feat: kyora council — summon other model families over MCP

[eliahilse]

Your agent is one lineage with one set of blind spots. This gives it a way to ask the others — mid-work, on the subscriptions you already pay for.

The MCP server

tool
council_convene	one question → several model families in parallel, independent takes back
council_convene_async	same, backgrounded; collect with council_result while you keep working
council_ask	one specific family's perspective, read-only
council_task	delegate concrete work to one family — write: true permits file edits
council_result	collect a background council or task
council_status	who can be summoned now, with live remaining quota

Backed by the same engine table as kyora-review, so every member is its vendor's own CLI under your own login. Seating is quota-aware — highest remaining quota first, never an engine that's cooling down or spent, and size caps a council so one question doesn't drain every subscription. The prompts push members to disagree rather than defer, and the summary tells the caller to treat disagreement as the signal rather than averaging votes.

Engines gained write-capable invocations for delegation (codex workspace-write, claude/kimi/glm/qwen with edit tools restored but CI suites and destructive git still denied). The review package now exports ./engines, ./usage, ./extract, ./types.

The high-stakes watcher

Optional PostToolUse hook: a cheap model reads the transcript tail and the pending action, and when it judges the moment genuinely consequential — schema migration, auth change, force push, irreversible op — it injects a note nudging the agent to consult the council before committing.

Built to be cheap and quiet, since a noisy or expensive watcher is one you'd turn off: a keyword/size pre-filter runs in-process so most tool calls never reach a model, classification is rate-limited (2m default), nudges harder (15m default), and every failure path exits silently so it can never block or break the agent. Model is any OpenAI-compatible endpoint — defaults to glm-4.5-air or qwen3.6-plus depending on which key is present.

Verification

• MCP handshake + all six tools registered (stdio round-trip)
• Live council with glm + grok on a real design question in this repo — both read the actual source and cited specific lines; seating correctly ordered glm (99% quota) ahead of claude (41%)
• 6 unit tests for the watcher's pre-filter and verdict parsing; types clean across the workspace

Follow-ups (not in this PR)

• Kimi still needs a Moonshot key to validate its probe and seat
• The watcher classifies text only; wiring vision for screenshot-bearing transcripts is a natural next step

🤖 Generated with Claude Code

https://claude.ai/code/session_01CQy3ZJ5MyUxo4qjZwA93Na

[eliahilse]

Added the plain subagent shape plus model/effort control:

• agent_spawn — delegate to another model family without naming one; the least-spent capable family is chosen automatically (prefer is a hint, ignored when that family has no quota). Read-only by default, write: true for edits, background: true for a job id.
• agent_fanout — several different tasks in parallel, one per family, so independent workstreams spread across subscriptions instead of draining one. Assignment logic is pure and unit-tested (pinned-first, one family each, reuse only when families run out, and a task pinned to an unavailable family is reported rather than silently rerouted).
• council_models — lists which families are summonable and the models/efforts each accepts.
• model and effort on every delegation tool, applied per engine as CLI flags (codex -m + -c model_reasoning_effort, grok -m + --reasoning-effort, claude --model) or the model env var for endpoint-routed families. A caller-supplied effort replaces the engine's built-in flag rather than duplicating it.

Two bugs this surfaced, both fixed:

1. Council calls inherited the review path's --output-schema and passed it an empty schema — codex 400'd on every ask. Added a chat mode so free-form answers use unconstrained invocations.
2. The codex model list I first wrote was invented and rejected by the account. It now reads the model from your own ~/.codex/config.toml (here: gpt-5.6-sol), and the other families list only ids actually run in testing.

Verified live: codex chat with effort: low answers correctly in 10s, glm honors model: glm-5.2 through the env route, all 9 tools register over stdio, 11 unit tests green.