Other

/plan-council

Cross-family judge council over ONE authored plan or spec: 3 visible seats from 3 model families, self-invented rubrics, live-repo read-only validation, measured same-family bias, lift round, exact-head review loops. Triggers: council, review my plan, score the plan, judge the plan, cross-family judges, lift round, exact-head review, conditional GO, MUST_FIX. NOT for anonymized candidate ranking (/adversarial-council) or bulk verdict fleets (/judge-fleet).

$ golems-cli skills install plan-council

3 workflows

Updated 1 week ago

Review one authored plan or spec with declared cross-family judges. Disagreement is evidence: preserve each judge's independent rubric, validate claims against live local state, and let the conservative non-author-family score govern disputed lanes.

Output contract

EVERY artifact this protocol emits — ballot, lift table, convener report — leads with its table; environment caveats and preamble go BELOW the first table, never above it.

Route first

Need	Use
Review one authored plan/spec with declared judge families and measured bias	`/plan-council`
Rank multiple anonymized candidates without authorship signals	`/adversarial-council`
Fan out bulk verdict production over many prompts/items	`/judge-fleet`

Five laws

1. Seats

Convene 3 voting seats from 3 model families in visible cmux panes: R1 Opus 5, R2 GPT-5.6-Sol at xhigh, and R3 Fable 5. Never use background subagents.
Pin every seat explicitly and verify the effective model from the pane. The Fable seat uses raw claude --dangerously-skip-permissions --model claude-fable-5 because the repoGolem launcher cannot express it.
Supply no shared rubric. Each judge invents named dimensions and weights totaling 100.
Each judge scores every lane 1-10, checks the live repo read-only, makes findings falsifiable, gives GO / NO-GO / CONDITIONAL GO per gated unit, names its top three changes, signs its family, and ends with its agreed sentinel.
Each judge's FIRST output line after the ballot heading must be the scorecard TABLE header; judges write no prose before it.

Use references/seat-briefs.md and workflows/run-council.md.

2. Merge

The plan author holds no seat and scores nothing.
Do not average disagreement away. For each disputed lane: merged(lane) = min(non-author-family judge scores).
Report same-family bias per lane and the round mean. Flag a same-family judge at ≥ +2 above the conservative non-family read.

python3 <plan-council-skill-dir>/council_bias.py <ballots-dir> --author-family <author-family>

3. Lift round

For every lane below 8, ask every judge for THE ONE artifact that would lift it above 8 and an explicit research call. The governing instruction is:

one artifact each, and it must be the thing whose absence is why the score is low — not a wish list.

Run workflows/lift-round.md. Conditions addressed to the author are work: apply them and continue. They are not ratification theater. Irreversible state changes retain their own receipt-first HARD STOP.

4. Exact-head review

Every execution verdict names the exact head SHA. ACCEPT / ITERATE / MUST_FIX without that SHA is void. The ACCEPTED head is the merged head; any later head movement makes the verdict stale and requires re-review.

Every review round declares its axis and method. A stale-only round counts toward a published stop-line only when it explores the same axis as the prior round and uses an adversarial method that attempts to defeat the rules. A confirmatory happy-path round never counts; changing axis resets the counter. Publish axis, method, counter, and stop-line before dispatch. See workflows/exact-head-review.md.

5. Conditions are work

Apply council conditions assigned to the author and go. Do not park approved work behind another ratification step. This does not weaken independent data-safety gates for migrations, backfills, or other irreversible operations.

Full SKILL.md source — includes LLM directives, anti-patterns, and technical instructions stripped from the Overview tab.

Output contract

EVERY artifact this protocol emits — ballot, lift table, convener report — leads with its table; environment caveats and preamble go BELOW the first table, never above it.

Route first

Need	Use
Review one authored plan/spec with declared judge families and measured bias	`/plan-council`
Rank multiple anonymized candidates without authorship signals	`/adversarial-council`
Fan out bulk verdict production over many prompts/items	`/judge-fleet`

Five laws

1. Seats

Convene 3 voting seats from 3 model families in visible cmux panes: R1 Opus 5, R2 GPT-5.6-Sol at xhigh, and R3 Fable 5. Never use background subagents.
Pin every seat explicitly and verify the effective model from the pane. The Fable seat uses raw claude --dangerously-skip-permissions --model claude-fable-5 because the repoGolem launcher cannot express it.
Supply no shared rubric. Each judge invents named dimensions and weights totaling 100.
Each judge scores every lane 1-10, checks the live repo read-only, makes findings falsifiable, gives GO / NO-GO / CONDITIONAL GO per gated unit, names its top three changes, signs its family, and ends with its agreed sentinel.
Each judge's FIRST output line after the ballot heading must be the scorecard TABLE header; judges write no prose before it.

Use references/seat-briefs.md and workflows/run-council.md.

2. Merge

The plan author holds no seat and scores nothing.
Do not average disagreement away. For each disputed lane: merged(lane) = min(non-author-family judge scores).
Report same-family bias per lane and the round mean. Flag a same-family judge at ≥ +2 above the conservative non-family read.

python3 <plan-council-skill-dir>/council_bias.py <ballots-dir> --author-family <author-family>

3. Lift round

For every lane below 8, ask every judge for THE ONE artifact that would lift it above 8 and an explicit research call. The governing instruction is:

one artifact each, and it must be the thing whose absence is why the score is low — not a wish list.

4. Exact-head review

5. Conditions are work

Validate ballots

python3 <plan-council-skill-dir>/council_lint_cli.py ballot.md \
  --lane "W3 preconditions" --lane "3a source_class" \
  --author-seat <author-voting-seat-or-nonseat-id> --sentinel DONE_COUNCIL_R1
python3 -m pytest <plan-council-skill-dir>/tests/ -q

Rule	Required invariant
`table-not-first`	First content after the ballot heading is a markdown table.
`missing-rubric`	Multiple named rubric dimensions carry weights.
`rubric-weights-not-100`	Weights sum to 100 within ±0.5.
`unscored-lane`	Every CLI-declared lane has a numeric score.
`score-out-of-range`	Every score is within 1-10.
`unfalsifiable-finding`	At least one F-numbered finding has a file:line, command, query, PR, or issue locator. If the ballot is checkable overall, each remaining uncited finding is reported as a warning.
`missing-refuter` (warning)	No explicit `Refute me with:` / `Falsifier:` guidance; reported but never exit-1.
`missing-live-receipts` (warning)	No distinct self-verified receipts section; reported but never exit-1.
`missing-conditional-verdict`	The ballot gives at least one explicit uppercase gating verdict: GO / NO-GO / CONDITIONAL GO / BLOCK / HOLD / RESHAPE / RESCOPE / REDESIGN.
`missing-family-signature`	Signature is `— R<n> · <declared family> · <harness>`; any declared family is legal.
`missing-sentinel`	The agreed sentinel is the final non-empty line.
`author-scored`	The signed voting-seat ID is not the seat held by the plan author (`--author-seat R<n>`). When the author holds no seat, pass its non-seat ID; family equality never fails lint.

author-scored enforces the operator-declared author seat; detecting covert authorship would require adding agent identity to the signature contract.

The mechanical contract has 10 gate rule IDs + 3 warning types. Checkability is ballot-scoped: unfalsifiable-finding rejects a ballot when none of its findings has a concrete locator, then warns on individual uncited findings when at least one sibling is cited. The refuter and receipts warnings remain visible but non-gating; the source corpus proves that strong ballots can omit those house-style forms and can include a paper-contract finding beside live verified findings.

Anti-patterns

Don't	Evidence
Run judges as background subagents	Etan distrusted the council; visible panes from round one.
`tail -f` a collab for judge output	It dies silently; use marker-count polling, one monitor per lane, killed with the lane.
Assume a spawned seat booted	A lane was lost to a silently dead launcher; verify with `read_screen` within 30 seconds.
Re-brief a running judge by editing a file	Files do not bind running agents; re-brief in-pane.
Accept prose-first ballots	Use table-first, one page: running / needs-receipt / needs-rework.
Average disputed scores	Same-family bias ran +3 to +5; the conservative non-family read governs.
Park approved work behind ratification	Conditions to the author are work; apply and go.
Merge a head the reviewer never ruled on	The ACCEPTED head is the merged head.
Stop after quiet rounds that changed axis	An axis-changing round found a hole twelve prior heads missed.

Receipts

The numeric claims and worked example come from the read-only source trail documented in references/worked-example.md.

Workflows

/plan-council:exact-head-review/plan-council:lift-round/plan-council:run-council