v0.14.0

What's New

• extraArgs support in tools.json — append shell-quoted CLI flags per role via an optional extraArgs array on each tool. Use for per-role MCP server wiring in /converge, letting implementer and reviewer load different configurations without duplicating the command.

v0.13.0

What's New

• /converge command for stateful multi-model deliberation — Claude drafts, Codex challenges, coordinator tracks issues across bounded rounds until resolution or cap.
• Canonical issue ledger owned by the coordinator prevents either model from manufacturing convergence.
• Stop conditions: no-open-blockers, stable-with-disputes, or round cap (default 3 rounds).
• Stalemate detection auto-closes disputed nits and concerns after one re-raise; blockers on correctness never auto-resolve.
• Strict JSON wire format for model outputs — coordinator validates schema and rejects prose.
• Session state written to disk — every prompt, artifact, critique, and ledger snapshot is auditable.
• MVP binds Claude as implementer and Codex as reviewer; tool swapping planned for phase 2.

Improvements

• README reframed around compare/converge mental model instead of shell-script positioning.
• Added "What it's not" section to distinguish Conclave from agent-soup frameworks, hosted orchestrators, and always-on services.
• Configuration, custom commands, and engine details moved to docs/advanced.md so README stays focused on first success.
• docs/converge.md design document covers wire formats, ledger schema, fingerprinting rules, round semantics, and stop conditions.

v0.12.0

What's New

• Core engine extracted — Conclave is now a single shell script (conclave-run.sh) that reads config, runs tools in parallel, and returns JSON. Everything else is an optional example.
• Commands moved to examples — /review and /consult are now example implementations you can fork and customize instead of opinionated built-ins.
• Agents simplified — Example agents reduced from ~650 lines to ~126 lines total. Multi-model-executor replaced with direct background Bash.
• Config-driven — /consult now uses config instead of hardcoded models. Add tools to ~/.config/conclave/tools.json and they're available everywhere.
• Explicit input field — Tools can now declare "input": "argument" in config if they take the prompt as an argument instead of stdin.
• Separate registration — bun run register installs core only. bun run register:agents adds optional investigators.
• Semantic-release AI notes — Automated AI-generated changelogs on release.

Improvements

• /review simplified from 626 lines to 221 lines.
• /consult simplified from 224 lines to 123 lines.
• Example config reduced from 16 tools to 3, easier to understand and extend.
• Landing page updated with "build your own" section and changelog link.
• Installation and customization docs clarified.

What's Different

Conclave is no longer an opinionated code review tool. It's a consensus engine — use it however you want. The old command and agent implementations live in examples/ for reference. Fork them, simplify them further, or build something new. The only required part is the engine script.

v0.11.1

Improvements

• Background execution fix — replaced executor sub-agent with direct background Bash calls for more reliable parallel tool execution
• Foreground parallelism — switched from background+TaskOutput pattern to parallel foreground Bash calls, fixing output capture reliability

v0.11.0

What's New

• Codex GPT-5.3 — upgraded the default Codex model to gpt-5.3-codex with xhigh reasoning effort for deeper reviews

v0.10.2

Improvements

• Ollama runs directly — switched from Claude Code proxy to ollama run directly, cutting out the middleman
• Executor sub-agent — tool spawning is now delegated to the multi-model-executor for cleaner parallel execution

v0.10.1

Fixes

• Direct env var approach — replaced the Claude proxy method with direct environment variable injection for Ollama cloud models. Simpler and more reliable

v0.10.0

What's New

• Ollama cloud via Claude proxy — Ollama cloud models now run through a Claude Code proxy for better authentication handling

v0.9.1

Fixes

• Reverted inline Claude review — the v0.9.0 approach caused session conflicts. Reverted to external CLI spawning with CLAUDECODE=0 to allow nested Claude sessions cleanly

v0.9.0

What's New

• Claude as internal reviewer — Claude now performs its own review inline instead of spawning as an external CLI tool, reducing overhead and improving context sharing