Skip to content

CHANGELOG.md

Latest commit

 

History

History
140 lines (112 loc) · 10.9 KB

CHANGELOG.md

File metadata and controls

140 lines (112 loc) · 10.9 KB

Changelog

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

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

Unreleased

0.5.0 - 2026-04-20

Added

  • Supervisor evaluation loop — the Supervisor is now an active quality gate, not advisory only. When it rejects an agent output (approved: false), the orchestrator re-invokes the agent with the flags and recommendation appended as feedback, up to COUNCIL_EVAL_RETRIES times (default: 2 → 3 total attempts). If the retry budget is exhausted and the output is still flagged, the result is surfaced anyway with the flags visible — no silent passes.
  • COUNCIL_EVAL_RETRIES env var — clamped to [0, 5]. Set to 0 to restore pre-0.5 advisory-only behaviour. Non-integer values fall back to the default.
  • Retry count recorded in session.metrics.eval_retries, visible via get_council_state and in the result summary footer.
  • supervisor_feedback field added to AgentInvokeOptions; Executor and Aide prompts render a clearly-delimited --- SUPERVISOR FEEDBACK --- block when set.
  • First test suite for the repo — 160 tests across 15 files covering config, schemas, stores, agent invokers, the feedback formatter, the eval loop, orchestrator routing, and integration tests for FileStore and SQLiteStore against real temp dirs.
  • CI now runs npm test on every push (new Test job in ci.yml) and as part of PR validation (pr-check.yml).

Fixed

  • Spurious first-call INVALID_JSON_RESPONSE failures — the claude CLI occasionally emits JSON wrapped in prose ("Here's the output: {...} Let me know!") or produces a single flaky sample (missing field, wrong enum) on the first try. A strict JSON.parse + schema check rejected both, surfacing as INVALID_JSON_RESPONSE before the Supervisor evaluation loop ever ran. Fixed with two new layers:
    • parseAgentJson extracts JSON via fence-match → balanced-brace extraction (string-aware, so braces inside JSON string values are not mis-counted) → raw trimmed, returning the first candidate that parses.
    • runAgentWithValidation wraps runAgent + parse + Zod validation with a one-shot retry on parse/validate failure. Transient flakiness recovers silently; CLI transport errors still propagate unchanged.

Changed

  • All 4 agent invokers (Chancellor, Executor, Aide, Supervisor) now route through runAgentWithValidation — removes ~80 lines of duplicated fence-match + parse + validate boilerplate. External behaviour is unchanged on success; failures are rarer and still produce the same CouncilError code (INVALID_JSON_RESPONSE for Chancellor/Executor/Aide, SUPERVISOR_ERROR for Supervisor) on final give-up.
  • FileStore and SQLiteStore accept a storage path via constructor argument. Default is still ~/.council/sessions and ~/.council/council.db respectively; the parameter exists so integration tests can use temp dirs instead of polluting the user's home directory.

0.4.0 - 2026-04-20

Added

  • Per-agent tool access — Chancellor and Aide now have tool access, configurable via AGENT_TOOLS constants:
    • Chancellor: Read, Glob, Grep (read-only — inspects codebase before planning, never writes)
    • Executor: Read, Write, Edit, Bash, Glob, Grep (unchanged)
    • Aide: Read (can read files before transforming them)
    • Supervisor: none (pure review, no side effects)
  • AGENT_TOOLS constant in src/domain/constants/index.ts — single source of truth for all per-agent tool sets
  • Chancellor and Aide system prompts updated to describe their tool capabilities and when to use them
  • Removed runExecutorWithTools convenience wrapper — Executor now uses runAgent + AGENT_TOOLS.EXECUTOR directly, consistent with all other agents
  • Caveman token compression — reduce internal agent output tokens by 50-60% with no accuracy loss. Inspired by Caveman. Set COUNCIL_CAVEMAN in the MCP env block:
    • off (default) — no compression, unchanged behaviour
    • lite — drops filler and pleasantries, keeps grammar (~20% savings)
    • full — fragments, flat bullets, explicit 50% word budget (~50-60% savings, recommended)
    • ultra — telegraphic abbreviations and symbols (~60-70% savings)
  • Compression applies to Chancellor, Executor, and Aide. Supervisor is exempt — its recommendation field is user-facing prose.
  • Active mode recorded in session.metrics.caveman_mode, visible via get_council_state

0.3.0 - 2026-04-18

Added

  • Persistent session memory — sessions now survive MCP server restarts via an optional COUNCIL_PERSIST env var:
    • memory (default) — in-process LRU Map, no breaking change
    • file — JSON files at ~/.council/sessions/<id>.json, zero dependencies
    • sqlite — SQLite at ~/.council/council.db via better-sqlite3, WAL mode for safe concurrent access
  • SessionStore interface — all backends implement a common contract; swappable without touching orchestration code
  • 7-day session TTL — file and SQLite backends auto-expire sessions older than 7 days on startup
  • Startup validation — unknown COUNCIL_PERSIST values emit a warning and fall back to memory

0.2.3 - 2026-04-13

Changed

  • README diagrams replaced with hosted images — architecture, orchestration flow, and session lifecycle diagrams now render correctly on all platforms including GitHub, npm, and PyPI mirrors

0.2.2 - 2026-04-13

Fixed

  • install.sh now also runs claude mcp add to register the server with Claude Code CLI — previously only Claude Desktop was configured, so the tools were invisible in the CLI
  • install.ps1 rewritten to match: detects claude binary, falls back to ANTHROPIC_API_KEY, configures both Claude Desktop and Claude Code CLI
  • runner.ts strips ANTHROPIC_API_KEY when set to an empty string — Claude Desktop injects an empty key into the MCP server env, causing the child claude process to attempt API key auth and fail with exit 1

0.2.1 - 2026-04-12

Fixed

  • Replace @anthropic-ai/claude-agent-sdk with direct claude CLI subprocess calls — eliminates 401 auth errors for users authenticated via Claude.ai OAuth (no separate API key needed)
  • Use --system-prompt-file instead of --system-prompt CLI arg — prevents exit 1 failures caused by long system prompts with XML tags and special characters
  • Startup check fails fast with a clear message if claude CLI is not in PATH and no ANTHROPIC_API_KEY is set
  • install.sh and install.ps1 now detect the claude binary location and add its directory to the MCP server PATH automatically

0.2.0 - 2026-04-12

Added

  • Supervisor agent (Claude Haiku 4.5) — reviews every Executor step result and Aide task output before they surface to the caller. Non-blocking: if the Supervisor errors, orchestration continues and a warning is logged.
  • get_supervisor_verdicts MCP tool — retrieve all Supervisor verdicts for a session, with optional flagged_only filter for quick triage
  • Supervisor flags in result summary — flagged outputs appear under a ## Supervisor Flags section in the orchestrate result
  • supervisor added to AgentRole — session metrics now track Supervisor invocations alongside Chancellor, Executor, and Aide
  • PR template — standardised pull request checklist for contributions
  • Issue templates — bug report and feature request templates; security reports redirect to GitHub Security Advisories
  • CODEOWNERS — @iamvirul set as required reviewer on all files

0.1.2 - 2026-04-12

Fixed

  • Release workflow now reconfigures npm registry to npm.pkg.github.com before publishing to GitHub Packages, fixing ENEEDAUTH on the GitHub Packages publish step

0.1.1 - 2026-04-12

Added

  • Published to GitHub Packages as @iamvirul/council-mcp in addition to npm — package now appears in the GitHub repository sidebar

Security

  • Zod runtime schema validation on all agent JSON responses — prevents malformed or injected agent output from propagating to downstream agents
  • Hard cap of 10 delegated tasks per Executor response — prevents prompt-injection-driven Aide invocation amplification
  • UUID format validation on session_id and task_id MCP tool inputs
  • Max length limits added to context and plan_context tool input fields (previously unbounded)
  • Code fence extraction regex corrected — non-greedy match prevents incorrect JSON extraction from multi-fence responses
  • Silent catch {} blocks replaced with logger.warn — state recording failures now visible in logs
  • Pino async destination flushed on beforeExit, uncaughtException, and unhandledRejection — prevents log loss on crash

0.1.0 - 2026-04-11

Added

  • MCP server with five tools: orchestrate, consult_chancellor, execute_with_executor, delegate_to_aide, get_council_state
  • Chancellor agent (Claude Opus 4.6) — strategic analysis, risk assessment, and step-by-step planning via the Agent SDK
  • Executor agent (Claude Sonnet 4.6) — plan implementation with access to Read, Write, Edit, Bash, Glob, Grep tools
  • Aide agent (Claude Haiku 4.5) — simple tasks: formatting, data transformation, utilities
  • Complexity-based routing in orchestrate — trivial problems go to Aide, simple to Executor, complex through the full Chancellor → Executor → Aide pipeline
  • In-memory session state with LRU eviction cap of 500 sessions to prevent OOM
  • Structured JSON logging via pino to stderr (stdout reserved for MCP JSON-RPC)
  • Graceful shutdown on SIGINT/SIGTERM
  • GitHub Actions workflows: CI (type-check + build + audit), PR check, and release (GitHub Release + npm publish with provenance)
  • MIT license

Security

  • All logs routed to stderr — MCP stdout never contaminated
  • Stack traces never exposed to MCP tool callers
  • Session IDs generated with crypto.randomUUID()
  • Executor runs with explicit permissionMode: 'acceptEdits' rather than relying on inherited default
  • @anthropic-ai/claude-agent-sdk pinned to ^0.2.101 (no latest in production)