Panel

Run a role-asymmetric advisory panel. Each role analyzes the same task from a single expert lens; the agent synthesizes a PM-style trade-off resolution. Use this when a decision spans multiple domains that pull in different directions and one of them shouldn't silently win. For peer-style brainstorming with no role separation, use /collab. For multi-model review of an existing diff with identical prompts, use /review-panel.

Load the consult-llm skill before proceeding — it defines the invocation contract (heredocs, timeouts, --run, prompt files, thread IDs). Do not call the CLI without loading it first.

Available models

Selectors resolvable in this environment (depends on configured API keys):

!`consult-llm models`

Argument handling

Arguments: $ARGUMENTS

Check $ARGUMENTS for flags:

Role selection:

• --roles <comma-separated> — explicit kebab-case role list (e.g. --roles security,api-design,ops-readiness). Trim whitespace; reject duplicates and empty entries.
• (none) — the agent picks the roles based on the task focus and the Phase 1 context. Default starting point is architect, security, maintainability, test-strategist; swap individual roles to fit the actual work. See "Example role bundles" near the bottom of this file as a reference. Pick 3–5 roles; do not exceed 5.

After picking, show the chosen roles to the user before Round 1 so they can override with --roles if the inference is off (e.g. "Roles: migration-architect, security, data-integrity, rollback-operator. Override with --roles if needed.").

Mode flags:

• --mode design|review — default design.
  • design is a forward-looking proposal panel. Uses --task plan internally.
  • review is a critical pass on an existing diff. Uses --task review internally and adds --diff-files/--diff-base.
• --diff-base <ref> — review mode only. Default is auto-detected (see Phase 1). Pass an explicit ref to override.
• --react — opt in to a second round where each role responds to the agent's draft synthesis on its own thread.

Model assignment: any --<selector> from the Models block selects a backing model. Repeat for multiple. With no selectors, use all listed selectors in their listed order. Roles map to selectors positionally — first role to first selector, etc. This workflow requires each role to end up on a distinct resolved model; enforce that before invoking --run.

Strip all flags; the remainder is the panel focus. In design mode, treat it as the proposal/decision under analysis. In review mode, treat it as optional review focus.

Fail-fast preconditions

Stop before Round 1 with a clear error if any of these fail:

• Role count > 5 (--run max).
• Role count exceeds the number of distinct selected/available models.
• Selectors resolve to duplicate underlying models.

Error format:

panel: <N> roles requested, only <M> distinct models available.
Roles:     <role list>
Selectors: <selector list>
Either pick fewer roles (--roles ...) or pass distinct --<selector> flags.

Phase 0: Load consult-llm skill

Load it now. Follow its invocation contract for every CLI call.

Phase 1: Gather shared context

Build a context summary that every role sees verbatim. Reasonable assumptions only — do not ask the user clarifying questions.

Both modes:

• Use Glob/Grep/Read to find files, patterns, conventions, constraints related to the focus.
• Before planning or consulting, do enough research to understand how the requested behavior actually works. Before starting, think about what resources would be useful to obtain first: relevant source files, tests, logs, generated files, config, examples, command output, external docs, or authoritative upstream source. Gather the cheapest useful evidence before forming a plan.
• Do not stop at the first plausible file, definition, setting, or example. Follow references, callers, related tests, and runtime usage until you can explain the current behavior and the likely impact of changing it.
• Note compatibility requirements, security boundaries, deployment concerns, prior decisions, known unknowns.
• Exclude generated files, lockfiles, vendored dependencies unless central.

Design mode — pass relevant source files as shared -f <path> to the panel call.

Review mode — resolve <diff-base> using the same logic as review-panel/SKILL.md Phase 1 (prefer @{upstream}, fall back to merge-base with the detected main branch, fall back to HEAD). Show the resolved base to the user before running the panel.

List changed files:

git diff --name-only --diff-filter=d <diff-base>
git ls-files --others --exclude-standard

Pass tracked changed files as shared --diff-files <path> with --diff-base <ref>. Pass untracked files as -f <path> so roles see full contents. Exclude binary files and lockfiles. If there is no diff, stop and report nothing to review.

Prepare the shared context summary as plain markdown — focus, mode, key files, constraints, known unknowns.

Phase 2: Round 1 — parallel panel

Write one prompt file per role with mktemp. Each file contains the persona framing, the panel focus, and the shared context summary.

Persona prompt — design mode:

You are the {role} specialist on an advisory panel. Speak only from your assigned perspective. Do not assume other roles' responsibilities.

Mode: design — analyze this proposed work before implementation.

Panel focus:
[focus]

Shared context:
[context summary]

Output exactly these sections:
1) Key observations — what stands out from your domain's viewpoint?
2) Risks / opportunities — what could go wrong or be improved?
3) Non-negotiable requirements — what must be true (or must not be true) for the solution to be acceptable from your perspective?
4) Open questions for other roles — what do you need from other specialists?

Acknowledge conflicts with other perspectives but do not resolve them — that is the agent's job.

Persona prompt — review mode:

Same structure, but framing changes to "critical pass on an existing diff". Replace section 1 with "what stands out when reading this diff (cite file:line where possible)" and section 2 with "concrete defects, regressions, safety issues, compatibility breaks, missing tests from your domain". Add: "Do not propose fixes. Do not soften findings."

Invoke consult-llm once with one --run per role. For each run pass model=<selector>,prompt-file=<tempfile>. Add --task plan for design mode or --task review --diff-base <ref> for review mode. Do not pass -m alongside --run.

Save the per-role thread IDs from each section header ([thread_id:<id>]) keyed by role name — needed for --react and the final artifact.

Phase 3: React round (only with --react)

Draft a provisional synthesis using the Phase 4 structure. This draft is internal; do not show the user yet.

Write one continuation prompt file per role and invoke consult-llm again with one --run per role, passing model=<selector>,thread=<role-thread-id>,prompt-file=<tempfile>. Keep the same shared -f/--diff-* context.

Reaction prompt template:

The agent has drafted this synthesis from the Round 1 panel:

[draft synthesis]

From your assigned {role} perspective:
1) Are your domain concerns adequately addressed? If not, what is missing?
2) What is truly non-negotiable — a hard requirement you would block on?
3) Which trade-offs do you reject? Be specific about why.
4) Any disagreement that should remain unresolved?

Do not resolve cross-role conflicts. Flag them for the agent.

Update the synthesis using the reactions. New non-negotiables and rejected trade-offs go into the final artifact; if a role rejects on a domain-critical point, either resolve it explicitly or preserve it under unresolved disagreements.

Phase 4: Synthesize trade-offs

The agent owns the resolution. The roles advise; you decide.

Resolution rules:

• Defer to security on safety conflicts: auth, data exposure, integrity, destructive actions, privacy, abuse, supply chain risk. If security (or any equivalent role) flags a concrete risk, the resolution must neutralize it.
• Prefer simpler when trade-offs balance.
• Preserve unresolved disagreements explicitly — do not fake consensus. If you cannot resolve a conflict with high confidence, name it as Unresolved and state what input would settle it.
• If the final recommendation accepts a risk, name the risk and the reason it's acceptable.
• If the recommendation depends on later validation, make that dependency explicit.

Phase 5: Save and report

Save the synthesis to history/<YYYY-MM-DD>-panel-<topic>.md (the history/ convention from CLAUDE.md). Derive <topic> from the current branch name (sanitized to kebab-case); fall back to a short slug from the panel focus when on the main branch or detached HEAD. Print the saved path.

Artifact template:

# Panel: <topic>

**Mode:** design|review
**Roles:** <role> (<selector>), <role> (<selector>)
**Diff base:** <ref>   _(review mode only)_

## Panel summary

- **<role>:** one-bullet summary of this role's position.
- **<role>:** one-bullet summary.

## Areas of agreement

- <point of multi-role consensus>

## Conflicts and resolutions

### <short conflict title>
- **<role A>** said: <concise paraphrase or quote>
- **<role B>** said: <position>

**Resolution:** <agent decision>. **Reasoning:** <why this trade-off wins, citing the resolution rules>.

## Unresolved disagreements

- **<issue>:** <roles involved>. <what remains unresolved; what decision or input would resolve it>.
- If none, write `None`.

## Final recommendation

<One paragraph: chosen direction, required constraints, rejected alternatives, validation needed before implementation or merge.>

## Thread map

- **<role>:** `<selector>` / `<thread_id>`

The thread map lets a follow-up --react run or manual consult-llm -t <id> continue any role's conversation later.

Print the saved path and a short final-recommendation summary to the user.

Critical rules

• Strict asymmetry. Each role gets only its own persona prompt. Never leak other roles' prompts. Roles drifting into general commentary weaken the panel — enforce focus through the persona framing.
• Distinct models per role. Validate this workflow's distinct-model requirement before invoking --run. Fail fast with the error format above.
• Mode → task mapping. design ⇒ --task plan. review ⇒ --task review plus --diff-files/--diff-base.
• --react continues threads. Do not start fresh threads for the reaction round.
• Defer to security on safety conflicts. Prefer simpler when trade-offs balance.
• Preserve unresolved disagreements. No fake consensus.
• Advisory only. The panel produces a report. Do not modify source files or commit as part of this skill — hand the artifact back to the user.

Example role bundles

Reference starting points when the agent is picking roles. Use them as inspiration, not canonical sets — swap individual roles to fit the actual task. Each bundle is 4 roles; trim or extend within the 3–5 range.

A role label is just a kebab-case persona handed to the LLM — no validation, no special casing. Coin new ones freely (e.g. i18n, bundle-size, fuzzing-strategist) when the task warrants.