Skip to content

leek/agent-skills / skills/panel

Convene a panel of agent CLIs (grok, claude, codex, agy) on a prompt or a code review, then report one consensus grouped by how much they agree.

Panel

Put one task in front of four agent CLIs (grok, claude, codex, agy) run each in its own subagent, headless, then report a single consensus grouped by how much the panel agrees.

Each CLI runs as its own latest/default model. Never pass --model to any of them. The point of the panel is four independent models; pinning a model defeats it.

Process

1. Read the task and pick the shape

  • Review when the user asks to review a branch, PR, WIP, or "review since X". Pin the fixed point with code-review's script, before fanning out: bash <this-skill-dir>/../code-review/scripts/review-scope.sh <fixed-point> (default main when the user names none; add --wip for uncommitted work in progress, so the diff also covers staged, unsaved, and untracked changes). It prints JSON: diff_cmd is the pinned diff (three-dot, against the merge-base), head_sha the HEAD step 3 uses. Exit 1 is a bad ref or an empty diff: stop here, not inside four subagents. Then read references/review-shape.md: it holds the review task string, step 3's read-only guard, and step 5's ReportFindings mapping.
  • Prompt otherwise; the user's text is the task, passed word for word.

Write the one task string now. Every CLI receives the identical string, so the four results compare cleanly. Start it with this line, so a panelist never convenes its own panel when the task mentions panels, panelists, or other CLIs:

You are one panelist. Answer the task below yourself, directly. Do not run a panel, launch another agent CLI, or start subagents.

$SCRATCH is the session's scratchpad directory where the harness has one, otherwise one mktemp -d; hand subagents its absolute path, since they do not share your shell. Save the string to $SCRATCH/panel-task.txt with the file-write tool (where the harness has none, a quoted heredoc: cat > "$SCRATCH/panel-task.txt" <<'PANEL_EOF'). The launcher reads the file, so no launch command contains the task's words, and a guard hook that reads the command (task text that names git commit, say) cannot block the launch.

2. Check who is on the panel

Run command -v for each of grok, claude, codex, agy. Drop any that is absent and record it for the final report. If fewer than two remain, there is no consensus to form: say which are missing and stop.

3. Fan out: one subagent per CLI, in parallel

Spawn one subagent per present CLI, all in a single message so they run at once (in Claude Code with this plugin installed, subagent_type=leek-skills:panelist: a small, cheap wrapper, since the thinking happens inside the CLI it runs; in other harnesses, dispatch and wait per harness-tools, one per CLI, all at once). Give each subagent the path of the same task file and, in full, the brief in references/panelist-brief.md, read on every run: launch once through scripts/launch.sh with literal absolute paths, run it detached, wait with one call and never poll, distill, return { cli, status, points[] }.

In review shape, apply Keep review read-only from references/review-shape.md: worktree isolation or a working-tree baseline before the fan-out, a check for stray writes after every subagent returns.

Per-CLI notes: why each flag in scripts/launch.sh is there, read-only options and their traps, structured-output flags, absent-vs-errored, auth: live in references/cli-invocations.md. Read it only when a CLI misbehaves or the panel roster changes.

4. Cluster into consensus

First check the panel had a quorum: at least two CLIs must have responded, not just been present. With fewer, there is no consensus: report the single result plainly and say the panel was short, with the step 5 ping.

Collect the results. Only a returned { cli, status, points[] } object is a result: a panelist pauses while its CLI runs, and a notification without that object is interim, so end the turn with no reply and wait for the next one. Match points that make the same claim, even when worded differently or aimed at the same area. Grade each cluster by the share of the responders (N) that raised it, and state N so a grade is never read as more agreement than it holds:

  • Unanimous: all N.
  • Majority: more than half, but not all.
  • Split: exactly half (a tie; only when N is even).
  • Lone: fewer than half.
  • Conflict: CLIs directly disagree on the same point.

Clustering is a judgement call. When two points are plausibly the same, join them; a borderline join that lifts a point to a higher grade is the whole value of the panel.

5. Report the consensus

Print the clusters, most-agreed first, under ## Unanimous, ## Majority, ## Split, ## Lone, ## Conflict (skip any that are empty). One line per point: the claim, its location, the CLIs that raised it in brackets, e.g. [claude, codex, grok], and its one-line reason. For a Conflict, state both sides and who holds each.

In review shape, where the host has ReportFindings, report the clusters through it instead, per references/review-shape.md. Otherwise print every cluster.

Then a one-line summary: how many points at each grade, the responder count N, and any CLI that was absent or errored. Ping the user that it is ready (PushNotification where available, otherwise the report itself is the signal).

Close with a Recommended next steps block:

  1. Fix list: every point at every grade, Unanimous through Lone, as one checklist. Do not curate: a Lone finding is a real finding until verified otherwise. For a Conflict, list both sides and say which one the code supports. After ReportFindings, the reported findings are the list; name their count instead of repeating them.
  2. Offer the shortcut: ask whether to verify and apply the whole list as one commit, naming the count, e.g. "Reply fix all to verify and apply all 11 points."

On fix all, work every point: verify each against the code, apply the ones that hold, and report the ones that do not with a one-line reason. Never apply only the confident subset and hand the rest back as "weigh individually"; the user asked for the panel so they would not have to re-adjudicate it. Ask only about a point that conflicts with a recorded decision.

Safety

Headless mode auto-approves every tool, so each CLI could write files even though the brief says review only. Two guards, in order: run the panel on a clean working tree, and in review shape isolate or revert stray writes per step 3. Per-CLI read-only flags exist but behave inconsistently (see references/cli-invocations.md), so step 3, not a flag, is what enforces read-only.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 469kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 321kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 321kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 321kUpdated Claude CodeCodexVerifiedInstall