Completions including OpenAI)
Contributors: @praneeth999 @qidanrui @sonichi @ncrispino @Henry-811 MassGen development team

[0.0.19] - 2025-09-15#

Added#

Coordination Tracking System: Comprehensive tracking of multi-agent coordination events
- New coordination_tracker.py with CoordinationTracker class for capturing agent state transitions
- Event-based tracking with timestamps and context preservation
- Support for recording answers, votes, and coordination phases
- New create_coordination_table.py utility in massgen/frontend/displays/ for generating coordination reports
Enhanced Agent Status Management: New enums for better state tracking
- Added ActionType enum in massgen/utils.py: NEW_ANSWER, VOTE, VOTE_IGNORED, ERROR, TIMEOUT, CANCELLED
- Added AgentStatus enum in massgen/utils.py: STREAMING, VOTED, ANSWERED, RESTARTING, ERROR, TIMEOUT, COMPLETED
- Improved state machine for agent coordination lifecycle

Changed#

Frontend Display Enhancements: Improved terminal interface with coordination visualization
- Modified massgen/frontend/displays/rich_terminal_display.py to add coordination table display method
- Added new terminal menu option ‘r’ to display coordination table
- Enhanced menu system with better organization of debugging tools
- Support for rich-formatted tables showing agent interactions across rounds

Technical Details#

Commits: 20+ commits including coordination tracking system and frontend enhancements
Files Modified: 5+ files across coordination tracking, frontend displays, and utilities
New Features: Coordination event tracking with visualization capabilities
Contributors: @ncrispino @qidanrui @sonichi @a5507203 @Henry-811 and the MassGen team

[0.0.18] - 2025-09-12#

Added#

Chat Completions MCP Support: Extended MCP (Model Context Protocol) integration to ChatCompletions-based backends
- Full MCP support for all Chat Completions providers (Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, OpenRouter)
- Filesystem support through MCP servers (FilesystemSupport.MCP) for Chat Completions backend
- Cross-provider function calling compatibility enabling seamless MCP tool execution across different providers
- Universal MCP server compatibility with existing stdio and streamable-http transports
New MCP Configuration Examples: Added 9 new Chat Completions MCP configurations
- GPT-OSS configurations: gpt_oss_mcp_example.yaml, gpt_oss_mcp_test.yaml, gpt_oss_streamable_http_test.yaml
- Qwen API configurations: qwen_api_mcp_example.yaml, qwen_api_mcp_test.yaml, qwen_api_streamable_http_test.yaml
- Qwen Local configurations: qwen_local_mcp_example.yaml, qwen_local_mcp_test.yaml, qwen_local_streamable_http_test.yaml
Enhanced LMStudio Backend: Improved local model support
- Better tracking of attempted model loads
- Improved server output handling and error reporting

Changed#

Backend Architecture: Major MCP framework expansion
- Extended existing v0.0.15 MCP infrastructure to support all ChatCompletions providers
- Refactored chat_completions.py with 1200+ lines of MCP integration code
- Enhanced error handling and retry mechanisms for provider-specific quirks
CLI Improvements: Better backend creation and provider detection
- Enhanced backend creation logic for improved provider handling
- Better system message handling for different backend types

Technical Details#

Main Feature: Chat Completions MCP integration enabling all providers to use MCP tools
Files Modified: 20+ files across backend, mcp_tools, configurations, and CLI
Contributors: @praneeth999 @qidanrui @sonichi @a5507203 @ncrispino @Henry-811 and the MassGen team

[0.0.17] - 2025-09-10#

Added#

OpenAI Backend MCP Support: Extended MCP (Model Context Protocol) integration to OpenAI backend
- Full MCP tool discovery and execution capabilities for OpenAI models
- Support for both stdio and HTTP-based MCP servers with OpenAI
- Seamless integration with existing OpenAI function calling
- Robust error handling and retry mechanisms
MCP Configuration Examples: New YAML configurations for OpenAI MCP usage
- gpt5_mini_mcp_test.yaml: Basic OpenAI MCP testing with test server
- gpt5_mini_mcp_example.yaml: Weather service integration example for OpenAI
- gpt5_mini_streamable_http_test.yaml: HTTP transport testing for OpenAI MCP
- Enhanced existing multi-agent configurations with OpenAI MCP support
Documentation: Added case studies and technical documentation
- unified-filesystem-mcp-integration.md: Case study demonstrating unified filesystem capabilities with MCP integration across multiple backends (from v0.0.16)
- MCP_INTEGRATION_RESPONSE_BACKEND.md: Technical documentation for MCP integration with response backends

Changed#

Backend Enhancements: Improved MCP support across backends
- Extended MCP integration from Gemini and Claude Code to include OpenAI backend
- Unified MCP tool handling across all supported backends
- Enhanced error reporting and debugging for MCP operations

Technical Details#

New Features: OpenAI backend MCP integration
Documentation: Added case study for unified filesystem MCP integration
Contributors: @praneeth999 @qidanrui @sonichi @ncrispino @a5507203 @Henry-811 and the MassGen team

[0.0.16] - 2025-09-08#

Added#

Unified Filesystem Support with MCP Integration: Advanced filesystem capabilities designed for all backends
- Complete FilesystemManager class providing unified filesystem access with extensible backend support
- Currently supports Gemini and Claude Code backends, designed for seamless expansion to all backends
- MCP-based filesystem operations enabling file manipulation, workspace management, and cross-agent collaboration
Expanded Configuration Library: New YAML configurations for various use cases
- Gemini MCP Filesystem Testing: gemini_mcp_filesystem_test.yaml, gemini_mcp_filesystem_test_sharing.yaml, gemini_mcp_filesystem_test_single_agent.yaml, gemini_mcp_filesystem_test_with_claude_code.yaml
- Hybrid Model Setups: geminicode_gpt5nano.yaml
Case Studies: Added comprehensive case studies from previous versions
- gemini-mcp-notion-integration.md: Gemini MCP Notion server integration and productivity workflows
- claude-code-workspace-management.md: Claude Code context sharing and workspace management demonstrations

Technical Details#

Commits: 30+ commits including workspace redesign and orchestrator enhancements
Files Modified: 40+ files across orchestrator, mcp_tools, configurations, and case studies
New Architecture: Complete workspace management system with FilesystemManager
Contributors: @ncrispino @a5507203 @sonichi @Henry-811 and the MassGen team

[0.0.15] - 2025-09-05#

Added#

MCP (Model Context Protocol) Integration Framework: Complete implementation for external tool integration
- New massgen/mcp_tools/ package with 8 core modules for MCP support
- Multi-server MCP client supporting simultaneous connections to multiple MCP servers
- Two transport types: stdio (process-based) and streamable-http (web-based)
- Circuit breaker patterns for fault tolerance and reliability
- Comprehensive security framework with command sanitization and validation
- Automatic tool discovery with name prefixing for multi-server setups
Gemini MCP Support: Full MCP integration for Gemini backend
- Session-based tool execution via Gemini SDK
- Automatic tool discovery and calling capabilities
- Robust error handling with exponential backoff
- Support for both stdio and HTTP-based MCP servers
- Integration with existing Gemini function calling
Test Infrastructure for MCP: Development and testing utilities
- Simple stdio-based MCP test server (mcp_test_server.py)
- FastMCP streamable-http test server (test_http_mcp_server.py)
- Comprehensive test suite for MCP integration
MCP Configuration Examples: New YAML configurations for MCP usage
- gemini_mcp_test.yaml: Basic Gemini MCP testing
- gemini_mcp_example.yaml: Weather service integration example
- gemini_streamable_http_test.yaml: HTTP transport testing
- multimcp_gemini.yaml: Multi-server MCP configuration
- Additional Claude Code MCP configurations

Changed#

Dependencies: Updated package requirements
- Added mcp>=1.12.0 for official MCP protocol support
- Added aiohttp>=3.8.0 for HTTP-based MCP communication
- Updated pyproject.toml and requirements.txt
Documentation: Enhanced project documentation
- Created technical analysis documents for Gemini MCP integration
- Added comprehensive MCP tools README with architecture diagrams
- Added security and troubleshooting guides for MCP

Technical Details#

Commits: 40+ commits including MCP integration, documentation, and bug fixes
Files Modified: 35+ files across MCP modules, backends, configurations, and tests
Security Features: Configurable security levels (strict/moderate/permissive)
Contributors: @praneeth999 @qidanrui @sonichi @a5507203 @ncrispino @Henry-811 and the MassGen team

[0.0.14] - 2025-09-02#

Added#

Enhanced Logging System: Improved logging infrastructure with add_log feature
- Better log organization and preservation for multi-agent workflows
- Enhanced workspace management for Claude Code agents
- New final answer directory structure in Claude Code and logs for storing final results

Documentation#

Release Documents: Updated release documentation and materials
- Updated CHANGELOG.md for better release tracking
- Removed unnecessary use case documentation

Technical Details#

Commits: 19 commits
Files Modified: Logging system enhancements, documentation updates
New Features: Enhanced logging, improved final presentation logging for Claude Code
Contributors: @qidanrui @sonichi and the MassGen team

[0.0.13] - 2025-08-28#

Added#

Unified Logging System: Better logging infrastructure for better debugging and monitoring
- New centralized logger_config.py with colored console output and file logging
- Debug mode support via --debug CLI flag for verbose logging
- Consistent logging format across all backends, including Claude, Gemini, Grok, Azure OpenAI, and other providers
- Color-coded log levels for better visibility (DEBUG: cyan, INFO: green)
Windows Platform Support: Enhanced cross-platform compatibility
- Windows-specific fixes for terminal display and color output
- Improved path handling for Windows file systems
- Better process management on Windows platform

Changed#

Frontend Improvements: Refined display
- Enhanced rich terminal display formatting to not show debug info in the final presentation
Documentation Updates: Improved project documentation
- Updated CONTRIBUTING.md with better guidelines
- Enhanced README with logging configuration details
- Renamed roadmap from v0.0.13 to v0.0.14 for future planning

Technical Details#

Commits: 35+ commits including new logging system and Windows support
Files Modified: 24+ files across backend, frontend, logging, and CLI modules
New Features: Unified logging system with debug mode, Windows platform support
Contributors: @qidanrui @sonichi @Henry-811 @JeffreyCh0 @voidcenter and the MassGen team

[0.0.12] - 2025-08-27#

Added#

Enhanced Claude Code Agent Context Sharing: Improved multiple Claude Code agent coordination with workspace sharing
- New workspace snapshot stored in orchestrator’s space for better context management
- New temporary working directory for each agent, stored in orchestrator’s space
- Claude Code agents can now share context by referencing their own temporary working directory in the orchestrator’s workspace
- Anonymous agent context mapping when referencing temporary directories
- Improved context preservation across agent coordination cycles
Advanced Orchestrator Configurations: Enhanced orchestrator configurations
- Configurable system message support for orchestrator
- New snapshot and temporary workspace settings for better context management

Changed#

Documentation Updates: documentation improvements
- Updated README with current features and usage examples
- Improved configuration examples and setup instructions

Technical Details#

Commits: 10+ commits including context sharing enhancements, workspace management, and configuration improvements
Files Modified: 20+ files across orchestrator, backend, configuration, and documentation
New Features: Enhanced Claude Code agent workspace sharing with temporary working directories and snapshot mechanisms
Contributors: @qidanrui @sonichi @Henry-811 @JeffreyCh0 @voidcenter and the MassGen team

[0.0.11] - 2025-08-25#

Known Issues#

System Message Handling in Multi-Agent Coordination: Critical issues affecting Claude Code agents
- Lost System Messages During Final Presentation (orchestrator.py:1183)
  - Claude Code agents lose domain expertise during final presentation
  - ConfigurableAgent doesn’t properly expose system messages via agent.system_message
- Backend Ignores System Messages (claude_code.py:754-762)
  - Claude Code backend filters out system messages from presentation_messages
  - Only processes user messages, causing loss of agent expertise context
  - System message handling only works during initial client creation, not with reset_chat=True
- Ambiguous Configuration Sources
  - Multiple conflicting system message sources: custom_system_instruction, system_prompt, append_system_prompt
  - Backend parameters silently override AgentConfig settings
  - Unclear precedence and behavior documentation
- Architecture Violations
  - Orchestrator contains Claude Code-specific implementation details
  - Tight coupling prevents easy addition of new backends
  - Violates separation of concerns principle

Fixed#

Custom System Message Support: Enhanced system message configuration and preservation
- Added base_system_message parameter to conversation builders for agent’s custom system message
- Orchestrator now passes agent’s get_configurable_system_message() to conversation builders
- Custom system messages properly combined with MassGen coordination instructions instead of being overwritten
- Backend-specific system prompt customization (system_prompt, append_system_prompt)
Claude Code Backend Enhancements: Improved integration and configuration
- Better system message handling and extraction
- Enhanced JSON structured response parsing
- Improved coordination action descriptions
Final Presentation & Agent Logic: Enhanced multi-agent coordination (#135)
- Improved final presentation handling for Claude Code agents
- Better coordination between agents during final answer selection
- Enhanced CLI presentation logic
- Agent configuration improvements for workflow coordination
Evaluation Message Enhancement: Improved synthesis instructions
- Changed to “digest existing answers, combine their strengths, and do additional work to address their weaknesses”
- Added “well” qualifier to evaluation questions
- More explicit guidance for agents to synthesize and improve upon existing answers

Changed#

Documentation Updates: Enhanced project documentation
- Renamed roadmap from v0.0.11 to v0.0.12 for future planning
- Updated README with latest features and improvements
- Improved CONTRIBUTING guidelines
- Enhanced configuration examples and best practices

Added#

New Configuration Files: Introduced additional YAML configuration files
- Added multi_agent_playwright_automation.yaml for browser automation workflows

Removed#

Deprecated Configurations: Cleaned up configuration files
- Removed gemini_claude_code_paper_search_mcp.yaml
- Removed gpt5_claude_code_paper_search_mcp.yaml
Gemini CLI Tests: Removed Gemini CLI related tests

Technical Details#

Commits: 25+ commits including bug fixes, feature additions, and improvements
Files Modified: 35+ files across backend, orchestrator, frontend, configuration, and documentation
New Configuration: multi_agent_playwright_automation.yaml for browser automation workflows
Contributors: @qidanrui @Leezekun @sonichi @voidcenter @Daucloud @Henry-811 and the MassGen team

[0.0.10] - 2025-08-22#

Added#

Azure OpenAI Support: Integration with Azure OpenAI services
- New azure_openai.py backend with async streaming capabilities
- Support for Azure-hosted GPT-4.1 and GPT-5-chat models
- Configuration examples for single and multi-agent Azure setups
- Test suite for Azure OpenAI functionality
Enhanced Claude Code Backend: Major refactoring and improvements
- Simplified MCP (Model Context Protocol) integration
Final Presentation Support: New orchestrator presentation capabilities
- Support for final answer presentation in multi-agent scenarios
- Fallback mechanisms for presentation generation
- Test coverage for presentation functionality

Fixed#

Claude Code MCP: Cleaned up and simplified MCP implementation
- Removed redundant MCP server and transport modules
Configuration Management: Improved YAML configuration handling
- Fixed Azure OpenAI deployment configurations
- Updated model mappings for Azure services

Changed#

Backend Architecture: Significant refactoring of backend systems
- Consolidated Azure OpenAI implementation using AsyncAzureOpenAI
- Improved error handling and streaming capabilities
- Enhanced async support across all backends
Documentation Updates: Enhanced project documentation
- Updated README with Azure OpenAI setup instructions
- Renamed roadmap from v0.0.10 to v0.0.11
- Improved presentation materials for DataHack Summit 2025
Test Infrastructure: Expanded test coverage
- Added comprehensive Azure OpenAI backend tests
- Integration tests for final presentation functionality
- Simplified test structure with better coverage

Removed#

Deprecated MCP Components: Removed unused MCP modules
- Removed standalone MCP client, transport, and server implementations
- Cleaned up MCP test files and testing checklist
- Simplified Claude Code backend by removing redundant MCP code

Technical Details#

Commits: 35+ commits including Azure OpenAI integration and Claude Code improvements
Files Modified: 30+ files across backend, configuration, tests, and documentation
New Backend: Azure OpenAI backend with full async support
Contributors: @qidanrui @Leezekun @sonichi and the MassGen team

[0.0.9] - 2025-08-22#

Added#

Quick Start Guide: Comprehensive quickstart documentation in README
- Streamlined setup instructions for new users
- Example configurations for getting started quickly
- Clear installation and usage steps
Multi-Agent Configuration Examples: New configuration files for various setups
- Paper search configuration with GPT-5 and Claude Code
- Multi-agent setups with different model combinations
Roadmap Documentation: Added comprehensive roadmap for version 0.0.10
- Focused on Claude Code context sharing between agents
- Multi-agent context synchronization planning
- Enhanced backend features and CLI improvements roadmap

Fixed#

Web Search Processing: Fixed bug in response handling for web search functionality
- Improved error handling in web search responses
- Better streaming of search results
Rich Terminal Display: Fixed rendering issues in terminal UI
- Resolved display formatting problems
- Improved message rendering consistency

Changed#

Claude Code Integration: Optimized Claude Code implementation
- MCP (Model Context Protocol) integration
- Streamlined Claude Code backend configuration
Documentation Updates: Enhanced project documentation
- Updated README with quickstart guide
- Added CONTRIBUTING.md guidelines
- Improved configuration examples

Technical Details#

Commits: 10 commits including bug fixes, code cleanup, and documentation updates
Files Modified: Multiple files across backend, configurations, and documentation
Contributors: @qidanrui @sonichi @Leezekun @voidcenter @JeffreyCh0 @stellaxiang

[0.0.8] - 2025-08-18#

Added#

Timeout Management System: Timeout capabilities for better control and time management
- New TimeoutConfig class for configuring timeout settings at different levels
- Orchestrator-level timeout with graceful fallback
- Added fast_timeout_example.yaml configuration demonstrating conservative timeout settings
- Test suite for timeout mechanisms in test_timeout.py
- Timeout indicators in Rich Terminal Display showing remaining time
Enhanced Display Features: Improved visual feedback and user experience
- Optimized message display formatting for better readability
- Enhanced status indicators for timeout warnings and fallback notifications
- Improved coordination UI with better multi-agent status tracking

Fixed#

Display Optimization: Multiple improvements to message rendering
- Fixed message display synchronization issues
- Optimized terminal display refresh rates
- Improved handling of concurrent agent outputs
- Better formatting for multi-line responses
Configuration Management: Enhanced robustness of configuration loading
- Fixed import ordering issues in CLI module
- Improved error handling for missing configurations
- Better validation of timeout settings

Changed#

Orchestrator Architecture: Simplified and enhanced timeout implementation
- Refactored timeout handling to be more efficient and maintainable
- Improved graceful degradation when timeouts occur
- Better integration with frontend displays for timeout notifications
- Enhanced error messages for timeout scenarios
Code Cleanup: Removed deprecated configurations and improved code organization
- Removed obsolete two_agents_claude_code configuration
- Cleaned up unused imports and redundant code
- Reformatted files for better consistency
CLI Enhancements: Improved command-line interface functionality
- Better timeout configuration parsing
- Enhanced error reporting for timeout scenarios
- Improved help documentation for timeout settings

Technical Details#

Commits: 18 commits including various optimizations and bug fixes
Files Modified: 13+ files across orchestrator, frontend, configuration, and test modules
Key Features: Timeout management system with graceful fallback, enhanced display optimizations
New Configuration: fast_timeout_example.yaml for time-conscious usage
Contributors: @qidanrui @Leezekun @sonichi @voidcenter

[0.0.7] - 2025-08-15#

Added#

Local Model Support: Complete integration with LM Studio for running open-weight models locally
- New lmstudio.py backend with automatic server management
- Automatic model downloading and loading capabilities
- Zero-cost reporting for local model usage
Extended Provider Support: Enhanced ChatCompletionsBackend to support multiple providers
- Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, OpenRouter
- Provider-specific environment variable detection
- Automatic provider name inference from base URLs
New Configuration Files: Added configurations for local and hybrid model setups
- lmstudio.yaml: Single agent configuration for LM Studio
- two_agents_opensource_lmstudio.yaml: Hybrid setup with GPT-5 and local Qwen model
- gpt5nano_glm_qwen.yaml: Three-agent setup combining Cerebras, ZAI GLM-4.5, and local Qwen
- Updated three_agents_opensource.yaml for open-source model combinations

Fixed#

Backend Stability: Improved error handling across all backend systems
- Fixed API key resolution and client initialization
- Enhanced provider name detection and configuration
- Resolved streaming issues in ChatCompletionsBackend
Documentation: Corrected references and updated model naming conventions
- Fixed GPT model references in documentation diagrams
- Updated case study file naming consistency

Changed#

Backend Architecture: Refactored ChatCompletionsBackend for better extensibility
- Improved provider registry and configuration management
- Enhanced logging and debugging capabilities
- Streamlined message processing and tool handling
Dependencies: Added new requirements for local model support
- Added lmstudio==1.4.1 for LM Studio Python SDK integration
Documentation Updates: Enhanced documentation for local model usage
- Updated environment variables documentation
- Added setup instructions for LM Studio integration
- Improved backend configuration examples

Technical Details#

Commits: 16 commits including merge pull requests #80 and #100
Files Modified: 17+ files across backend, configuration, documentation, and CLI modules
New Dependencies: LM Studio SDK (lmstudio==1.4.1)
Contributors: @qidanrui @sonichi @Leezekun @praneeth999 @voidcenter

[0.0.6] - 2025-08-13#

Added#

GLM-4.5 Model Support: Integration with ZhipuAI’s GLM-4.5 model family
- Added GLM-4.5 backend support in chat_completions.py
- New configuration file zai_glm45.yaml for GLM-4.5 agent setup
- Updated zai_coding_team.yaml with GLM-4.5 integration
- Added GLM-4.5 model mappings and environment variable support
Enhanced Reasoning Display: Improved reasoning presentation for GLM models
- Added reasoning start and completion indicators in frontend displays
- Enhanced coordination UI to show reasoning progress
- Better visual formatting for reasoning states in terminal display

Fixed#

Claude Code Backend: Updated default allowed tools configuration
- Fixed default tools setup in claude_code.py backend

Changed#

Documentation Updates: Updated README.md with GLM-4.5 support information
- Added GLM-4.5 to supported models list
- Updated environment variables documentation for ZhipuAI integration
- Enhanced model comparison and configuration examples
Configuration Management: Enhanced agent configuration system
- Updated agent_config.py with GLM-4.5 support
- Improved CLI integration for GLM models
- Better model parameter handling in utils.py

Technical Details#

Commits: 6 major commits including merge pull requests #90 and #94
Files Modified: 12+ files across backend, frontend, configuration, and documentation
New Dependencies: ZhipuAI GLM-4.5 model integration
Contributors: @Stanislas0 @qidanrui @sonichi @Leezekun @voidcenter

[0.0.5] - 2025-08-11#

Added#

Claude Code Integration: Complete integration with Claude Code CLI backend
- New claude_code.py backend with streaming capabilities and tool support
- Support for Claude Code SDK with stateful conversation management
- JSON tool call functionality and proper tool result handling
- Session management with append system prompt support
New Configuration Files: Added Claude Code specific YAML configurations
- claude_code_single.yaml: Single agent setup using Claude Code backend
- claude_code_flash2.5.yaml: Multi-agent setup with Claude Code and Gemini Flash 2.5
- claude_code_flash2.5_gptoss.yaml: Multi-agent setup with Claude Code, Gemini Flash 2.5, and GPT-OSS
Test Coverage: Added test suite for Claude Code functionality
- test_claude_code_orchestrator.py: orchestrator testing
- Backend-specific test coverage for Claude Code integration

Fixed#

Backend Stability: Multiple critical bug fixes across all backend systems
- Fixed parameter handling in chat_completions.py, claude.py, gemini.py, grok.py
- Resolved response processing issues in response.py
- Improved error handling and client existence validation
Tool Call Processing: Enhanced tool call parsing and execution
- Deduplicated tool call parsing logic across backends
- Fixed JSON tool call functionality and result formatting
- Improved builtin tool result handling in streaming contexts
Message Handling: Resolved system message processing issues
- Fixed SystemMessage to StreamChunk conversion
- Proper session info extraction from system messages
- Cleaned up message formatting and display consistency
Frontend Display: Fixed output formatting and presentation
- Improved rich terminal display formatting
- Better coordination UI integration and multi-turn conversation display
- Enhanced status message display with proper newline handling

Changed#

Code Architecture: Significant refactoring and cleanup across the codebase
- Renamed and consolidated backend files for consistency
- Simplified chat agent architecture and removed redundant code
- Streamlined orchestrator logic with improved error handling
Configuration Management: Updated and cleaned up configuration files
- Updated agent configuration with Claude Code support
Backend Infrastructure: Enhanced backend parameter handling
- Improved stateful conversation management across all backends
- Better integration with orchestrator for multi-agent coordination
- Enhanced streaming capabilities with proper chunk processing
Documentation: Updated project documentation
- Added Claude Code setup instructions in README
- Updated backend architecture documentation
- Improved reasoning and streaming integration notes

Technical Details#

Commits: 50+ commits since version 0.0.4
Files Modified: 25+ files across backend, configuration, frontend, and test modules
Major Components Updated: Backend systems, orchestrator, frontend display, configuration management
New Dependencies: Added Claude Code SDK integration
Contributors: @qidanrui @randombet @sonichi

[0.0.4] - 2025-08-08#

Added#

GPT-5 Series Support: Full support for OpenAI’s GPT-5 model family
- GPT-5: Full-scale model with advanced capabilities
- GPT-5-mini: Efficient variant for faster responses
- GPT-5-nano: Lightweight model for resource-constrained deployments
New Model Parameters: Introduced GPT-5 specific configuration options
- text.verbosity: Control response detail level (low/medium/high)
- reasoning.effort: Configure reasoning depth (minimal/medium/high)
- Note: reasoning parameter is mutually exclusive with web search capability
Configuration Files: Added dedicated YAML configurations
- gpt5.yaml: Three-agent setup with GPT-5, GPT-5-mini, and GPT-5-nano
- gpt5_nano.yaml: Three GPT-5-nano agents with different reasoning levels
Extended Model Support: Added GPT-5 series to model mappings in utils.py
Reasoning for All Models: Extended reasoning parameter support beyond GPT-5 models

Fixed#

Tool Output Formatting: Added proper newline formatting for provider tool outputs
- Web search status messages now display on new lines
- Code interpreter status messages now display on new lines
- Search query display formatting improved
YAML Configuration: Fixed configuration syntax in GPT-5 related YAML files
Backend Response Handling: Multiple bug fixes in response.py for proper parameter handling

Changed#

Documentation Updates:
- Updated README.md to highlight GPT-5 series support
- Changed example commands to use GPT-5 models
- Added new backend configuration examples with GPT-5 specific parameters
- Updated models comparison table to show GPT-5 as latest OpenAI model
Parameter Handling: Improved backend parameter validation
- Temperature parameter now excluded for GPT-5 series models (like o-series)
- Max tokens parameter now excluded for GPT-5 series models
- Added conditional logic for GPT-5 specific parameters (text, reasoning)
Version Number: Updated to 0.0.4 in massgen/init.py

Technical Details#

Commits: 9 commits since version 0.0.3
Files Modified: 6 files (response.py, utils.py, README.md, init.py, and 2 new config files)
Contributors: @qidanrui @sonichi @voidcenter @JeffreyCh0 @praneeth999

[0.0.3] - 2025-08-03#

Added#

Complete architecture with foundation release
Multi-backend support: Claude (Messages API), Gemini (Chat API), Grok (Chat API), OpenAI (Responses API)
Builtin tools: Code execution and web search with streaming results
Async streaming with proper chat agent interfaces and tool result handling
Multi-agent orchestration with voting and consensus mechanisms
Real-time frontend displays with multi-region terminal UI
CLI with file-based YAML configuration and interactive mode
Proper StreamChunk architecture separating tool_calls from builtin_tool_results
Multi-turn conversation support with dynamic context reconstruction
Chat interface with orchestrator supporting async streaming
Case study configurations and specialized YAML configs
Claude backend support with production-ready multi-tool API and streaming
OpenAI builtin tools support for code execution and web search streaming

Fixed#

Grok backend testing and compatibility issues
CLI multi-turn conversation display with coordination UI integration
Claude streaming handler with proper tool argument capture
CLI backend parameter passing with proper ConfigurableAgent integration

Changed#

Restructured codebase with new architecture
Improved message handling and streaming capabilities
Enhanced frontend features and user experience

[0.0.1] - Initial Release#

Added#

Basic multi-agent system framework
Support for OpenAI, Gemini, and Grok backends
Simple configuration system
Basic streaming display
Initial logging capabilities

Changelog

Contents

Changelog#

Full Changelog#

Changelog#

[Unreleased]#

Recent Releases#

[0.1.87] - 2026-05-15#

Added#

Fixed#

Documentations, Configurations and Resources#

Notes#

Technical Details#

[0.1.86] - 2026-05-13#

Added#

Fixed#

Documentations, Configurations and Resources#

Tests#

Notes#

Technical Details#

[0.1.85] - 2026-05-11#

Added#

Why This Matters#

Documentations, Configurations and Resources#

Tests#

Notes#

Technical Details#

[0.1.84] - 2026-05-08#

Added#

Documentations, Configurations and Resources#

Tests#

Notes#

Technical Details#

[0.1.83] - 2026-05-01#

Added#

Changed#

Documentations, Configurations and Resources#

Tests#

Notes#

Technical Details#

[0.1.82] - 2026-04-29#

Added#

Changed#

Fixed#

Documentations, Configurations and Resources#

Technical Details#

[0.1.81] - 2026-04-27#

Added#

Technical Details#

[0.1.80] - 2026-04-22#

Added#

Changed#

Fixed#

Documentation, Configurations and Resources#

Technical Details#

[0.1.79] - 2026-04-20#

Added#

Changed#

Documentation, Configurations and Resources#

Technical Details#

[0.1.78] - 2026-04-17#

Added#

Tests#

Documentation, Configurations and Resources#

Technical Details#

[0.1.77] - 2026-04-15#

Added#

Changed#

Technical Details#

[0.1.76] - 2026-04-13#

Added#

Fixed#

Documentation, Configurations and Resources#

Technical Details#

[0.1.75] - 2026-04-10#

Added#

Changed#

Fixed#

Documentation, Configurations and Resources#

Technical Details#