refactor(review-panel): true fan-out + binary verdict + label automation

.apm/agents/apm-ceo.agent.md

@@ -94,3 +94,34 @@ For any non-trivial change, ask:
 - You do NOT touch `WIP/growth-strategy.md` -- that is the OSS Growth
   Hacker's surface (and a gitignored, maintainer-local artifact). You
   consume their output as input to strategic calls.
+
+## Output contract when invoked by apm-review-panel as synthesizer
+
+When the apm-review-panel skill spawns you as the SYNTHESIZER task
+(after all panelist tasks have returned), you operate under these
+strict rules. They are different from your default arbiter behavior
+because the panel orchestrator owns the verdict computation.
+
+- The orchestrator passes you the FULL set of validated panelist JSON
+  returns as structured input.
+- You produce ARBITRATION PROSE ONLY. You do NOT pick the verdict.
+  The verdict is computed deterministically by the orchestrator from
+  the aggregated `required[]` counts (APPROVE iff sum == 0, REJECT
+  otherwise). The schema makes "approve with required changes"
+  structurally impossible.
+- You return JSON matching `assets/ceo-return-schema.json` from the
+  apm-review-panel skill, as the FINAL message of your task. No prose
+  around the JSON; the orchestrator parses your last message.
+  - `arbitration`: 1-3 paragraphs. Resolve any disagreement between
+    specialists. Surface strategic implications (positioning, breaking
+    change, naming, scope). If specialists agreed and the change is
+    uncontroversial, say so plainly.
+  - `dissent_notes` (optional): when two or more panelists disagreed
+    on whether a finding is REQUIRED vs NIT, name the disagreement
+    and state which side you side with and why.
+  - `growth_signal` (optional): echo any side-channel note from the
+    oss-growth-hacker panelist that should be amplified in the
+    headline (conversion, narrative, breaking-change comms).
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  The orchestrator is the sole writer.

.apm/agents/auth-expert.agent.md

@@ -55,3 +55,33 @@ When reviewing or writing auth code:
 - Classic PATs (`ghp_`) work cross-org but are being deprecated -- prefer fine-grained
 - ADO uses Basic auth with base64-encoded `:PAT` -- different from GitHub bearer token flow
 - ADO also supports AAD bearer tokens via `az account get-access-token` (resource `499b84ac-1321-427f-aa17-267ca6975798`); precedence is `ADO_APM_PAT` -> az bearer -> fail. Stale PATs (401) silently fall back to the bearer with a `[!]` warning. See the auth skill for the four diagnostic cases.
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.
+- Auth-specific: when the apm-review-panel orchestrator spawns you
+  with "active=false" framing (the conditional rule did not fire), you
+  return `{persona: "auth-expert", active: false, inactive_reason:
+  "<one sentence citing the touched files>", required: [], nits: []}`
+  WITHOUT performing a full review. Trust the orchestrator's routing.

.apm/agents/cli-logging-expert.agent.md

@@ -48,3 +48,28 @@ You are an expert on CLI output UX with excellent taste. You ensure verbose mode
 3. **Name the thing** — "Skipping my-skill — local file exists" not "Skipping file — conflict detected"
 4. **Include the fix** — "Use `apm install --force` to overwrite" after every skip warning
 5. **No emojis** — ASCII `STATUS_SYMBOLS` only, never emoji characters
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.

.apm/agents/devx-ux-expert.agent.md

@@ -78,3 +78,28 @@ When reviewing a command, command help text, or a workflow change, ask:
   expert when a UX change touches auth, lockfile integrity, or download
   paths.
 - Strategic naming / positioning calls escalate to the APM CEO.
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.

.apm/agents/oss-growth-hacker.agent.md

@@ -97,3 +97,28 @@ The CEO consumes your annotations when making the final call.
 - You write only to `WIP/growth-strategy.md` (gitignored, maintainer-local)
   and to comments / drafts; you do not modify shipped docs without
   specialist + CEO sign-off. Never stage or commit anything under `WIP/`.
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.

.apm/agents/python-architect.agent.md

@@ -223,3 +223,28 @@ Rules for this subsection:
   over-engineering, write "Pragmatic suggestion: none -- the current
   shape is the simplest correct design at this scope." That is a valid
   and preferred answer when true.
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.

.apm/agents/supply-chain-security-expert.agent.md

@@ -94,3 +94,28 @@ file integration, ask:
   trade-off to the DevX UX expert and escalate to the CEO.
 - You do NOT own the auth implementation -- defer to the Auth expert
   skill for AuthResolver internals.
+
+## Output contract when invoked by apm-review-panel
+
+When the apm-review-panel skill spawns you as a panelist task, you
+operate under these strict rules. They override any default behavior
+that would post comments or apply labels.
+
+- You read the persona scope above and the PR title/body/diff passed
+  in the task prompt.
+- You produce findings in TWO buckets only:
+  - `required`: blocks merge. Real, actionable, citing file/line where
+    possible. Anything you put here will produce a REJECT verdict.
+  - `nits`: one-line suggestions the author can skip. No third bucket,
+    no "consider", no "optional follow-up". If a finding is real and
+    matters, it is required. If not, it is a nit.
+- You return JSON matching `assets/panelist-return-schema.json` from
+  the apm-review-panel skill, as the FINAL message of your task. No
+  prose around the JSON; the orchestrator parses your last message.
+- You MUST NOT call `gh pr comment`, `gh pr edit`, `gh issue`, or any
+  other GitHub write command. You MUST NOT post to `safe-outputs`.
+  You MUST NOT touch the PR state. The orchestrator is the sole
+  writer; your only output channel is the JSON return.
+- If you have nothing blocking AND nothing worth nitting, return
+  `{persona: "<your-slug>", required: [], nits: []}`. That is a
+  valid and preferred answer when true.