sam-council

name

sam-council

description

Rapidly triage or fully falsify consequential system-development plans through blind specialist reviews, explicit rebuttals, bounded revision rounds, and evidence-weighted decisions. Use for architecture, features, migrations, incidents, releases, security-sensitive work, expensive changes, or uncertain and hard-to-reverse technical decisions on any agent platform; use multi-provider confrontation only when explicitly requested.

Sam Council

Purpose

Turn a consequential software plan into a falsifiable thesis, expose its weakest assumptions through independent review, and issue a traceable result. Optimize for decisive evidence per model call. Remain independent of any vendor, model, CLI, agent API, or fixed concurrency limit.

Non-Negotiable Contract

Remain read-only in the target system. Use distinct workers for every seat and keep first-pass reviews blind. Answer every objection explicitly. Decide by evidence and risk, never by vote count. Never report full approval while a supported blocker, untreated high risk, unverified critical assumption, or unvalidated correction remains. Retain raw responses in scratch space until the machine report validates. Never invent an unavailable worker or provider.

fast is a triage profile, not a weaker approval path. It may return only TRIAGE_PASS, ESCALATE_TO_FULL, or BLOCKED. Only full may return APPROVED, APPROVED_WITH_CONDITIONS, REVISE, or BLOCKED.

Token Saver inheritance: when the host provides RC_TOKEN_SAVER_EXECUTION_RECEIPT_V1, every controlled seat and verifier must inherit that content-free receipt and its authorized capability/lane environment unchanged. Never reconstruct or widen admission. A missing, malformed, denied, cross-user, or provider-mismatched receipt is raw fail-open input. Never put Skills, exact-output commands, prompts, transcripts, secrets, or full reviewer responses into the receipt. Skills and exact-output evidence remain lossless. Do not claim billing or quota savings.

Every controlled seat/verifier lifetime must be bracketed by the provider-neutral telemetry bridge:

telemetry_command="${REMOTE_CODE_SUBAGENT_TELEMETRY_COMMAND:-distill}"
child_run="$("$telemetry_command" subagent begin --node '<stable-seat-id>')"
# run the seat or verifier
"$telemetry_command" subagent end --run-id "$child_run" --status completed

Use failed or cancelled on the corresponding terminal path. Preserve the host receipt and returned child run id through retries. Bridge unavailability means raw execution plus an explicit Subagents proof gap — never invent a Done row. Do not require Distill to process Skill bodies or exact output; the bridge is lifetime telemetry only.

Required resources

Read these files completely before running the council:

1. references/reviewer-lenses.md for profile seats and conditional specialists.
2. references/provider-matrix.md for portable runtime discovery, topology, effort, and scheduling.
3. references/prompt-contract.md for compact packets and response limits.
4. references/output-contract.md before writing the report.

Select profile and topology

Choose the smallest valid profile before forming the thesis.

Profile	Use when	First pass	Verification	Rounds
fast	Reversible, bounded work without a triggered specialist domain	3 composite seats	1 fresh triage arbiter	exactly 1
full	Explicitly requested, costly, irreversible, production-critical, security/privacy, migration, compliance, or cross-provider	6 required seats plus applicable specialists	3 fresh verifiers	1 by default, at most 3

Default an explicit sam-council request to fast only when every full trigger is absent. Escalate before dispatch when uncertain. A fast result that finds a blocker, high risk, critical unknown, displaced material risk, or applicable conditional specialist returns ESCALATE_TO_FULL; do not imply approval.

Topology is independent of profile:

• single-host is always the default and uses the active host runtime.
• multi-provider is explicit opt-in only, requires at least two independent providers, and forces full.
• Provider identifiers are runtime-supplied lowercase slugs. Examples include codex, claude-code, and grok; they are not an allowlist.
• If exactly one provider is named, stay single-host on it.

Portable execution policy

Discover and record the host's actual adapter, model label, supported effort controls, and maximum safe parallel workers. Never require a named model or vendor-specific command.

• Request the host's closest available medium tier for blind reviewers, conditional specialists, closure verification, and system verification.
• Request the closest available high tier only for the arbiter or meta-arbiter. If effort is not configurable, use host-default and record it.
• Dispatch independent seats concurrently up to the discovered safe capacity. Run one wave when capacity covers all seats; otherwise use the minimum number of batches. Never serialize independent seats unnecessarily.
• Evaluate conditional-seat applicability once in the controller. In fast, any applicable specialist triggers escalation rather than another seat. In full, dispatch every applicable specialist.
• Use a relevant-only packet. Prefer exact file ranges, diffs, receipts, and user constraints; exclude unrelated repository or conversation history.
• Cap each reviewer at 3 material objections and 1,000 words. Prefer fewer load-bearing objections. A reviewer may return NO_MATERIAL_OBJECTION.
• Do not run the report validator or skill harness inside reviewer workers.

Record the policy and actual runtime in the report. A host limitation is not a reason to fabricate compliance: use the nearest supported capability and make the deviation explicit.

Seats

Fast

Create three blind workers:

1. frame-evidence: problem frame, logic, and load-bearing assumptions.
2. delivery-failure: execution, operations, abuse, edge cases, and failure.
3. simplification: cheaper, smaller, reversible, or existing alternatives.

After revision, use one fresh triage-arbiter. It checks closure and displaced risk, but cannot grant full approval.

Full

Create six blind workers: logic, assumptions, execution, adversarial, alternatives, and problem-frame. Add every applicable conditional seat from references/reviewer-lenses.md.

After revision, use fresh closure-verifier, system-verifier, and arbiter. In multi-provider mode, run the required blind panel per provider with {provider}/{seat} IDs and replace arbiter with meta-arbiter.

Evidence rules

Assign stable IDs: evidence E-###, assumptions A-###, objections O-R<round>-###, and theses T-###. Classify evidence as VERIFIED, OBSERVED, INFERRED, ASSUMED, or UNKNOWN; repetition never upgrades it.

Every objection must state a falsifiable claim, failure mode, severity, confidence, premise/evidence IDs, required proof, and smallest correction. Deduplicate by failure mechanism while preserving all supporting_reviewer_ids. A supported blocker outweighs any number of passes.

Workflow

1. Freeze the objective, decision, scope, invariants, constraints, no-go surfaces, owner, evidence, assumptions, selected profile/topology, execution policy, runtime capability, seat count, and batch plan.
2. Write T-001 as a concise falsifiable thesis with approach, interfaces, state/failure boundaries, steps, alternatives, tests, rollout/recovery, observability, risks, measurable success, and recheck triggers.
3. Build one frozen relevant-only packet. Give every worker the same core plus only its lens. Preserve each raw response.
4. Dispatch blind seats at maximum safe concurrency. Do not reveal peer output until every seat in the pass is terminal.
5. Normalize and deduplicate objections without voting. Preserve unsupported claims as UNSUPPORTED rather than silently dropping them.
6. In explicit multi-provider runs, confront provider claims as described in references/provider-matrix.md.
7. Cross-examine the author. Use exactly one disposition per objection: ACCEPT, PARTIAL, REJECT, INVESTIGATE, or ACCEPT_RISK.
8. Publish the revised thesis and map every objection to its response, change, validation, and status.
9. Run the fresh verification panel at maximum safe concurrency.
10. Stop after round one by default. Return a terminal result instead of automatically starting another round. Continue to round two or three only when the user explicitly asks to validate a revised thesis. Never exceed three rounds; stop after two rounds without material progress.
11. Validate the scratch report with python3 -B scripts/validate_council_report.py council-report.json.

Decision rules

For fast:

• TRIAGE_PASS: no supported blocker/high, no critical unknown, no specialist trigger, and the fresh triage arbiter found no material displaced risk.
• ESCALATE_TO_FULL: any blocker/high, critical unknown, specialist trigger, multi-provider requirement, or material new risk exists.
• BLOCKED: distinct workers, evidence, runtime capability, or authority is unavailable.

For full:

• APPROVED: no supported blocker/high remains; critical assumptions and corrections are verified.
• APPROVED_WITH_CONDITIONS: no blocker remains; every high is mitigated or explicitly accepted by the decision owner, with owned gated conditions.
• REVISE: an actionable material objection or unknown remains after the current authorized round.
• BLOCKED: required evidence, independence, authority, provider, or runtime capability is unavailable.

Never treat TRIAGE_PASS as approval. Never accept a blocker as residual risk or a high risk on the user's behalf.

Output

Follow references/output-contract.md. Report the terminal result, confidence, profile, topology, providers, actual runtime bindings, batch plan, thesis, evidence, objections, responses, verification, conditions, residual uncertainty, and validator result. Disclose missing raw history in historical_record_limitations.

Run scripts/test_council_harness.py only when changing this skill. Never call a blocked or invalid result approved.