council
Convene configured Lite, Standard or Strict council review; the owning writer records exact-subject approval evidence.
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- council
- model
- opus
- context
- fork
- metadata
- { "tags": "verification, multi-model, council, anti-hallucination", "bundles": "crux-verification, crux-docs", "triggers": "run the council | convene the council | ask the council | get a multi-model opinion", "risk_level": "medium", "requires_access": "write-run" }
- description
- Convene configured Lite, Standard or Strict council review; the owning writer records exact-subject approval evidence.
02 资源文件 1
- 参考文档references/native-lite.md4.9 KB
03 SKILL.md 正文
Crux Council Vote (Async-First)
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.
<!-- BEGIN GENERATED: runtime-compat -->
Runtime compatibility
This skill is portable across Claude Code, Codex, OpenCode, and OMP (Oh My Pi). This section overrides platform-specific labels below.
- Before running a command that uses
CRUX_PLUGIN_ROOT, set it to the installed plugin root. In Claude Code, use the value ofCLAUDE_PLUGIN_ROOT. In Codex, OpenCode, and OMP, derive it from the absolute path of this selectedSKILL.md: the plugin root is the parent of itsskills/directory (in OMPbash,skill://<name>resolves to the skill directory). In a source checkout, use the checkoutcrux/directory. - For project-local skills, use
.claude/skillsin Claude Code,.agents/skillsin Codex,.opencode/skillsin OpenCode (which also reads the singular.opencode/skill), and.omp/skillsin OMP. SetCRUX_LOCAL_SKILLS_DIRto that path before following any command below that uses it. - Translate Claude Code tool labels such as
Agent,Read,Write,Bash,WebSearch, andWebFetchto the matching capability in the current session. Codex names its own capabilities; OpenCode uses the lowercase formssubagent,read,edit,shell,websearch, andwebfetch, whereeditcovers bothEditandWrite; OMP usestask,read,edit,write,bash,grep,glob,find,todo, andweb_search(wherereadalso handles URLs andskill://URIs). Do not attempt to invoke the Claude Code labels as literal commands on another host. - Install the generated role agents before delegating:
install-codex-agentsin Codex,install-opencode-agentsin OpenCode,install-omp-agentsin OMP. Codex names themcrux_architect,crux_brainstormer,crux_commander,crux_dev_lead,crux_developer,crux_historian,crux_librarian,crux_night_gardener,crux_reviewer, andcrux_wayfinder; OpenCode and OMP use the bare role namesarchitect,brainstormer,commander,dev-lead,developer,historian,librarian,night-gardener,reviewer, andwayfinder(in OMP,scoutmay also be used for reconnaissance alongsidewayfinder). If a required role or capability is unavailable, report that truthfully instead of claiming it ran. - Argument placeholders such as
$adrand$bookbind 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 ofOPENCODE_GUIDE.mdnames the invocation-control fields OpenCode ignores. <!-- END GENERATED: runtime-compat -->
Configured councils
Read .bionic.yml through the canonical policy reader. Council mode and human
acceptance are independent. Existing projects without a council setting retain
Strict. New setup explicitly selects Standard or an owner-selected Lite mode.
- Lite: native configured Apex challenge, architect response, fresh final review. Final reviewer approval and exact-final architect agreement are both required. No OpenRouter account is required. Follow native-lite.md.
- Standard: three gateway provider seats, two approvals and architect agreement.
Address every dissent with a typed disposition. Resolved/refuted findings need
retained checkable evidence; a non-binding trade-off needs an explanation.
Single-seat dissent rule: when two seats approve and one dissents, revise
only if the finding is a binding failure or missing authorization, or states
a plausible failure, consequence and smallest fix that the architect cannot
refute and no later phase would catch. For an unresolved required safety
fact, obtain a targeted check; if it stays unresolved, withdraw the dependent
choice. A required safety fact about data policy or publishing is not a council
dispute: it is the outward-facing stop, and the one question put to the owner.
Otherwise answer on the same revision with reasoned agreement. Record
refutedwith a named check when the finding fails the blocking standard. Recordaccepted-tradeoffwith the reason when a later phase would catch it or its risk is lower than the cost of another round. When all three seats approve with nits only, record agreement; do not revise for nits. - Strict: all three gateway seats approve with no recorded blocking findings.
Standard and Strict require OPENROUTER_API_KEY through crux-env. Missing
credentials or native capabilities refuse with their cause. Never change mode,
provider, model or seat count as a fallback. Demonstrated binding failures,
missing authorization and unresolved required safety facts prevent approval.
A minority label alone is not a Standard veto.
Use cycle-council.py for ADR, IDR, ADR-creation, consultation and disposition
purposes. Consultation approves no decision. Its --identity stays stable across
renamed files. IDRs retain their cycle name across revisions. The configured
attempt limit defaults to seven for new mode-aware decisions. Stop on stalled
progress. A council dispute never goes to the owner. At the limit, or when the
same blocking finding recurs on two consecutive attempts after a revision that
answered it, the architect withdraws the decision with withdraw-decision;
no agent requests an extension. A Standard-mode single-seat dissent that the
architect disposes as accepted-tradeoff on an approved attempt is answered,
not unanswered; it does not count toward the recurrence trigger, and neither
does its repetition on a later attempt over the same subject bytes. A finding
counts toward recurrence only when it blocks approval of an attempt. A narrowed replacement names the record it
replaces, and its attempts continue that record's count. An extension or mode
change the owner records on their own initiative keeps its provenance, findings
and total count. Historical reviews keep their original policy and limit.
Gateway convening is detached by default. Initial success names a running attempt
and await_command; it is not a vote or approval. Run the returned await command
for the terminal outcome, then the existing architect disposition. --foreground
is the synchronous diagnostic form. --probe ATTEMPT reads the attempt flock.
--recover ATTEMPT sends no request: it refuses while the flock is held, recovers
the original persisted result, or marks an unowned execution interrupted. An
interrupted claim still counts, and a later convening is a separate counted attempt.
Full cycle inspect exposes exact subject/attempt/result references, recovery argv
and prerequisites without requests or writes. New gateway attempts always persist
request/outcome checkpoints under existing ownership. A persistence failure stops
further calls and positive approval. Durable negatives survive; missing outcomes
stay unknown. Recovery sends no request and resets no budget. Partial events never
approve; later convening uses prior findings as historical context beside fresh votes.
New approval/application checks current bound evidence. Pure sealed historical
validation and applied acceptance retain their standing. The cycle commands reference
gives the interfaces; historical/native-Lite callers keep their original protocols.
The runner holds ownership through result persistence; never hand-edit or delete
a claim. A missing launcher acknowledgement proves nothing: after the exact child
terminates, check its exact preallocated claim identity before reporting no claim.
Use recovery for an existing claim; do not request a substitute owner approval.
Native Lite follows its separate sealed-dispatch procedure.
Settle a dispute under the configured mode. Record the architect agreement even
when every seat approves: it names the approved sha256 and each nit id with a
disposition, apply-in-build (a note on a Build task) or declined with its
reason, in the nits list of the --architect JSON. A nit never starts another
attempt, and the approved revision stays as approved. srde output may be kept as
evidence for a refuted disposition of an IDR dissent; srde never runs on the
ADR path.
A new revision of an IDR whose earlier revision was approved is reviewed as a delta. The writer computes each top-level contents key added, removed or changed against the latest approved revision, stores it in the attempt, and checks it again against both revisions' bytes at convening and at approval. The seats receive the full new revision with the changed keys marked and judge the change. Approval binds the new revision's sha256, and a delta attempt counts toward the same limit.
A Standard result waits for the architect even when its provider votes suffice.
That result reports the outcome awaiting-architect, exits 1 and prints the command
to run; the sealed record keeps changes-required. Use --architect PATH --record <result>, where PATH names a JSON file, to record exact-revision agreement and
typed dissent; do not call another council just to add that agreement. The writer
resolves evidence artifacts safely and hashes their bytes. A targeted check for a
required fact is not another mandatory council round.
Fix-round discipline, for every mode: fix all findings (or refute one with a named check) in one revision, apply each fix at every site that states the same concept, and rerun the pre-council self-check's mechanical checks before reconvening.
The following promptbook procedures preserve historical gateway-only behavior. They do not override configured prospective modes.
Gate council or ad-hoc council
Council 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.
Inside a promptbook run, a council gate uses the council runner, run-council.py. Run it as a background command:
uv run "${CRUX_PLUGIN_ROOT}/scripts/run-council.py" <run-RUN-NNN.yaml> [--prompt N] --round N \
--question <question file> --subject <subject file> [--subject <subject file> ...] \
[--book <book file>] [--retain-subjects] [--max-tokens 64000] [--timeout 600]
--max-tokens is the first-attempt output budget for every seat (default 64000). A seat
whose reply stops on that budget is retried once at its own output limit from the model
registry. A value above the smallest selected seat limit is refused before any call.
--timeout is the per-seat-call deadline in seconds (default 600). When a reasoning seat
truncates, pass --timeout 1200: the escalated retry under 600 s is likely to time out.
Finish any open Git merge or other sequence before council. Both book formats
use the current attempt-aware gate. An implementation module uses council kind
implementation, including in a combined architectural book; combined is never
a council kind. Verify and patch retain their existing kinds and dimensions.
Select an exact formal revision with --implementation-revision <decision-path>.
Also pass that exact path as --subject with --retain-subjects. Record each subject's
write-time witness before committing it; never attribute dirty bytes after a refusal.
--migration-batch is retired: it exits 2 with historical-migration-writes-retired
before any batch is loaded, any attempt is claimed or any council call is made. Propose
clause migrations in the prospective format through migrate-adrs.
New formal closes retain context three/profile four. Historical context two/profile three
and context one/profile two remain immutable and replay-only, never selected from the book format.
--prompt is optional and defaults to the run's current prompt. The run, --round, --question and at least one --subject are required.
--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.
The 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.
The 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 <n>; never convene another round until recovery reports.[^recover]
0: it wrote, committed and verified a council record that ran.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 blockedwith it attached. A refusal prints"record": nulland leaves nothing to attach."refused": "round"means correct--roundand run again."refused": "prompt"means--promptis 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'srepairand run again at the same round.retypemeans correct the path.commit-run-workmeans commit the subject, the run's own witnessed work, withrun-work-witness.py commit <run> --path <path>. The third preflight refusal at one convening prompt commits apreflight-retries-spentrecord instead, an escalation-loop stop.[^defer]"refused": "attempt-open"means an open attempt holds the module: recover it, never claim past it.2: no council record was committed. Exit 2 never means nothing was written. When stderr namestimeout, 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 staleindex.lockin 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 withrun-council.py --recover <run> --prompt <n> --probe, and once no live council runner holds it, runrun-council.py --recover <run> --prompt <n>. Pass--prompt: without it recovery cannot recognise a record whose pending copy is already removed, and it reportsnothing-open. Recovery makes no council request and commits or recognises only the original council record.[^recover] Route each recovery result by the recovery table inrun-promptbook'sreferences/gates.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.
A step that writes a file it later passes as a subject records a run-work witness after each write: run-work-witness.py record <run> --prompt N --path <path>. 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/gates.md describes how a conductor issues and advances a gate prompt, repairs a preflight refusal and recovers an open attempt.
The 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]
The 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]
When to Use
- Choosing between architectural approaches
- Evaluating tradeoffs (cost vs accuracy, speed vs correctness)
- Any decision where multi-model consensus reduces risk
- Before committing to a design pattern or library choice
When NOT to Use
- Factual lookups ("what's the default timeout?") — a council of three models burns three API calls to confirm one fact. Use
call-llmfor a single answer. - Obvious single answers — if the choice is clear, a council just rubber-stamps it at 3x the cost and latency.
- 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-llmfor in-loop reasoning and reserve the council for the one decision that frames the loop.
A 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.
File Locations
crux/scripts/crux/council/
├── __init__.py
├── async_council.py ← USE THIS for an ad-hoc council
└── council.py ← sync fallback only
Exports
From council __init__:
council_vote, quick_council, get_opinion, synthesize_opinions,
VotingMethod, Opinion, CouncilDecision, DEFAULT_COUNCIL, MODEL_WEIGHTS,
AsyncCouncil, AsyncVisualCouncil, AsyncCouncilConfig, VisualVoteResult,
create_async_council, create_visual_council
Running it (PEP 723 driver under uv, NOT bare python3)
The council needs httpx, the HTTP client — NOT present in the bare
system python3, so running there dies with ModuleNotFoundError: No module named 'httpx'. The
canonical pattern: write your driver to a temp .py file that
starts with the standard PEP 723 header, then run it with uv, which
provisions the dependencies into a cached, isolated environment:
# /// script
# requires-python = ">=3.10"
# dependencies = ["httpx>=0.27"]
# ///
import sys
sys.path.insert(0, "${CRUX_PLUGIN_ROOT}/scripts")
from crux.council import create_async_council
# ... your driver code (see Usage below)
d=$(mktemp -d) && uv run "$d/driver.py"
Each call to a seat has its own deadline: AsyncCouncilConfig and cycle-council.py use
1200 s, while the run-council.py command line keeps 600 s through --timeout. A seat that
fails with a timeout, provider, malformed-response or truncated fault is called once
more, so the worst case per seat is two deadlines, about 40 minutes for the driver. A
truncated retry uses the seat's registry output limit. With max_tokens left unset, each
seat's first-attempt budget is half its registry limit (at least 32000, at most the limit).
A given max_tokens is the first-attempt budget for every seat. Each entry in a vote's
attempts records the max_tokens it requested. The cycle-council record's seats carry
those entries; a run-council.py record does not carry the budget. A Claude Code foreground shell call
stops at 10 minutes, so run the driver as a background command and read its output when it
exits. A shell that kills the driver first loses every seat's answer. AsyncCouncilConfig
takes max_retries 0 (no retry) or 1 (the default); any other value is refused when the
configuration is built.
Write the driver into a private per-run directory as above, never a fixed shared
path like /tmp/driver.py — a predictable name in a world-writable directory is a
symlink hazard.
When writing the temp file, substitute ${CRUX_PLUGIN_ROOT} (crux's portable
plugin-root name — in Claude Code, the value of CLAUDE_PLUGIN_ROOT; in Codex,
derived from this SKILL.md's path per the Runtime compatibility note above)
with its actual value; in a source checkout, where the
variable is unset, substitute the checkout's crux/ directory. No
PYTHONPATH, no extras, no surrounding project needed — the inline metadata
makes uv ignore any enclosing project.
Without uv installed this fails at the shell — command not found: uv
(exit 127); the script never executes. Remediation: install uv
(https://docs.astral.sh/uv/). API keys are read from ~/.crux/env via
crux_env. One key backs every seat: crux-env set OPENROUTER_API_KEY ….
ALWAYS Use Async for an Ad-Hoc Council
import asyncio
from crux.council import create_async_council, AsyncCouncilConfig
async def decide(question: str, context: str):
council = create_async_council()
decision = await council.deliberate(
prompt=f"Question: {question}\n\nContext: {context}",
system="You are a technical architecture council member.",
)
# deliberate() returns a CouncilDeliberation (crux/scripts/crux/core/data_classes.py):
# votes, consensus, consensus_confidence, key_agreements, key_disagreements,
# final_recommendation, dissent_count, confidence_adjustment (+ .has_critical_dissent).
# final_recommendation keys: action, reason, degraded, errored_seats, off_scale,
# conditioned, conditions, nits, nit_items.
print(f"Consensus: {decision.consensus}") # e.g. UNANIMOUS_APPROVE
# Consensus labels: UNANIMOUS_APPROVE, UNANIMOUS_REJECT, MAJORITY_APPROVE, MAJORITY_REJECT,
# SPLIT, NO_QUORUM, and UNANIMOUS_<TOKEN> when every responding seat returned the same
# other decision. UNANIMOUS_OFF_SCALE replaces it when the token is not 1-40 characters of
# A-Z and underscore starting with a letter, contains APPROVE or REJECT, or is AUTO_EXECUTE
# or EXECUTE_WITH_MONITORING.
# Only UNANIMOUS_APPROVE and MAJORITY_APPROVE can route to an execute action; every other
# label routes to DEFER_TO_HUMAN or ABORT. Unanimity on REVISE is agreement, not approval.
print(decision.final_recommendation["off_scale"]) # ["<provider>:<decision>", ...] for
# responding seats outside the known
# scale; each entry is bounded and
# its unprintable characters replaced
print(f"Confidence: {decision.consensus_confidence}")
# A conditioned approval: a responding seat returned APPROVE_WITH_CONDITIONS. The label
# can still read UNANIMOUS_APPROVE or MAJORITY_APPROVE, so read the flag. The route is EXECUTE_WITH_MONITORING at
# most, never AUTO_EXECUTE. The flag fires only when the prompt puts APPROVE_WITH_CONDITIONS
# on the decision scale (the built-in envelope does not).
fr = decision.final_recommendation
if fr["conditioned"]: # decision.conditioned is the same flag
for seat, points in fr["conditions"].items(): # keyed by seat; untrusted model text, carried unchanged
print(f" [condition] {seat}: {points}")
# Nits: a responding seat returned APPROVE_WITH_NITS. Nits are reported and never change
# the route. Only conditions block.
if fr["nits"]: # decision.nits is the same flag
for seat, points in fr["nit_items"].items():
print(f" [nit] {seat}: {points}")
# Both flags read False, with empty items, when no responding seat returned them. An
# errored seat never contributes.
if decision.dissent_count:
print(f"{decision.dissent_count} dissent(s) — srde output is evidence only")
for v in decision.votes: # each vote is a CouncilVote
for d in v.dissenting_points: # NOTE: dissenting_points, not dissent_points
print(f" [{v.provider}/{v.model}] {d}")
# See srde skill
# Each vote also records how its seat fared. finish_reason is the reply's finish reason
# from a closed set (stop, length, content_filter, tool_calls, error, other, missing; None
# when the seat got no reply, such as a timeout). fault_label names an errored seat's cause
# (one of timeout, provider, malformed-response, truncated, auth, rate-limit, client-config,
# refused, insufficient-credit, unexpected-status or unreachable, or unknown). A gate council
# adds served-model and served-provider-mismatch, which are never retried. retried marks
# a seat called a second time; recovered marks a seat that answered on that second call;
# first_fault_label and first_finish_reason describe its first call.
# attempts lists each call to the seat with what its reply reported.
for v in decision.votes:
print(v.provider, v.finish_reason, v.fault_label, v.retried, v.recovered)
return decision
result = asyncio.run(decide(
question="Should we use approach A or B?",
context="Full context including constraints and requirements",
))
Async with Custom Config
config = AsyncCouncilConfig(...) # customize models, weights, etc.
council = create_async_council(config=config)
decision = await council.deliberate(prompt="...", system="...")
Sync Deliberation (from async council)
council = create_async_council()
decision = council.deliberate_sync(prompt="...", system="...")
Visual Council (image analysis)
from crux.council import create_visual_council
visual_council = create_visual_council()
votes = await visual_council.analyze_image(
image_b64="<base64-encoded-image>",
prompt="Does this dashboard render correctly?",
)
# Or sync:
votes = visual_council.analyze_sync(image_b64="...", prompt="...")
The visual council retries and labels a seat as the text council does. Its result for a seat
that errored reads passed=False at confidence 0.0 and carries errored=True. Skip errored
results when you count failed checks: an errored seat is a missing check, not a failed one.
Sync Fallback (only if async is impossible)
from crux.council import council_vote, quick_council, get_opinion, VotingMethod
from crux.core.llm_caller import get_default_model
# Full council with options
decision = council_vote(
question="Should we use approach A or B?",
context="Full context here",
voting_method=VotingMethod.ARBITER, # models and arbiter default to the registry's assignment
tracer=None, # optional Tracer instance
)
# Quick one-liner (returns string)
answer = quick_council("Should we cache at the API or DB layer?")
# Get a single model's opinion
opinion = get_opinion(get_default_model("council_arbiter"), "question", context="...")
Model Config
The 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}/scripts/crux/_config/llm_router_config.json; this skill names no model.[^seats]
The router reaches every provider through the OpenRouter gateway.
API keys are read via crux_env.require(...) from ~/.crux/env (managed by crux-env).
Rules
- ALWAYS prefer async for an ad-hoc council and for agent teams (parallel teammates = concurrent calls)
- ALWAYS check
dissent_count(and each vote'sdissenting_points) — never just readconsensus - ALWAYS read
final_recommendation["conditioned"]and["nits"]: a conditioned approval can readUNANIMOUS_APPROVEorMAJORITY_APPROVEbut never routes toAUTO_EXECUTE - 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 - ALWAYS provide rich context — models deliberate better with specifics
- 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] - 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
- 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
- Treat srde output as evidence: it never settles a gate and never runs on the ADR path (see srde skill)[^srde]
- Register results on Semantic Bridge for cross-teammate visibility
[^council]: rule:council-is-never-harness-native, rule:council-gate-needs-a-runner-record [^review]: rule:review-gate-needs-a-reviewer-report [^seats]: rule:every-seat-assesses-five-dimensions, rule:council-assignment-is-validated-before-spend [^served]: rule:seat-records-requested-and-served-model, rule:served-provider-matches-a-declared-set [^defer]: rule:only-a-preflight-refusal-is-retried [^attempt]: rule:an-open-council-attempt-stops-the-gate [^match]: rule:council-evidence-matches-its-attempt [^recover]: rule:recovery-never-deliberates-again [^srde]: rule:de-wire-srde-from-adr-path, rule:blocking-finding-names-a-plausible-failure [^blocker]: rule:blocking-finding-names-a-plausible-failure [^trust]: rule:crux-trusts-its-local-environment
04 同仓库更多 Skills
whiteboarding
Explore an idea through docs-informed dialogue before committing to a design. Capture the exploration for later filing as a brief.
verify-code-docs
Check generated code documentation against source and report drift without regenerating it.
transition-observation
Apply a human disposition to an observation's lifecycle state without changing its claim or body.
transition-invariant
Ratify, reject, or retire an invariant on human direction. Update lifecycle metadata without changing the invariant's body.
transition-decision
Apply a human ratify, reject, or defer verdict to a recovered decision candidate; ratification creates a Proposed ADR or observation.
transition-brief
Publish or abandon a draft brief, updating lifecycle metadata and indexes without changing its body.