panel
Skill63 starsChanged 4 days agoMITView on GitHub
Multi-provider deliberation panel for COMPLEX, multi-view questions — architecture, strategy, market research, and codebase sweeps. Runs a 4-stage deliberation across the provider fleet (diverge → contrarian red-team → adversarial verify → cited synthesis), the same multi-perspective rigour the plan-gate already applies to plan reviews, generalised to arbitrary questions. Use when one model's answer isn't enough and you want convergence THROUGH disagreement + verification, not a single opinion.
14 other projects use a version of this.
Adopted by
Most recently taken up by
- wollemiahq/skillstheir own version, first seen Nov 16, 2026
- expo/exacttheir own version, first seen Nov 15, 2026
- leek/agent-skillstheir own version, first seen Nov 10, 2026
- ilchemla/agent-skillstheir own version, first seen Oct 13, 2026
- raine/consult-llm-mcptheir own version, first seen Oct 8, 2026
- scrypster/muninndbtheir own version, first seen Oct 8, 2026
Used with
Projects with it also use implement, research and 1 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
Hugging FaceSkillimplement
Make tests pass. Invoke after /write-tests produces failing tests.
Also in 3 projects that use this
NVIDIASkillresearch
Before designing a feature, research how comparable libraries solve the problem, check best practices, known pitfalls, and upstream changes. Use when starting a new feature, investigating an unfamiliar area, or the user asks to research something.
Also in 3 projects that use this
GoogleSkilltdd
Test-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.
Also in 3 projects that use this
Other versions14 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 14
- expo/exact★ 4.3ksame name
- scrypster/muninndb★ 336same name
- raine/consult-llm-mcp★ 148same name
- Reef123/OpenVera★ 9same name
- dkpapadopoulos/auto-claude-skills★ 6same name
- leek/agent-skills★ 5same name
- niieani/caara★ 5same name
- ilchemla/agent-skills★ 4same name
- wollemiahq/skills★ 2same name
- diazhh/claude-expert-panel★ 1same name
4 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository56 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- .claude/terminals/T0/CLAUDE.mdCLAUDE.md
- .claude/terminals/T1/CLAUDE.mdCLAUDE.md
- .claude/terminals/T2/CLAUDE.mdCLAUDE.md
- .claude/terminals/T3/CLAUDE.mdCLAUDE.md
- api-developerSkill
- architectSkill
- backend-developerSkill
- control-centreSkill
- data-analystSkill
- database-engineerSkill
- debuggerSkill
- excel-reporterSkill
- fabric-referenceSkill
- featureplan-kickoffSkill
- frontend-developerSkill
- horizonSkill
- intelligence-engineerSkill
- monitoring-specialistSkill
- performance-profilerSkill
- plannerSkill
- pmSkill
- python-optimizerSkill
- quality-engineerSkill
- refactor-safeSkill
- reviewerSkill
- security-engineerSkill
- supabase-expertSkill
- t0-orchestratorSkill
- test-engineerSkill
- vnx-managerSkill
- api-developerSkill
- architectSkill
- backend-developerSkill
- data-analystSkill
- database-engineerSkill
- debuggerSkill
- excel-reporterSkill
- fabric-referenceSkill
- featureplan-kickoffSkill
- frontend-developerSkill
- horizonSkill
- intelligence-engineerSkill
- monitoring-specialistSkill
- performance-profilerSkill
- plannerSkill
- pmSkill
- python-optimizerSkill
- quality-engineerSkill
- refactor-safeSkill
- reviewerSkill
- security-engineerSkill
- supabase-expertSkill
- t0-orchestratorSkill
- test-engineerSkill
- vnx-managerSkill
The file
Files
/panel — multi-provider deliberation
For hard questions where you want several strong models to genuinely deliberate, not just answer in parallel. Each stage builds on the last, so the panel converges through disagreement and verification instead of averaging opinions.
When to use
- architecture — a feature/system design where the tradeoffs are non-obvious.
- strategy — a business/product call that rests on assumptions worth stress-testing.
- research — market/competitive questions where claims need refuting.
- sweep — a codebase audit (security / correctness / dead-code / refactor).
Reach for it when a single model's answer would be a guess, and you'd otherwise open five terminals yourself.
How it works (4 stages)
- Diverge — every fleet provider (codex / kimi / claude / glm-5.2 / deepseek-harness) analyses the SAME question through a DIFFERENT mode-specific lens.
- Contrarian — one designated seat red-teams the emerging consensus: what did everyone miss, which "this is fine" is wrong.
- Verify — the top claims are adversarially checked (against the CODE for sweeps — real
file:line; against SOURCES for research — try to refute). - Synthesis — one cited report: consensus + surviving dissent + verified/refuted claims, ranked and deduped.
How seat coverage is counted (OI-1519)
"N of M lenses present" is reconciled against the t0 receipt ledger
(t0_receipts.ndjson), never taken from the exit_code a seat writes about ITSELF in
its own report frontmatter. After each stage completes, every seat's dispatch-id is
looked up in the ledger via receipt_provenance.find_receipts_by_dispatch and lands in
one of THREE outcomes:
- present — the ledger holds a decisive SUCCESS record for the seat's dispatch-id
(
statusin the ADR-035 success setdone/success/complete/completed, orverdict.decision=accept). - failed — the ledger holds a decisive FAILURE record (
statusin the ADR-035 hard-failure setfailed/failure/error/blocked/timeout/contract_invalid, orverdict.decision=reject; multiple receipts for one dispatch-id reconcile failure-beats-success — a success record never launders a recorded failure). A dispatcher that RAISED before the seat completed is failed on direct local evidence. - unmeasured — the ledger has NO decisive record for the dispatch-id (no receipt
at all, or only indecisive ones). This is a third BRANCH, not a third value: the
seat counts as NEITHER present NOR failed and is named in the report
(
**Unmeasured seats:**,[SEAT UNMEASURED — no decisive ledger record]). Reading "unmeasured" as "probably fine" repeats exactly the bug OI-1519 repaired — a measurement gap is not a lens.
Where the ledger and the seat's own frontmatter DISAGREE, the ledger WINS the count and
the divergence itself is reported (## Ledger reconciliation — divergences in the
report, result.ledger_divergences) — a silent correction would produce a panel that
cannot be weighed. When no ledger exists or can be read at all (fresh checkout) the
tally falls back to the pre-OI-1519 frontmatter measurement, flagged loudly via
ledger_available=False in the report — never silently.
This holds for the diverge fan-out AND for the sequential stages (contrarian, verify,
synthesis), which accept a seat via _first_ok under the same measurement: a
ledger-failed seat is skipped; an unmeasured seat is kept only as a last-resort
fallback so a measurement gap cannot collapse a stage to [empty] when real content
exists.
Measured case (2026-08-29, the dispatch that opened OI-1519): seat
panel-sweep-diverge-0-a6421f (codex) wrote exit_code: 0 in its own report
frontmatter while the receipt ledger records status=timeout,
verdict.decision=reject after the 600s deadline. Pre-fix it counted PRESENT (4/5
lenses reported; really 3 + a timeout). Post-fix it counts FAILED, and the divergence
is printed in the report.
Run it
python3 scripts/panel.py <mode> "<question>" [--context-file FILE] [--timeout 900] [--out FILE]Examples:
# architecture decision, grounded on a design doc
python3 scripts/panel.py architecture "Should the judge run per-receipt or batched?" --context-file docs/adr-028.md
# code sweep grounded on a diff or file list
git diff origin/main > /tmp/d.diff
python3 scripts/panel.py sweep "Review this change for security + correctness" --context-file /tmp/d.diff
# market research
python3 scripts/panel.py research "Is there an underserved MKB market for local-LLM invoice processing?"The cited report lands in unified_reports/panel-<mode>-<id>.md and prints to stdout.
Notes
- Governed lane: each panelist dispatches through the review lane and emits a receipt. Respects the provider constraints (kimi-via-cli-only, glm via OpenRouter/harness, deepseek-harness with its own key + hardening, no-anthropic-sdk) — the dispatcher routes each provider correctly.
- A dead/absent provider (e.g. deepseek without its key) degrades the panel gracefully; it never blocks the run.
- The panel is the fabric's general multi-view tool. It complements the plan-gate panel (which is scoped to plan reviews) and the t0-orchestrator skill (which owns orchestration decisions). Implementation:
scripts/lib/deliberation_panel.py.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

