[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7C9WEDREBPGF8DMDP887YP1":3,"skill-same:01M7C9WEDREBPGF8DMDP887YP1":61},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":32,"commitSha":33,"skillDirectory":34,"forksCount":14,"repositoryHtmlUrl":35,"repositoryHomepageUrl":36,"ownerHtmlUrl":37,"ownerBlogUrl":38,"ownerDescription":39,"updatedAt":40,"install":41,"resources":54},"01M7C9WEDREBPGF8DMDP887YP1","council",null,"Convene configured Lite, Standard or Strict council review; the owning writer records exact-subject approval evidence.","bionic-coding\u002Fcrux","bionic-coding","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F301040008?v=4","01M7C9T2NA1V1V0EF1M552HRB3","Bionic Coding","01M7C9T2P72H1AQV5T5PTBFNHN",0,"MIT","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Fcouncil\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: council\ndescription: \"Convene configured Lite, Standard or Strict council review; the owning writer records exact-subject approval evidence.\"\ncontext: fork\nmodel: opus\nmetadata:\n  tags: \"verification, multi-model, council, anti-hallucination\"\n  bundles: \"crux-verification, crux-docs\"\n  risk_level: \"medium\"\n  requires_access: \"write-run\"\n  triggers: \"run the council | convene the council | ask the council | get a multi-model opinion\"\n---\n\n# Crux Council Vote (Async-First)\n\n> **Execution context:** this skill runs in a forked subagent and returns a summary to the caller. The fork does not see the main-thread conversation, so pass any needed context explicitly at invocation.\n\n\u003C!-- BEGIN GENERATED: runtime-compat -->\n## Runtime compatibility\n\nThis skill is portable across Claude Code, Codex, OpenCode, and OMP (Oh My Pi). This section overrides platform-specific labels below.\n\n- Before running a command that uses `CRUX_PLUGIN_ROOT`, set it to the installed plugin root. In Claude Code, use the value of `CLAUDE_PLUGIN_ROOT`. In Codex, OpenCode, and OMP, derive it from the absolute path of this selected `SKILL.md`: the plugin root is the parent of its `skills\u002F` directory (in OMP `bash`, `skill:\u002F\u002F\u003Cname>` resolves to the skill directory). In a source checkout, use the checkout `crux\u002F` directory.\n- For project-local skills, use `.claude\u002Fskills` in Claude Code, `.agents\u002Fskills` in Codex, `.opencode\u002Fskills` in OpenCode (which also reads the singular `.opencode\u002Fskill`), and `.omp\u002Fskills` in OMP. Set `CRUX_LOCAL_SKILLS_DIR` to that path before following any command below that uses it.\n- Translate Claude Code tool labels such as `Agent`, `Read`, `Write`, `Bash`, `WebSearch`, and `WebFetch` to the matching capability in the current session. Codex names its own capabilities; OpenCode uses the lowercase forms `subagent`, `read`, `edit`, `shell`, `websearch`, and `webfetch`, where `edit` covers both `Edit` and `Write`; OMP uses `task`, `read`, `edit`, `write`, `bash`, `grep`, `glob`, `find`, `todo`, and `web_search` (where `read` also handles URLs and `skill:\u002F\u002F` URIs). Do not attempt to invoke the Claude Code labels as literal commands on another host.\n- Install the generated role agents before delegating: `install-codex-agents` in Codex, `install-opencode-agents` in OpenCode, `install-omp-agents` in OMP. Codex names them `crux_architect`, `crux_brainstormer`, `crux_commander`, `crux_dev_lead`, `crux_developer`, `crux_historian`, `crux_librarian`, `crux_night_gardener`, `crux_reviewer`, and `crux_wayfinder`; OpenCode and OMP use the bare role names `architect`, `brainstormer`, `commander`, `dev-lead`, `developer`, `historian`, `librarian`, `night-gardener`, `reviewer`, and `wayfinder` (in OMP, `scout` may also be used for reconnaissance alongside `wayfinder`). If a required role or capability is unavailable, report that truthfully instead of claiming it ran.\n- Argument placeholders such as `$adr` and `$book` bind only in Claude Code. On a host without argument binding they are unset — take the value from the user's phrase. The \"Fields OpenCode ignores\" section of `OPENCODE_GUIDE.md` names the invocation-control fields OpenCode ignores.\n\u003C!-- END GENERATED: runtime-compat -->\n\n## Configured councils\n\nRead `.bionic.yml` through the canonical policy reader. Council mode and human\nacceptance are independent. Existing projects without a council setting retain\nStrict. New setup explicitly selects Standard or an owner-selected Lite mode.\n\n- Lite: native configured Apex challenge, architect response, fresh final review.\n  Final reviewer approval and exact-final architect agreement are both required.\n  No OpenRouter account is required. Follow [native-lite.md](references\u002Fnative-lite.md).\n- Standard: three gateway provider seats, two approvals and architect agreement.\n  Address every dissent with a typed disposition. Resolved\u002Frefuted findings need\n  retained checkable evidence; a non-binding trade-off needs an explanation.\n  Single-seat dissent rule: when two seats approve and one dissents, revise\n  only if the finding is a binding failure or missing authorization, or states\n  a plausible failure, consequence and smallest fix that the architect cannot\n  refute and no later phase would catch. For an unresolved required safety\n  fact, obtain a targeted check; if it stays unresolved, withdraw the dependent\n  choice. A required safety fact about data policy or publishing is not a council\n  dispute: it is the outward-facing stop, and the one question put to the owner.\n  Otherwise answer on the same revision with reasoned agreement. Record `refuted` with a named check\n  when the finding fails the blocking standard. Record `accepted-tradeoff` with\n  the reason when a later phase would catch it or its risk is lower than the\n  cost of another round. When all three seats approve with nits only, record\n  agreement; do not revise for nits.\n- Strict: all three gateway seats approve with no recorded blocking findings.\n\nStandard and Strict require `OPENROUTER_API_KEY` through `crux-env`. Missing\ncredentials or native capabilities refuse with their cause. Never change mode,\nprovider, model or seat count as a fallback. Demonstrated binding failures,\nmissing authorization and unresolved required safety facts prevent approval.\nA minority label alone is not a Standard veto.\n\nUse `cycle-council.py` for ADR, IDR, ADR-creation, consultation and disposition\npurposes. Consultation approves no decision. Its `--identity` stays stable across\nrenamed files. IDRs retain their cycle name across revisions. The configured\nattempt limit defaults to seven for new mode-aware decisions. Stop on stalled\nprogress. A council dispute never goes to the owner. At the limit, or when the\nsame blocking finding recurs on two consecutive attempts after a revision that\nanswered it, the architect withdraws the decision with `withdraw-decision`;\nno agent requests an extension. A Standard-mode single-seat dissent that the\narchitect disposes as `accepted-tradeoff` on an approved attempt is answered,\nnot unanswered; it does not count toward the recurrence trigger, and neither\ndoes its repetition on a later attempt over the same subject bytes. A finding\ncounts toward recurrence only when it blocks approval of an attempt. A narrowed replacement names the record it\nreplaces, and its attempts continue that record's count. An extension or mode\nchange the owner records on their own initiative keeps its provenance, findings\nand total count. Historical reviews keep their original policy and limit.\n\nGateway convening is detached by default. Initial success names a running attempt\nand `await_command`; it is not a vote or approval. Run the returned await command\nfor the terminal outcome, then the existing architect disposition. `--foreground`\nis the synchronous diagnostic form. `--probe ATTEMPT` reads the attempt flock.\n`--recover ATTEMPT` sends no request: it refuses while the flock is held, recovers\nthe original persisted result, or marks an unowned execution interrupted. An\ninterrupted claim still counts, and a later convening is a separate counted attempt.\nFull cycle inspect exposes exact subject\u002Fattempt\u002Fresult references, recovery argv\nand prerequisites without requests or writes. New gateway attempts always persist\nrequest\u002Foutcome checkpoints under existing ownership. A persistence failure stops\nfurther calls and positive approval. Durable negatives survive; missing outcomes\nstay unknown. Recovery sends no request and resets no budget. Partial events never\napprove; later convening uses prior findings as historical context beside fresh votes.\nNew approval\u002Fapplication checks current bound evidence. Pure sealed historical\nvalidation and applied acceptance retain their standing. The cycle commands reference\ngives the interfaces; historical\u002Fnative-Lite callers keep their original protocols.\nThe runner holds ownership through result persistence; never hand-edit or delete\na claim. A missing launcher acknowledgement proves nothing: after the exact child\nterminates, check its exact preallocated claim identity before reporting no claim.\nUse recovery for an existing claim; do not request a substitute owner approval.\nNative Lite follows its separate sealed-dispatch procedure.\n\nSettle a dispute under the configured mode. Record the architect agreement even\nwhen every seat approves: it names the approved sha256 and each nit id with a\ndisposition, `apply-in-build` (a note on a Build task) or `declined` with its\nreason, in the `nits` list of the `--architect` JSON. A nit never starts another\nattempt, and the approved revision stays as approved. `srde` output may be kept as\nevidence for a `refuted` disposition of an IDR dissent; `srde` never runs on the\nADR path.\n\nA new revision of an IDR whose earlier revision was approved is reviewed as a\ndelta. The writer computes each top-level contents key added, removed or changed\nagainst the latest approved revision, stores it in the attempt, and checks it\nagain against both revisions' bytes at convening and at approval. The seats\nreceive the full new revision with the changed keys marked and judge the change.\nApproval binds the new revision's sha256, and a delta attempt counts toward the\nsame limit.\n\nA Standard result waits for the architect even when its provider votes suffice.\nThat result reports the outcome `awaiting-architect`, exits 1 and prints the command\nto run; the sealed record keeps `changes-required`. Use `--architect PATH --record\n\u003Cresult>`, where PATH names a JSON file, to record exact-revision agreement and\ntyped dissent; do not call another council just to add that agreement. The writer\nresolves evidence artifacts safely and hashes their bytes. A targeted check for a\nrequired fact is not another mandatory council round.\n\nFix-round discipline, for every mode: fix all findings (or refute one with a named check) in one revision, apply each\nfix at every site that states the same concept, and rerun the pre-council\nself-check's mechanical checks before reconvening.\n\nThe following promptbook procedures preserve historical gateway-only behavior.\nThey do not override configured prospective modes.\n\n## Gate council or ad-hoc council\n\nCouncil deliberation runs only through the council runner, and its council record is the only evidence a council gate accepts.[^council] Independent review is a reviewer's examination, and its reviewer report is the only evidence an independent-review gate accepts.[^review] Neither satisfies the other's gate.\n\nInside a promptbook run, a council gate uses the council runner, `run-council.py`. Run it as a background command:\n\n```bash\nuv run \"${CRUX_PLUGIN_ROOT}\u002Fscripts\u002Frun-council.py\" \u003Crun-RUN-NNN.yaml> [--prompt N] --round N \\\n  --question \u003Cquestion file> --subject \u003Csubject file> [--subject \u003Csubject file> ...] \\\n  [--book \u003Cbook file>] [--retain-subjects] [--max-tokens 64000] [--timeout 600]\n```\n\n`--max-tokens` is the first-attempt output budget for every seat (default 64000). A seat\nwhose reply stops on that budget is retried once at its own output limit from the model\nregistry. A value above the smallest selected seat limit is refused before any call.\n`--timeout` is the per-seat-call deadline in seconds (default 600). When a reasoning seat\ntruncates, pass `--timeout 1200`: the escalated retry under 600 s is likely to time out.\n\nFinish any open Git merge or other sequence before council. Both book formats\nuse the current attempt-aware gate. An implementation module uses council kind\n`implementation`, including in a combined architectural book; `combined` is never\na council kind. Verify and patch retain their existing kinds and dimensions.\nSelect an exact formal revision with `--implementation-revision \u003Cdecision-path>`.\nAlso pass that exact path as `--subject` with `--retain-subjects`. Record each subject's\nwrite-time witness before committing it; never attribute dirty bytes after a refusal.\n`--migration-batch` is retired: it exits 2 with `historical-migration-writes-retired`\nbefore any batch is loaded, any attempt is claimed or any council call is made. Propose\nclause migrations in the prospective format through `migrate-adrs`.\nNew formal closes retain context three\u002Fprofile four. Historical context two\u002Fprofile three\nand context one\u002Fprofile two remain immutable and replay-only, never selected from the book format.\n\n`--prompt` is optional and defaults to the run's current prompt. The run, `--round`, `--question` and at least one `--subject` are required.\n\n`--round` is the number of council records with outcome `ran` already in this module (in a patch book, this prompt), plus one. In an adr module a valid adjudicator refutation record fills place 3. A could-not-run record takes no place, so after you fix its cause you reconvene with the same number. `run-council.py` refuses a mismatch before any call: exit 1, a JSON error on stdout, `\"record\": null`, and no record written.\n\nThe council runner commits the attempt record and the council record itself. Before any council request it commits an attempt record that claims the round. The claim begins a council attempt, one execution of the round, which is distinct from a seat's calls below. A round that began never vanishes: until its own council record resolves it, the gate stops on the open attempt.[^attempt] It verifies that the committed council record holds the bytes it computed.[^match] The gate check (`advance-run.py`) reads only a committed record. Do not commit, edit or delete either record yourself.\n\nThe council runner exits with one of three codes. A council runner that ends without an exit code, with a signal or with a code other than 0, 1 and 2 stopped before it reported. Treat it as an exit 2 that names a claimed attempt, and follow the steps under `2` with `--prompt \u003Cn>`; never convene another round until recovery reports.[^recover]\n\n- `0`: it wrote, committed and verified a council record that ran.\n- `1`: it committed a council record the gate check stops on, or it refused the call and wrote no council record. The record is a could-not-run record, whose action is DEFER_TO_HUMAN, or a council record that ran but whose full write the secret scan refused: advance `--outcome blocked` with it attached. A refusal prints `\"record\": null` and leaves nothing to attach. `\"refused\": \"round\"` means correct `--round` and run again. `\"refused\": \"prompt\"` means `--prompt` is not the run's current prompt: run again with the current prompt or omit `--prompt`. `\"refused\": \"preflight\"` means a question or subject failed the input checks before any request: apply each cause's `repair` and run again at the same round. `retype` means correct the path. `commit-run-work` means commit the subject, the run's own witnessed work, with `run-work-witness.py commit \u003Crun> --path \u003Cpath>`. The third preflight refusal at one convening prompt commits a `preflight-retries-spent` record instead, an escalation-loop stop.[^defer] `\"refused\": \"attempt-open\"` means an open attempt holds the module: recover it, never claim past it.\n- `2`: no council record was committed. Exit 2 never means nothing was written. When stderr names `timeout`, or names outside work the commit moved, the owner's remedy comes before recovery: report a contradicted-premise stop for the owner. The owner first restores the set-aside work (`git stash list`, or a pre-commit framework's backup patch) and then removes a stale `index.lock` in the git directory. Start recovery only after the owner reports both steps done. When stderr names a claimed attempt, run the process check, probe the lock with `run-council.py --recover \u003Crun> --prompt \u003Cn> --probe`, and once no live council runner holds it, run `run-council.py --recover \u003Crun> --prompt \u003Cn>`. Pass `--prompt`: without it recovery cannot recognise a record whose pending copy is already removed, and it reports `nothing-open`. Recovery makes no council request and commits or recognises only the original council record.[^recover] Route each recovery result by the recovery table in `run-promptbook`'s `references\u002Fgates.md`. Never convene another round over a claimed attempt. When stderr names no attempt, the prompt cannot advance, and the conductor reports why, as a contradicted-premise stop for the owner.\n\nA step that writes a file it later passes as a subject records a run-work witness after each write: `run-work-witness.py record \u003Crun> --prompt N --path \u003Cpath>`. Never record a witness to repair a preflight refusal: a subject with no witness taken at its write goes to the owner. `run-promptbook`'s `references\u002Fgates.md` describes how a conductor issues and advances a gate prompt, repairs a preflight refusal and recovers an open attempt.\n\nThe async driver below serves an ad-hoc council only, such as a design question outside a run. It satisfies no promptbook council gate. Outside a run (an ad-hoc council, the retrospective, the night gardener) council deliberation holds by prose only. No native agent and no reviewer ever casts a seat's vote.[^council] A council that cannot run defers to a human.[^defer]\n\nThe council runner records the requested and the served model for each seat. Two gate-mode fault labels, `served-model` and `served-provider-mismatch`, mark a seat whose served model or served provider is absent or outside its registry entry's accepted set. The council runner never retries such a seat and never replaces its model.[^served]\n\n## When to Use\n- Choosing between architectural approaches\n- Evaluating tradeoffs (cost vs accuracy, speed vs correctness)\n- Any decision where multi-model consensus reduces risk\n- Before committing to a design pattern or library choice\n\n## When NOT to Use\n- **Factual lookups** (\"what's the default timeout?\") — a council of three models burns three API calls to confirm one fact. Use `call-llm` for a single answer.\n- **Obvious single answers** — if the choice is clear, a council just rubber-stamps it at 3x the cost and latency.\n- **Tight loops** — never put a council call inside a per-step or per-item loop; the parallel-but-multi-model cost compounds fast. Use `call-llm` for in-loop reasoning and reserve the council for the one decision that frames the loop.\n\nA good council question is a genuine *fork* — two or more defensible approaches where multiple perspectives change the outcome. If there's no fork, skip it.\n\n## File Locations\n```\ncrux\u002Fscripts\u002Fcrux\u002Fcouncil\u002F\n├── __init__.py\n├── async_council.py   ← USE THIS for an ad-hoc council\n└── council.py         ← sync fallback only\n```\n\n## Exports\n```\nFrom council __init__:\n  council_vote, quick_council, get_opinion, synthesize_opinions,\n  VotingMethod, Opinion, CouncilDecision, DEFAULT_COUNCIL, MODEL_WEIGHTS,\n  AsyncCouncil, AsyncVisualCouncil, AsyncCouncilConfig, VisualVoteResult,\n  create_async_council, create_visual_council\n```\n\n## Running it (PEP 723 driver under `uv`, NOT bare `python3`)\n\nThe council needs `httpx`, the HTTP client — **NOT** present in the bare\nsystem `python3`, so running there dies with `ModuleNotFoundError: No module named 'httpx'`. The\ncanonical pattern: write your driver to a temp `.py` file that\nstarts with the standard PEP 723 header, then run it with **`uv`**, which\nprovisions the dependencies into a cached, isolated environment:\n\n```python\n# \u002F\u002F\u002F script\n# requires-python = \">=3.10\"\n# dependencies = [\"httpx>=0.27\"]\n# \u002F\u002F\u002F\nimport sys\nsys.path.insert(0, \"${CRUX_PLUGIN_ROOT}\u002Fscripts\")\n\nfrom crux.council import create_async_council\n# ... your driver code (see Usage below)\n```\n\n```bash\nd=$(mktemp -d) && uv run \"$d\u002Fdriver.py\"\n```\n\nEach call to a seat has its own deadline: `AsyncCouncilConfig` and `cycle-council.py` use\n1200 s, while the `run-council.py` command line keeps 600 s through `--timeout`. A seat that\nfails with a `timeout`, `provider`, `malformed-response` or `truncated` fault is called once\nmore, so the worst case per seat is two deadlines, about 40 minutes for the driver. A\n`truncated` retry uses the seat's registry output limit. With `max_tokens` left unset, each\nseat's first-attempt budget is half its registry limit (at least 32000, at most the limit).\nA given `max_tokens` is the first-attempt budget for every seat. Each entry in a vote's\n`attempts` records the `max_tokens` it requested. The cycle-council record's `seats` carry\nthose entries; a `run-council.py` record does not carry the budget. A Claude Code foreground shell call\nstops at 10 minutes, so **run the driver as a background command** and read its output when it\nexits. A shell that kills the driver first loses every seat's answer. `AsyncCouncilConfig`\ntakes `max_retries` 0 (no retry) or 1 (the default); any other value is refused when the\nconfiguration is built.\n\nWrite the driver into a private per-run directory as above, never a fixed shared\npath like `\u002Ftmp\u002Fdriver.py` — a predictable name in a world-writable directory is a\nsymlink hazard.\n\nWhen writing the temp file, substitute `${CRUX_PLUGIN_ROOT}` (crux's portable\nplugin-root name — in Claude Code, the value of `CLAUDE_PLUGIN_ROOT`; in Codex,\nderived from this `SKILL.md`'s path per the Runtime compatibility note above)\nwith its actual value; in a source checkout, where the\nvariable is unset, substitute the checkout's `crux\u002F` directory. No\n`PYTHONPATH`, no extras, no surrounding project needed — the inline metadata\nmakes `uv` ignore any enclosing project.\n\n**Without `uv` installed this fails at the shell** — `command not found: uv`\n(exit 127); the script never executes. Remediation: install uv\n(https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002F). API keys are read from `~\u002F.crux\u002Fenv` via\n`crux_env`. One key backs every seat: `crux-env set OPENROUTER_API_KEY …`.\n\n## ALWAYS Use Async for an Ad-Hoc Council\n\n```python\nimport asyncio\n\nfrom crux.council import create_async_council, AsyncCouncilConfig\n\nasync def decide(question: str, context: str):\n    council = create_async_council()\n\n    decision = await council.deliberate(\n        prompt=f\"Question: {question}\\n\\nContext: {context}\",\n        system=\"You are a technical architecture council member.\",\n    )\n\n    # deliberate() returns a CouncilDeliberation (crux\u002Fscripts\u002Fcrux\u002Fcore\u002Fdata_classes.py):\n    #   votes, consensus, consensus_confidence, key_agreements, key_disagreements,\n    #   final_recommendation, dissent_count, confidence_adjustment (+ .has_critical_dissent).\n    # final_recommendation keys: action, reason, degraded, errored_seats, off_scale,\n    #   conditioned, conditions, nits, nit_items.\n    print(f\"Consensus: {decision.consensus}\")              # e.g. UNANIMOUS_APPROVE\n    # Consensus labels: UNANIMOUS_APPROVE, UNANIMOUS_REJECT, MAJORITY_APPROVE, MAJORITY_REJECT,\n    # SPLIT, NO_QUORUM, and UNANIMOUS_\u003CTOKEN> when every responding seat returned the same\n    # other decision. UNANIMOUS_OFF_SCALE replaces it when the token is not 1-40 characters of\n    # A-Z and underscore starting with a letter, contains APPROVE or REJECT, or is AUTO_EXECUTE\n    # or EXECUTE_WITH_MONITORING.\n    # Only UNANIMOUS_APPROVE and MAJORITY_APPROVE can route to an execute action; every other\n    # label routes to DEFER_TO_HUMAN or ABORT. Unanimity on REVISE is agreement, not approval.\n    print(decision.final_recommendation[\"off_scale\"])      # [\"\u003Cprovider>:\u003Cdecision>\", ...] for\n                                                            # responding seats outside the known\n                                                            # scale; each entry is bounded and\n                                                            # its unprintable characters replaced\n    print(f\"Confidence: {decision.consensus_confidence}\")\n\n    # A conditioned approval: a responding seat returned APPROVE_WITH_CONDITIONS. The label\n    # can still read UNANIMOUS_APPROVE or MAJORITY_APPROVE, so read the flag. The route is EXECUTE_WITH_MONITORING at\n    # most, never AUTO_EXECUTE. The flag fires only when the prompt puts APPROVE_WITH_CONDITIONS\n    # on the decision scale (the built-in envelope does not).\n    fr = decision.final_recommendation\n    if fr[\"conditioned\"]:                                   # decision.conditioned is the same flag\n        for seat, points in fr[\"conditions\"].items():       # keyed by seat; untrusted model text, carried unchanged\n            print(f\"  [condition] {seat}: {points}\")\n    # Nits: a responding seat returned APPROVE_WITH_NITS. Nits are reported and never change\n    # the route. Only conditions block.\n    if fr[\"nits\"]:                                          # decision.nits is the same flag\n        for seat, points in fr[\"nit_items\"].items():\n            print(f\"  [nit] {seat}: {points}\")\n    # Both flags read False, with empty items, when no responding seat returned them. An\n    # errored seat never contributes.\n\n    if decision.dissent_count:\n        print(f\"{decision.dissent_count} dissent(s) — srde output is evidence only\")\n        for v in decision.votes:                            # each vote is a CouncilVote\n            for d in v.dissenting_points:                   # NOTE: dissenting_points, not dissent_points\n                print(f\"  [{v.provider}\u002F{v.model}] {d}\")\n        # See srde skill\n\n    # Each vote also records how its seat fared. finish_reason is the reply's finish reason\n    # from a closed set (stop, length, content_filter, tool_calls, error, other, missing; None\n    # when the seat got no reply, such as a timeout). fault_label names an errored seat's cause\n    # (one of timeout, provider, malformed-response, truncated, auth, rate-limit, client-config,\n    # refused, insufficient-credit, unexpected-status or unreachable, or unknown). A gate council\n    # adds served-model and served-provider-mismatch, which are never retried. retried marks\n    # a seat called a second time; recovered marks a seat that answered on that second call;\n    # first_fault_label and first_finish_reason describe its first call.\n    # attempts lists each call to the seat with what its reply reported.\n    for v in decision.votes:\n        print(v.provider, v.finish_reason, v.fault_label, v.retried, v.recovered)\n\n    return decision\n\nresult = asyncio.run(decide(\n    question=\"Should we use approach A or B?\",\n    context=\"Full context including constraints and requirements\",\n))\n```\n\n### Async with Custom Config\n\n```python\nconfig = AsyncCouncilConfig(...)  # customize models, weights, etc.\ncouncil = create_async_council(config=config)\ndecision = await council.deliberate(prompt=\"...\", system=\"...\")\n```\n\n### Sync Deliberation (from async council)\n\n```python\ncouncil = create_async_council()\ndecision = council.deliberate_sync(prompt=\"...\", system=\"...\")\n```\n\n### Visual Council (image analysis)\n\n```python\nfrom crux.council import create_visual_council\n\nvisual_council = create_visual_council()\nvotes = await visual_council.analyze_image(\n    image_b64=\"\u003Cbase64-encoded-image>\",\n    prompt=\"Does this dashboard render correctly?\",\n)\n# Or sync:\nvotes = visual_council.analyze_sync(image_b64=\"...\", prompt=\"...\")\n```\n\nThe visual council retries and labels a seat as the text council does. Its result for a seat\nthat errored reads `passed=False` at confidence 0.0 and carries `errored=True`. Skip errored\nresults when you count failed checks: an errored seat is a missing check, not a failed one.\n\n## Sync Fallback (only if async is impossible)\n\n```python\nfrom crux.council import council_vote, quick_council, get_opinion, VotingMethod\nfrom crux.core.llm_caller import get_default_model\n\n# Full council with options\ndecision = council_vote(\n    question=\"Should we use approach A or B?\",\n    context=\"Full context here\",\n    voting_method=VotingMethod.ARBITER,   # models and arbiter default to the registry's assignment\n    tracer=None,  # optional Tracer instance\n)\n\n# Quick one-liner (returns string)\nanswer = quick_council(\"Should we cache at the API or DB layer?\")\n\n# Get a single model's opinion\nopinion = get_opinion(get_default_model(\"council_arbiter\"), \"question\", context=\"...\")\n```\n\n## Model Config\n\nThe router registry assigns each seat by role: `openai_top`, `anthropic_top` and `google_top`. The sync council reads `council_default` and `council_arbiter`. Read the assignments in `${CRUX_PLUGIN_ROOT}\u002Fscripts\u002Fcrux\u002F_config\u002Fllm_router_config.json`; this skill names no model.[^seats]\n\nThe router reaches every provider through the OpenRouter gateway.\n\nAPI keys are read via `crux_env.require(...)` from `~\u002F.crux\u002Fenv` (managed by `crux-env`).\n\n## Rules\n- ALWAYS prefer async for an ad-hoc council and for agent teams (parallel teammates = concurrent calls)\n- ALWAYS check `dissent_count` (and each vote's `dissenting_points`) — never just read `consensus`\n- ALWAYS read `final_recommendation[\"conditioned\"]` and `[\"nits\"]`: a conditioned approval can read `UNANIMOUS_APPROVE` or `MAJORITY_APPROVE` but never routes to `AUTO_EXECUTE`\n- ALWAYS run an ad-hoc council driver in the background (inside a cycle run, the council gate uses `run-council.py`): the worst case per seat is about 40 minutes\n- ALWAYS provide rich context — models deliberate better with specifics\n- Every seat prompt states the operating assumptions: a trusted owner, cooperating agents, a trusted local checkout and one release coordinator. Plausible mistakes and crashes are in scope; deliberate local tampering and identity authentication are outside. Crux runs locally with a trusted caller, environment and interpreter, so denial of service from hostile input is never a blocker,[^trust] while an accidental malformed or oversized input that hangs or crashes Crux is a plausible mistake. A blocker states a plausible failure within those assumptions, its consequence and the smallest necessary fix. Everything else is an advisory or a documented limit, and \"safety-related\" alone does not justify a block. Every builder appends the shared text from `council_assumptions.py`. Cycle councils' vote instruction applies the same blocker test; historical promptbook gate councils keep their original instruction. Approval thresholds are unchanged[^blocker]\n- A recommendation that adds complexity (code, tests, instructions, execution, configuration or concepts) names the concrete failure it prevents, why a simpler fix is insufficient, and its maintenance and testing cost; otherwise it is an advisory\n- On a simplify cycle, every configured mode evaluates only the proposals within the agreed scope: no new features, speculative hardening or unrelated requirements. An addition the council believes necessary is reported as different work, not made a condition of approval. The council may still block or drop a proposal whose retained behavior or verification is inadequate; only remedies that add complexity are out of scope. Voting modes, thresholds, seats and attempt limits are unchanged\n- Treat srde output as evidence: it never settles a gate and never runs on the ADR path (see srde skill)[^srde]\n- Register results on Semantic Bridge for cross-teammate visibility\n\n[^council]: rule:council-is-never-harness-native, rule:council-gate-needs-a-runner-record\n[^review]: rule:review-gate-needs-a-reviewer-report\n[^seats]: rule:every-seat-assesses-five-dimensions, rule:council-assignment-is-validated-before-spend\n[^served]: rule:seat-records-requested-and-served-model, rule:served-provider-matches-a-declared-set\n[^defer]: rule:only-a-preflight-refusal-is-retried\n[^attempt]: rule:an-open-council-attempt-stops-the-gate\n[^match]: rule:council-evidence-matches-its-attempt\n[^recover]: rule:recovery-never-deliberates-again\n[^srde]: rule:de-wire-srde-from-adr-path, rule:blocking-finding-names-a-plausible-failure\n[^blocker]: rule:blocking-finding-names-a-plausible-failure\n[^trust]: rule:crux-trusts-its-local-environment\n","# Crux Council Vote (Async-First)\n\n> **Execution context:** this skill runs in a forked subagent and returns a summary to the caller. The fork does not see the main-thread conversation, so pass any needed context explicitly at invocation.\n\n\u003C!-- BEGIN GENERATED: runtime-compat -->\n## Runtime compatibility\n\nThis skill is portable across Claude Code, Codex, OpenCode, and OMP (Oh My Pi). This section overrides platform-specific labels below.\n\n- Before running a command that uses `CRUX_PLUGIN_ROOT`, set it to the installed plugin root. In Claude Code, use the value of `CLAUDE_PLUGIN_ROOT`. In Codex, OpenCode, and OMP, derive it from the absolute path of this selected `SKILL.md`: the plugin root is the parent of its `skills\u002F` directory (in OMP `bash`, `skill:\u002F\u002F\u003Cname>` resolves to the skill directory). In a source checkout, use the checkout `crux\u002F` directory.\n- For project-local skills, use `.claude\u002Fskills` in Claude Code, `.agents\u002Fskills` in Codex, `.opencode\u002Fskills` in OpenCode (which also reads the singular `.opencode\u002Fskill`), and `.omp\u002Fskills` in OMP. Set `CRUX_LOCAL_SKILLS_DIR` to that path before following any command below that uses it.\n- Translate Claude Code tool labels such as `Agent`, `Read`, `Write`, `Bash`, `WebSearch`, and `WebFetch` to the matching capability in the current session. Codex names its own capabilities; OpenCode uses the lowercase forms `subagent`, `read`, `edit`, `shell`, `websearch`, and `webfetch`, where `edit` covers both `Edit` and `Write`; OMP uses `task`, `read`, `edit`, `write`, `bash`, `grep`, `glob`, `find`, `todo`, and `web_search` (where `read` also handles URLs and `skill:\u002F\u002F` URIs). Do not attempt to invoke the Claude Code labels as literal commands on another host.\n- Install the generated role agents before delegating: `install-codex-agents` in Codex, `install-opencode-agents` in OpenCode, `install-omp-agents` in OMP. Codex names them `crux_architect`, `crux_brainstormer`, `crux_commander`, `crux_dev_lead`, `crux_developer`, `crux_historian`, `crux_librarian`, `crux_night_gardener`, `crux_reviewer`, and `crux_wayfinder`; OpenCode and OMP use the bare role names `architect`, `brainstormer`, `commander`, `dev-lead`, `developer`, `historian`, `librarian`, `night-gardener`, `reviewer`, and `wayfinder` (in OMP, `scout` may also be used for reconnaissance alongside `wayfinder`). If a required role or capability is unavailable, report that truthfully instead of claiming it ran.\n- Argument placeholders such as `$adr` and `$book` bind only in Claude Code. On a host without argument binding they are unset — take the value from the user's phrase. The \"Fields OpenCode ignores\" section of `OPENCODE_GUIDE.md` names the invocation-control fields OpenCode ignores.\n\u003C!-- END GENERATED: runtime-compat -->\n\n## Configured councils\n\nRead `.bionic.yml` through the canonical policy reader. Council mode and human\nacceptance are independent. Existing projects without a council setting retain\nStrict. New setup explicitly selects Standard or an owner-selected Lite mode.\n\n- Lite: native configured Apex challenge, architect response, fresh final review.\n  Final reviewer approval and exact-final architect agreement are both required.\n  No OpenRouter account is required. Follow [native-lite.md](references\u002Fnative-lite.md).\n- Standard: three gateway provider seats, two approvals and architect agreement.\n  Address every dissent with a typed disposition. Resolved\u002Frefuted findings need\n  retained checkable evidence; a non-binding trade-off needs an explanation.\n  Single-seat dissent rule: when two seats approve and one dissents, revise\n  only if the finding is a binding failure or missing authorization, or states\n  a plausible failure, consequence and smallest fix that the architect cannot\n  refute and no later phase would catch. For an unresolved required safety\n  fact, obtain a targeted check; if it stays unresolved, withdraw the dependent\n  choice. A required safety fact about data policy or publishing is not a council\n  dispute: it is the outward-facing stop, and the one question put to the owner.\n  Otherwise answer on the same revision with reasoned agreement. Record `refuted` with a named check\n  when the finding fails the blocking standard. Record `accepted-tradeoff` with\n  the reason when a later phase would catch it or its risk is lower than the\n  cost of another round. When all three seats approve with nits only, record\n  agreement; do not revise for nits.\n- Strict: all three gateway seats approve with no recorded blocking findings.\n\nStandard and Strict require `OPENROUTER_API_KEY` through `crux-env`. Missing\ncredentials or native capabilities refuse with their cause. Never change mode,\nprovider, model or seat count as a fallback. Demonstrated binding failures,\nmissing authorization and unresolved required safety facts prevent approval.\nA minority label alone is not a Standard veto.\n\nUse `cycle-council.py` for ADR, IDR, ADR-creation, consultation and disposition\npurposes. Consultation approves no decision. Its `--identity` stays stable across\nrenamed files. IDRs retain their cycle name across revisions. The configured\nattempt limit defaults to seven for new mode-aware decisions. Stop on stalled\nprogress. A council dispute never goes to the owner. At the limit, or when the\nsame blocking finding recurs on two consecutive attempts after a revision that\nanswered it, the architect withdraws the decision with `withdraw-decision`;\nno agent requests an extension. A Standard-mode single-seat dissent that the\narchitect disposes as `accepted-tradeoff` on an approved attempt is answered,\nnot unanswered; it does not count toward the recurrence trigger, and neither\ndoes its repetition on a later attempt over the same subject bytes. A finding\ncounts toward recurrence only when it blocks approval of an attempt. A narrowed replacement names the record it\nreplaces, and its attempts continue that record's count. An extension or mode\nchange the owner records on their own initiative keeps its provenance, findings\nand total count. Historical reviews keep their original policy and limit.\n\nGateway convening is detached by default. Initial success names a running attempt\nand `await_command`; it is not a vote or approval. Run the returned await command\nfor the terminal outcome, then the existing architect disposition. `--foreground`\nis the synchronous diagnostic form. `--probe ATTEMPT` reads the attempt flock.\n`--recover ATTEMPT` sends no request: it refuses while the flock is held, recovers\nthe original persisted result, or marks an unowned execution interrupted. An\ninterrupted claim still counts, and a later convening is a separate counted attempt.\nFull cycle inspect exposes exact subject\u002Fattempt\u002Fresult references, recovery argv\nand prerequisites without requests or writes. New gateway attempts always persist\nrequest\u002Foutcome checkpoints under existing ownership. A persistence failure stops\nfurther calls and positive approval. Durable negatives survive; missing outcomes\nstay unknown. Recovery sends no request and resets no budget. Partial events never\napprove; later convening uses prior findings as historical context beside fresh votes.\nNew approval\u002Fapplication checks current bound evidence. Pure sealed historical\nvalidation and applied acceptance retain their standing. The cycle commands reference\ngives the interfaces; historical\u002Fnative-Lite callers keep their original protocols.\nThe runner holds ownership through result persistence; never hand-edit or delete\na claim. A missing launcher acknowledgement proves nothing: after the exact child\nterminates, check its exact preallocated claim identity before reporting no claim.\nUse recovery for an existing claim; do not request a substitute owner approval.\nNative Lite follows its separate sealed-dispatch procedure.\n\nSettle a dispute under the configured mode. Record the architect agreement even\nwhen every seat approves: it names the approved sha256 and each nit id with a\ndisposition, `apply-in-build` (a note on a Build task) or `declined` with its\nreason, in the `nits` list of the `--architect` JSON. A nit never starts another\nattempt, and the approved revision stays as approved. `srde` output may be kept as\nevidence for a `refuted` disposition of an IDR dissent; `srde` never runs on the\nADR path.\n\nA new revision of an IDR whose earlier revision was approved is reviewed as a\ndelta. The writer computes each top-level contents key added, removed or changed\nagainst the latest approved revision, stores it in the attempt, and checks it\nagain against both revisions' bytes at convening and at approval. The seats\nreceive the full new revision with the changed keys marked and judge the change.\nApproval binds the new revision's sha256, and a delta attempt counts toward the\nsame limit.\n\nA Standard result waits for the architect even when its provider votes suffice.\nThat result reports the outcome `awaiting-architect`, exits 1 and prints the command\nto run; the sealed record keeps `changes-required`. Use `--architect PATH --record\n\u003Cresult>`, where PATH names a JSON file, to record exact-revision agreement and\ntyped dissent; do not call another council just to add that agreement. The writer\nresolves evidence artifacts safely and hashes their bytes. A targeted check for a\nrequired fact is not another mandatory council round.\n\nFix-round discipline, for every mode: fix all findings (or refute one with a named check) in one revision, apply each\nfix at every site that states the same concept, and rerun the pre-council\nself-check's mechanical checks before reconvening.\n\nThe following promptbook procedures preserve historical gateway-only behavior.\nThey do not override configured prospective modes.\n\n## Gate council or ad-hoc council\n\nCouncil deliberation runs only through the council runner, and its council record is the only evidence a council gate accepts.[^council] Independent review is a reviewer's examination, and its reviewer report is the only evidence an independent-review gate accepts.[^review] Neither satisfies the other's gate.\n\nInside a promptbook run, a council gate uses the council runner, `run-council.py`. Run it as a background command:\n\n```bash\nuv run \"${CRUX_PLUGIN_ROOT}\u002Fscripts\u002Frun-council.py\" \u003Crun-RUN-NNN.yaml> [--prompt N] --round N \\\n  --question \u003Cquestion file> --subject \u003Csubject file> [--subject \u003Csubject file> ...] \\\n  [--book \u003Cbook file>] [--retain-subjects] [--max-tokens 64000] [--timeout 600]\n```\n\n`--max-tokens` is the first-attempt output budget for every seat (default 64000). A seat\nwhose reply stops on that budget is retried once at its own output limit from the model\nregistry. A value above the smallest selected seat limit is refused before any call.\n`--timeout` is the per-seat-call deadline in seconds (default 600). When a reasoning seat\ntruncates, pass `--timeout 1200`: the escalated retry under 600 s is likely to time out.\n\nFinish any open Git merge or other sequence before council. Both book formats\nuse the current attempt-aware gate. An implementation module uses council kind\n`implementation`, including in a combined architectural book; `combined` is never\na council kind. Verify and patch retain their existing kinds and dimensions.\nSelect an exact formal revision with `--implementation-revision \u003Cdecision-path>`.\nAlso pass that exact path as `--subject` with `--retain-subjects`. Record each subject's\nwrite-time witness before committing it; never attribute dirty bytes after a refusal.\n`--migration-batch` is retired: it exits 2 with `historical-migration-writes-retired`\nbefore any batch is loaded, any attempt is claimed or any council call is made. Propose\nclause migrations in the prospective format through `migrate-adrs`.\nNew formal closes retain context three\u002Fprofile four. Historical context two\u002Fprofile three\nand context one\u002Fprofile two remain immutable and replay-only, never selected from the book format.\n\n`--prompt` is optional and defaults to the run's current prompt. The run, `--round`, `--question` and at least one `--subject` are required.\n\n`--round` is the number of council records with outcome `ran` already in this module (in a patch book, this prompt), plus one. In an adr module a valid adjudicator refutation record fills place 3. A could-not-run record takes no place, so after you fix its cause you reconvene with the same number. `run-council.py` refuses a mismatch before any call: exit 1, a JSON error on stdout, `\"record\": null`, and no record written.\n\nThe council runner commits the attempt record and the council record itself. Before any council request it commits an attempt record that claims the round. The claim begins a council attempt, one execution of the round, which is distinct from a seat's calls below. A round that began never vanishes: until its own council record resolves it, the gate stops on the open attempt.[^attempt] It verifies that the committed council record holds the bytes it computed.[^match] The gate check (`advance-run.py`) reads only a committed record. Do not commit, edit or delete either record yourself.\n\nThe council runner exits with one of three codes. A council runner that ends without an exit code, with a signal or with a code other than 0, 1 and 2 stopped before it reported. Treat it as an exit 2 that names a claimed attempt, and follow the steps under `2` with `--prompt \u003Cn>`; never convene another round until recovery reports.[^recover]\n\n- `0`: it wrote, committed and verified a council record that ran.\n- `1`: it committed a council record the gate check stops on, or it refused the call and wrote no council record. The record is a could-not-run record, whose action is DEFER_TO_HUMAN, or a council record that ran but whose full write the secret scan refused: advance `--outcome blocked` with it attached. A refusal prints `\"record\": null` and leaves nothing to attach. `\"refused\": \"round\"` means correct `--round` and run again. `\"refused\": \"prompt\"` means `--prompt` is not the run's current prompt: run again with the current prompt or omit `--prompt`. `\"refused\": \"preflight\"` means a question or subject failed the input checks before any request: apply each cause's `repair` and run again at the same round. `retype` means correct the path. `commit-run-work` means commit the subject, the run's own witnessed work, with `run-work-witness.py commit \u003Crun> --path \u003Cpath>`. The third preflight refusal at one convening prompt commits a `preflight-retries-spent` record instead, an escalation-loop stop.[^defer] `\"refused\": \"attempt-open\"` means an open attempt holds the module: recover it, never claim past it.\n- `2`: no council record was committed. Exit 2 never means nothing was written. When stderr names `timeout`, or names outside work the commit moved, the owner's remedy comes before recovery: report a contradicted-premise stop for the owner. The owner first restores the set-aside work (`git stash list`, or a pre-commit framework's backup patch) and then removes a stale `index.lock` in the git directory. Start recovery only after the owner reports both steps done. When stderr names a claimed attempt, run the process check, probe the lock with `run-council.py --recover \u003Crun> --prompt \u003Cn> --probe`, and once no live council runner holds it, run `run-council.py --recover \u003Crun> --prompt \u003Cn>`. Pass `--prompt`: without it recovery cannot recognise a record whose pending copy is already removed, and it reports `nothing-open`. Recovery makes no council request and commits or recognises only the original council record.[^recover] Route each recovery result by the recovery table in `run-promptbook`'s `references\u002Fgates.md`. Never convene another round over a claimed attempt. When stderr names no attempt, the prompt cannot advance, and the conductor reports why, as a contradicted-premise stop for the owner.\n\nA step that writes a file it later passes as a subject records a run-work witness after each write: `run-work-witness.py record \u003Crun> --prompt N --path \u003Cpath>`. Never record a witness to repair a preflight refusal: a subject with no witness taken at its write goes to the owner. `run-promptbook`'s `references\u002Fgates.md` describes how a conductor issues and advances a gate prompt, repairs a preflight refusal and recovers an open attempt.\n\nThe async driver below serves an ad-hoc council only, such as a design question outside a run. It satisfies no promptbook council gate. Outside a run (an ad-hoc council, the retrospective, the night gardener) council deliberation holds by prose only. No native agent and no reviewer ever casts a seat's vote.[^council] A council that cannot run defers to a human.[^defer]\n\nThe council runner records the requested and the served model for each seat. Two gate-mode fault labels, `served-model` and `served-provider-mismatch`, mark a seat whose served model or served provider is absent or outside its registry entry's accepted set. The council runner never retries such a seat and never replaces its model.[^served]\n\n## When to Use\n- Choosing between architectural approaches\n- Evaluating tradeoffs (cost vs accuracy, speed vs correctness)\n- Any decision where multi-model consensus reduces risk\n- Before committing to a design pattern or library choice\n\n## When NOT to Use\n- **Factual lookups** (\"what's the default timeout?\") — a council of three models burns three API calls to confirm one fact. Use `call-llm` for a single answer.\n- **Obvious single answers** — if the choice is clear, a council just rubber-stamps it at 3x the cost and latency.\n- **Tight loops** — never put a council call inside a per-step or per-item loop; the parallel-but-multi-model cost compounds fast. Use `call-llm` for in-loop reasoning and reserve the council for the one decision that frames the loop.\n\nA good council question is a genuine *fork* — two or more defensible approaches where multiple perspectives change the outcome. If there's no fork, skip it.\n\n## File Locations\n```\ncrux\u002Fscripts\u002Fcrux\u002Fcouncil\u002F\n├── __init__.py\n├── async_council.py   ← USE THIS for an ad-hoc council\n└── council.py         ← sync fallback only\n```\n\n## Exports\n```\nFrom council __init__:\n  council_vote, quick_council, get_opinion, synthesize_opinions,\n  VotingMethod, Opinion, CouncilDecision, DEFAULT_COUNCIL, MODEL_WEIGHTS,\n  AsyncCouncil, AsyncVisualCouncil, AsyncCouncilConfig, VisualVoteResult,\n  create_async_council, create_visual_council\n```\n\n## Running it (PEP 723 driver under `uv`, NOT bare `python3`)\n\nThe council needs `httpx`, the HTTP client — **NOT** present in the bare\nsystem `python3`, so running there dies with `ModuleNotFoundError: No module named 'httpx'`. The\ncanonical pattern: write your driver to a temp `.py` file that\nstarts with the standard PEP 723 header, then run it with **`uv`**, which\nprovisions the dependencies into a cached, isolated environment:\n\n```python\n# \u002F\u002F\u002F script\n# requires-python = \">=3.10\"\n# dependencies = [\"httpx>=0.27\"]\n# \u002F\u002F\u002F\nimport sys\nsys.path.insert(0, \"${CRUX_PLUGIN_ROOT}\u002Fscripts\")\n\nfrom crux.council import create_async_council\n# ... your driver code (see Usage below)\n```\n\n```bash\nd=$(mktemp -d) && uv run \"$d\u002Fdriver.py\"\n```\n\nEach call to a seat has its own deadline: `AsyncCouncilConfig` and `cycle-council.py` use\n1200 s, while the `run-council.py` command line keeps 600 s through `--timeout`. A seat that\nfails with a `timeout`, `provider`, `malformed-response` or `truncated` fault is called once\nmore, so the worst case per seat is two deadlines, about 40 minutes for the driver. A\n`truncated` retry uses the seat's registry output limit. With `max_tokens` left unset, each\nseat's first-attempt budget is half its registry limit (at least 32000, at most the limit).\nA given `max_tokens` is the first-attempt budget for every seat. Each entry in a vote's\n`attempts` records the `max_tokens` it requested. The cycle-council record's `seats` carry\nthose entries; a `run-council.py` record does not carry the budget. A Claude Code foreground shell call\nstops at 10 minutes, so **run the driver as a background command** and read its output when it\nexits. A shell that kills the driver first loses every seat's answer. `AsyncCouncilConfig`\ntakes `max_retries` 0 (no retry) or 1 (the default); any other value is refused when the\nconfiguration is built.\n\nWrite the driver into a private per-run directory as above, never a fixed shared\npath like `\u002Ftmp\u002Fdriver.py` — a predictable name in a world-writable directory is a\nsymlink hazard.\n\nWhen writing the temp file, substitute `${CRUX_PLUGIN_ROOT}` (crux's portable\nplugin-root name — in Claude Code, the value of `CLAUDE_PLUGIN_ROOT`; in Codex,\nderived from this `SKILL.md`'s path per the Runtime compatibility note above)\nwith its actual value; in a source checkout, where the\nvariable is unset, substitute the checkout's `crux\u002F` directory. No\n`PYTHONPATH`, no extras, no surrounding project needed — the inline metadata\nmakes `uv` ignore any enclosing project.\n\n**Without `uv` installed this fails at the shell** — `command not found: uv`\n(exit 127); the script never executes. Remediation: install uv\n(https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002F). API keys are read from `~\u002F.crux\u002Fenv` via\n`crux_env`. One key backs every seat: `crux-env set OPENROUTER_API_KEY …`.\n\n## ALWAYS Use Async for an Ad-Hoc Council\n\n```python\nimport asyncio\n\nfrom crux.council import create_async_council, AsyncCouncilConfig\n\nasync def decide(question: str, context: str):\n    council = create_async_council()\n\n    decision = await council.deliberate(\n        prompt=f\"Question: {question}\\n\\nContext: {context}\",\n        system=\"You are a technical architecture council member.\",\n    )\n\n    # deliberate() returns a CouncilDeliberation (crux\u002Fscripts\u002Fcrux\u002Fcore\u002Fdata_classes.py):\n    #   votes, consensus, consensus_confidence, key_agreements, key_disagreements,\n    #   final_recommendation, dissent_count, confidence_adjustment (+ .has_critical_dissent).\n    # final_recommendation keys: action, reason, degraded, errored_seats, off_scale,\n    #   conditioned, conditions, nits, nit_items.\n    print(f\"Consensus: {decision.consensus}\")              # e.g. UNANIMOUS_APPROVE\n    # Consensus labels: UNANIMOUS_APPROVE, UNANIMOUS_REJECT, MAJORITY_APPROVE, MAJORITY_REJECT,\n    # SPLIT, NO_QUORUM, and UNANIMOUS_\u003CTOKEN> when every responding seat returned the same\n    # other decision. UNANIMOUS_OFF_SCALE replaces it when the token is not 1-40 characters of\n    # A-Z and underscore starting with a letter, contains APPROVE or REJECT, or is AUTO_EXECUTE\n    # or EXECUTE_WITH_MONITORING.\n    # Only UNANIMOUS_APPROVE and MAJORITY_APPROVE can route to an execute action; every other\n    # label routes to DEFER_TO_HUMAN or ABORT. Unanimity on REVISE is agreement, not approval.\n    print(decision.final_recommendation[\"off_scale\"])      # [\"\u003Cprovider>:\u003Cdecision>\", ...] for\n                                                            # responding seats outside the known\n                                                            # scale; each entry is bounded and\n                                                            # its unprintable characters replaced\n    print(f\"Confidence: {decision.consensus_confidence}\")\n\n    # A conditioned approval: a responding seat returned APPROVE_WITH_CONDITIONS. The label\n    # can still read UNANIMOUS_APPROVE or MAJORITY_APPROVE, so read the flag. The route is EXECUTE_WITH_MONITORING at\n    # most, never AUTO_EXECUTE. The flag fires only when the prompt puts APPROVE_WITH_CONDITIONS\n    # on the decision scale (the built-in envelope does not).\n    fr = decision.final_recommendation\n    if fr[\"conditioned\"]:                                   # decision.conditioned is the same flag\n        for seat, points in fr[\"conditions\"].items():       # keyed by seat; untrusted model text, carried unchanged\n            print(f\"  [condition] {seat}: {points}\")\n    # Nits: a responding seat returned APPROVE_WITH_NITS. Nits are reported and never change\n    # the route. Only conditions block.\n    if fr[\"nits\"]:                                          # decision.nits is the same flag\n        for seat, points in fr[\"nit_items\"].items():\n            print(f\"  [nit] {seat}: {points}\")\n    # Both flags read False, with empty items, when no responding seat returned them. An\n    # errored seat never contributes.\n\n    if decision.dissent_count:\n        print(f\"{decision.dissent_count} dissent(s) — srde output is evidence only\")\n        for v in decision.votes:                            # each vote is a CouncilVote\n            for d in v.dissenting_points:                   # NOTE: dissenting_points, not dissent_points\n                print(f\"  [{v.provider}\u002F{v.model}] {d}\")\n        # See srde skill\n\n    # Each vote also records how its seat fared. finish_reason is the reply's finish reason\n    # from a closed set (stop, length, content_filter, tool_calls, error, other, missing; None\n    # when the seat got no reply, such as a timeout). fault_label names an errored seat's cause\n    # (one of timeout, provider, malformed-response, truncated, auth, rate-limit, client-config,\n    # refused, insufficient-credit, unexpected-status or unreachable, or unknown). A gate council\n    # adds served-model and served-provider-mismatch, which are never retried. retried marks\n    # a seat called a second time; recovered marks a seat that answered on that second call;\n    # first_fault_label and first_finish_reason describe its first call.\n    # attempts lists each call to the seat with what its reply reported.\n    for v in decision.votes:\n        print(v.provider, v.finish_reason, v.fault_label, v.retried, v.recovered)\n\n    return decision\n\nresult = asyncio.run(decide(\n    question=\"Should we use approach A or B?\",\n    context=\"Full context including constraints and requirements\",\n))\n```\n\n### Async with Custom Config\n\n```python\nconfig = AsyncCouncilConfig(...)  # customize models, weights, etc.\ncouncil = create_async_council(config=config)\ndecision = await council.deliberate(prompt=\"...\", system=\"...\")\n```\n\n### Sync Deliberation (from async council)\n\n```python\ncouncil = create_async_council()\ndecision = council.deliberate_sync(prompt=\"...\", system=\"...\")\n```\n\n### Visual Council (image analysis)\n\n```python\nfrom crux.council import create_visual_council\n\nvisual_council = create_visual_council()\nvotes = await visual_council.analyze_image(\n    image_b64=\"\u003Cbase64-encoded-image>\",\n    prompt=\"Does this dashboard render correctly?\",\n)\n# Or sync:\nvotes = visual_council.analyze_sync(image_b64=\"...\", prompt=\"...\")\n```\n\nThe visual council retries and labels a seat as the text council does. Its result for a seat\nthat errored reads `passed=False` at confidence 0.0 and carries `errored=True`. Skip errored\nresults when you count failed checks: an errored seat is a missing check, not a failed one.\n\n## Sync Fallback (only if async is impossible)\n\n```python\nfrom crux.council import council_vote, quick_council, get_opinion, VotingMethod\nfrom crux.core.llm_caller import get_default_model\n\n# Full council with options\ndecision = council_vote(\n    question=\"Should we use approach A or B?\",\n    context=\"Full context here\",\n    voting_method=VotingMethod.ARBITER,   # models and arbiter default to the registry's assignment\n    tracer=None,  # optional Tracer instance\n)\n\n# Quick one-liner (returns string)\nanswer = quick_council(\"Should we cache at the API or DB layer?\")\n\n# Get a single model's opinion\nopinion = get_opinion(get_default_model(\"council_arbiter\"), \"question\", context=\"...\")\n```\n\n## Model Config\n\nThe router registry assigns each seat by role: `openai_top`, `anthropic_top` and `google_top`. The sync council reads `council_default` and `council_arbiter`. Read the assignments in `${CRUX_PLUGIN_ROOT}\u002Fscripts\u002Fcrux\u002F_config\u002Fllm_router_config.json`; this skill names no model.[^seats]\n\nThe router reaches every provider through the OpenRouter gateway.\n\nAPI keys are read via `crux_env.require(...)` from `~\u002F.crux\u002Fenv` (managed by `crux-env`).\n\n## Rules\n- ALWAYS prefer async for an ad-hoc council and for agent teams (parallel teammates = concurrent calls)\n- ALWAYS check `dissent_count` (and each vote's `dissenting_points`) — never just read `consensus`\n- ALWAYS read `final_recommendation[\"conditioned\"]` and `[\"nits\"]`: a conditioned approval can read `UNANIMOUS_APPROVE` or `MAJORITY_APPROVE` but never routes to `AUTO_EXECUTE`\n- ALWAYS run an ad-hoc council driver in the background (inside a cycle run, the council gate uses `run-council.py`): the worst case per seat is about 40 minutes\n- ALWAYS provide rich context — models deliberate better with specifics\n- Every seat prompt states the operating assumptions: a trusted owner, cooperating agents, a trusted local checkout and one release coordinator. Plausible mistakes and crashes are in scope; deliberate local tampering and identity authentication are outside. Crux runs locally with a trusted caller, environment and interpreter, so denial of service from hostile input is never a blocker,[^trust] while an accidental malformed or oversized input that hangs or crashes Crux is a plausible mistake. A blocker states a plausible failure within those assumptions, its consequence and the smallest necessary fix. Everything else is an advisory or a documented limit, and \"safety-related\" alone does not justify a block. Every builder appends the shared text from `council_assumptions.py`. Cycle councils' vote instruction applies the same blocker test; historical promptbook gate councils keep their original instruction. Approval thresholds are unchanged[^blocker]\n- A recommendation that adds complexity (code, tests, instructions, execution, configuration or concepts) names the concrete failure it prevents, why a simpler fix is insufficient, and its maintenance and testing cost; otherwise it is an advisory\n- On a simplify cycle, every configured mode evaluates only the proposals within the agreed scope: no new features, speculative hardening or unrelated requirements. An addition the council believes necessary is reported as different work, not made a condition of approval. The council may still block or drop a proposal whose retained behavior or verification is inadequate; only remedies that add complexity are out of scope. Voting modes, thresholds, seats and attempt limits are unchanged\n- Treat srde output as evidence: it never settles a gate and never runs on the ADR path (see srde skill)[^srde]\n- Register results on Semantic Bridge for cross-teammate visibility\n\n[^council]: rule:council-is-never-harness-native, rule:council-gate-needs-a-runner-record\n[^review]: rule:review-gate-needs-a-reviewer-report\n[^seats]: rule:every-seat-assesses-five-dimensions, rule:council-assignment-is-validated-before-spend\n[^served]: rule:seat-records-requested-and-served-model, rule:served-provider-matches-a-declared-set\n[^defer]: rule:only-a-preflight-refusal-is-retried\n[^attempt]: rule:an-open-council-attempt-stops-the-gate\n[^match]: rule:council-evidence-matches-its-attempt\n[^recover]: rule:recovery-never-deliberates-again\n[^srde]: rule:de-wire-srde-from-adr-path, rule:blocking-finding-names-a-plausible-failure\n[^blocker]: rule:blocking-finding-names-a-plausible-failure\n[^trust]: rule:crux-trusts-its-local-environment\n",{"name":5,"model":24,"context":25,"metadata":26,"description":7},"opus","fork",{"tags":27,"bundles":28,"triggers":29,"risk_level":30,"requires_access":31},"verification, multi-model, council, anti-hallucination","crux-verification, crux-docs","run the council | convene the council | ask the council | get a multi-model opinion","medium","write-run","main","dd4f16e7ac12ff940790660ec897677284ae4b6d","crux\u002Fskills\u002Fcouncil","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux","https:\u002F\u002Fbionic-coding.com\u002Fcrux\u002F","https:\u002F\u002Fgithub.com\u002Fbionic-coding","https:\u002F\u002Fbionic-coding.com","Durable intent, disposable code.","2026-11-12 09:36:33",{"source":8,"skillName":5,"commands":42,"prompt":52,"zipAvailable":53},[43,46,49],{"tool":44,"command":45},"npx","npx skills add bionic-coding\u002Fcrux --skill council",{"tool":47,"command":48},"bunx","bunx skills add bionic-coding\u002Fcrux --skill council",{"tool":50,"command":51},"pnpm","pnpm dlx skills add bionic-coding\u002Fcrux --skill council","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Fcouncil\u002FSKILL.md",true,[55],{"path":56,"resourceType":57,"resourceTypeDesc":58,"sizeBytes":59,"sourceHtmlUrl":60},"references\u002Fnative-lite.md","REFERENCE","参考文档",5068,"https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Fcouncil\u002Freferences\u002Fnative-lite.md",[62,69,76,83,90,97],{"skillNo":63,"name":64,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M7CA5PH4JT2V5WBTA5ZPJNS8","whiteboarding","Explore an idea through docs-informed dialogue before committing to a design. Capture the exploration for later filing as a brief.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Fwhiteboarding\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7CA5GZPSVNJJZYQMF0JT3EH","verify-code-docs","Check generated code documentation against source and report drift without regenerating it.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Fverify-code-docs\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":19},"01M7CA5A4HA73VF91N8VY6SDBN","transition-observation","Apply a human disposition to an observation's lifecycle state without changing its claim or body.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Ftransition-observation\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M7CA53X1ESGWKTPCEQ55JWEZ","transition-invariant","Ratify, reject, or retire an invariant on human direction. Update lifecycle metadata without changing the invariant's body.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Ftransition-invariant\u002FSKILL.md",[],[],{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":19},"01M7CA4Y2VC25NFZE0XJZ47VSX","transition-decision","Apply a human ratify, reject, or defer verdict to a recovered decision candidate; ratification creates a Proposed ADR or observation.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Ftransition-decision\u002FSKILL.md",[],[],{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":19},"01M7CA4QD9RFDC379VD9DNK90D","transition-brief","Publish or abandon a draft brief, updating lifecycle metadata and indexes without changing its body.","https:\u002F\u002Fgithub.com\u002Fbionic-coding\u002Fcrux\u002Fblob\u002Fmain\u002Fcrux\u002Fskills\u002Ftransition-brief\u002FSKILL.md",[],[]]