CHANGELOG.md

Changelog

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

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

[Unreleased]

[0.9.1] - 2026-03-18

Changed

• Refactored run_review() into 4 phase functions for maintainability.
• Replaced ArgumentParser.error monkeypatch with _StreamSafeParser subclass.
• Improved CLI help text for --file priority and per-provider vs global timeout clarity.

Fixed

• adapter.cancel() failures are now logged instead of being silently swallowed.
• Duplicate agent names in agents.yaml now emit a warning and only the first entry is registered.
• Removed the unused load_state import.
• mco agent check now rejects empty agent names.
• Debate rounds now skip early when zero findings exist.
• Added streaming + divide mode regression coverage.
• Documented --debate / --divide mutual exclusivity in the README.
• Documented agent config-file priority order in the README.

[0.9.0] - 2026-03-18

Added

• Consensus engine with agreement-based scoring (consensus_score, confirmed / needs-verification / unverified).
• --stream live for real-time terminal progress per provider.
• --debate mode for structured multi-agent challenge rounds.
• --divide files|dimensions for task splitting across providers.
• Custom agent ecosystem: .mco/agents.yaml config, Ollama adapter, mco agent list / mco agent check.
• has_consensus_fallback field in synthesis output.

Changed

• --synthesize now produces algorithm-driven consensus analysis plus optional agent narrative.
• Findings are sorted by consensus level and then consensus score.
• SARIF confidence now maps to consensus_score.
• Markdown PR output now groups findings by consensus level.

Fixed

• --divide files no longer collects non-source directories.
• Empty file-slice providers are skipped instead of scanning the full repository.
• synthesis.success now correctly reflects narrative provider status.
• max_provider_parallelism config file fallback is restored.
• --divide dimensions now assigns after provider filtering.
• OllamaAdapter no longer hijacks command+model agents.
• Incomplete agents without command or model are filtered from the registry.

[0.8.0] - 2026-03-17

Added

• --chain mode for sequential multi-agent analysis.
• --perspectives-json for per-provider review focus.
• Consensus badges ([N/M agree] and chain-specific confirmed-by labels) in human-readable outputs.
• Session result retrieval for queued or asynchronous session runs.

Changed

• Session retry behavior now classifies errors and applies exponential backoff.

Fixed

• Fixed a critical --no-wait session data-loss issue.

[0.7.0] - 2026-03-17

Added

• --file and stdin prompt ingestion for run / review.
• Temporary --agent registration for custom ACP-compatible agents.
• --quiet output mode and config-file loading support.
• mco session ensure, --no-wait, and Ctrl+C-aware session cancellation improvements.
• ACP expansion with structured rendering plus bidirectional filesystem and terminal handlers.

Fixed

• Fixed allow_paths passthrough, deep config merge behavior, and empty-stdin rejection.
• Hardened ACP permission keys and launch-flag handling.
• Improved thread safety around pending session state and streaming buffers.

[0.6.0] - 2026-03-16

Added

• Stateful multi-turn sessions via mco session.
• Session prompt queueing with cancellation support.
• ACP (Agent Client Protocol) transport layer.
• Data-driven ACP protocol conformance tests.

[0.5.0] - 2026-03-16

Added

• Diff-only review mode with --diff, --staged, and --unstaged.
• mco serve MCP server mode.
• Structured streaming via --stream jsonl.

[0.4.0] - 2026-03-12

Added

• Cross-session memory bridge with --memory and --space.
• Passive confirmation, forget-cleaner, and confidence-scored finding persistence.
• Agent reliability scoring, task classification, and tech-stack priors.
• mco findings and mco memory subcommands.

Fixed

• Correctness fixes for canonical latest-view deduplication, category-aware priors, memory_id propagation, and per-finding changed-file tracking.

[0.3.5] - 2026-03-10

Changed

• Default providers list now includes all 5 supported providers (claude, codex, gemini, opencode, qwen) instead of only claude and codex.
• Claude adapter: permission mode plan → bypassPermissions for full tool execution.
• Codex adapter: sandbox workspace-write → danger-full-access for full filesystem and network access.
• Gemini adapter: added -y (YOLO mode) for automatic tool approval in non-interactive mode.
• Qwen adapter: added -y (YOLO mode) for automatic tool approval in non-interactive mode.

[0.3.3] - 2026-02-27

Added

• Added mco doctor command with human-readable and --json outputs to probe provider binary/auth readiness.
• Added --format markdown-pr (review-only) to render PR-ready Markdown summaries from aggregated findings.
• Added opt-in --include-token-usage to include best-effort provider token usage and aggregate token summary in outputs.
• Added --format sarif (review-only) to emit SARIF 2.1.0 output for code scanning integrations.
• Added opt-in synthesis pass via --synthesize and --synth-provider, returning structured synthesis output (consensus/divergence/next steps) in JSON and artifacts.

Changed

• Added deterministic cross-provider findings deduplication in review aggregation and findings.json, with merged detected_by provenance and max-confidence rollup.

[0.3.2] - 2026-02-27

Changed

• Added run-mode answer extraction fields per provider: final_text, response_ok, and response_reason, while keeping output_text as raw output for debugging.
• Improved final_text extraction quality for event-stream outputs by preferring high-signal answer candidates over trailing low-signal tokens.

[0.3.1] - 2026-02-27

Changed

• Made stdout mode truly non-persistent by default: no artifact files are written unless --save-artifacts or --result-mode artifact/both is used.
• In stdout mode without artifact writes, artifact_root and provider output_path now return null.
• Unified adapter detect/probe binary resolution and environment handling with runtime execution (shutil.which + sanitized env) and refined auth probe reason classification (auth_check_failed, probe_config_error, probe_unknown_error).

[0.3.0] - 2026-02-27

Changed

• Disabled runtime idempotency/dispatch cache replay; repeated invocations now always re-execute providers.
• Extended stdout payloads and human-readable output to include full per-provider output text (not only excerpt).
• Removed legacy idempotency/state/cache knobs and fields (--idempotency-key, --state-file, created_new_task, deduped_dispatch, dispatch_key).

[0.2.1] - 2026-02-26

Changed

• Changed default CLI delivery mode to stdout-first (--result-mode stdout) so agent callers receive results directly without mandatory artifact reads.
• Added --save-artifacts to explicitly persist artifact files while keeping stdout result delivery.
• Updated benchmark script to explicitly opt into artifact persistence (--save-artifacts).
• Repositioned README (EN/CN) messaging around "Any Prompt. Any Agent. Any IDE." and clarified caller-agent orchestration scenarios.

[0.2.0] - 2026-02-26

Changed

• Removed config-file mode from CLI; mco now uses built-in defaults with flag-only overrides.
• Removed --config from mco run / mco review; passing it now errors as unsupported.
• Updated benchmark automation to run without config files and to report provider set directly.
• Updated README (EN/CN) to document zero-config usage with CLI flag overrides only.

Removed

• Removed config file loading path (load_review_config) and related YAML/JSON config parsing.
• Removed sample config files (mco.example.json, mco.step3-baseline.json).

[0.1.3] - 2026-02-26

Added

• Added full Simplified Chinese README (README.zh-CN.md) with language switch links.
• Added environment sanitization for provider subprocesses to strip CLAUDECODE.

Changed

• Aligned review findings schema and parser contract by making evidence.line and evidence.symbol optional keys.
• Clarified installation channels in docs: npm available now, PyPI pending Trusted Publisher setup.

Fixed

• Implemented real retry backoff sleep in runtime retry loop.
• Released adapter run handles in terminal and cancel paths to avoid in-memory handle growth.
• Switched CLI config parsing to fail-fast for invalid --provider-timeouts and --provider-permissions-json.

[0.1.2] - 2026-02-26

Added

• Added packaging metadata (pyproject.toml) and mco console entrypoint.
• Added npm wrapper package (@tt-a1i/mco) for Node-based environments.
• Added publishing workflows for PyPI and npm.

Changed

• Updated repository naming and distribution identity to mco.
• Updated README and release docs for install and usage guidance.

[0.1.1] - 2026-02-26

Added

• Added provider permission contract docs.
• Added release governance artifacts (CODEOWNERS, release notes updates).

Fixed

• Hardened npm publish workflow behavior when NPM_TOKEN is missing.
• Fixed npm workflow syntax/guard issues for reliable CI execution.

[0.1.0] - 2026-02-26

Added

• Initial runnable runtime for multi-provider orchestration (run and review commands).
• Provider adapters for claude, codex, gemini, opencode, and qwen.
• Progress-driven timeout handling, retry semantics, idempotent dispatch, and notification dedupe.
• Canonical findings normalization, review decisioning, and artifact outputs (summary.md, decision.md, findings.json, run.json).
• Runtime gate, adapter contract tests, and benchmark/probe scripts.