chore(panel-review): tighten one-comment contract; safety cap 1 -> 7 by danielmeppiel · Pull Request #905 · microsoft/apm

Daniel Meppiel (danielmeppiel) · 2026-04-24T14:54:19Z

Summary

Tighten the panel-review one-comment contract at the prompt layer, and convert the safe-outputs cap from a hard fail-stop into a soft ceiling.

Changes

Prompt (Step 3 — Output contract)

Make it explicit that the single allowed comment must be the CEO's final synthesized verdict, not any sub-agent's intermediate output.
Instruct the orchestrator to tell each sub-agent persona: "do not post any comment; return your findings to the orchestrator."
Keep the "no GitHub API directly — use the output channel" rule.

Bloat removal

Drop the workflow-only ASCII rule from the prompt — it's a project source-code rule, doesn't apply to PR-comment prose.
Drop meta references to safe-outputs.add-comment internals — the agent doesn't need to know it runs inside gh-aw.

Safe-outputs (frontmatter)

safe-outputs.add-comment.max: 1 → 7. With the strengthened prompt-level contract above, the cap becomes a fail-soft ceiling: occasional LLM drift no longer aborts the run at the safe-outputs gate and loses the verdict.

Lockfile

.github/workflows/pr-review-panel.lock.yml — recompiled via gh aw compile pr-review-panel (0 errors / 0 warnings).

Security surface

Unchanged. Only write surface remains the safe-outputs comment channel; the permission-isolated downstream job remains the sole publisher. No new secrets, actions, or permissions. Trigger gating untouched.

Copilot

Pull request overview

This PR increases the safe-outputs.add-comment.max limit for the pr-review-panel agentic workflow so the panel can emit multiple PR comments instead of being restricted to a single comment.

Changes:

Bumped .github/workflows/pr-review-panel.mdsafe-outputs.add-comment.max from 1 to 7 and updated related workflow instructions.
Recompiled .github/workflows/pr-review-panel.lock.yml to propagate the new safe-outputs configuration.

Show a summary per file

File	Description
.github/workflows/pr-review-panel.md	Raises the safe-outputs add-comment cap and updates the prompt text about comment emission.
.github/workflows/pr-review-panel.lock.yml	Mechanical regen reflecting `add_comment.max: 7` in the compiled workflow prompt/config.

Copilot's findings

Files reviewed: 2/2 changed files
Comments generated: 2

Copilot · 2026-04-24T14:57:07Z

+# - the only write surface is safe-outputs.add-comment (max 7,
+# one per persona)


The workflow frontmatter description: still says it posts a single synthesized verdict comment, but this workflow is now configured to allow up to 7 comments. Please update the description: (and any other "single comment" wording) to match the new behavior, then recompile the lockfile so the generated header stays consistent.

Copilot · 2026-04-24T14:57:08Z

@@ -139,8 +141,8 @@ owns the review behavior; this step owns only the emission boundary.
 ## Step 4: Emit the safe output

 Post the verdict by writing the comment body to the agent output channel.
-The `safe-outputs.add-comment` job (capped at 1) will pick it up and
-post it to PR #$PR.
+The `safe-outputs.add-comment` job (capped at 7) will pick them up and
+post them to PR #$PR.


Step 2 instructs the agent to follow the apm-review-panel skill's output contract exactly, but Step 3/4 now instruct emitting up to 7 safe-outputs.add-comment items. The skill currently specifies "exactly ONE comment" and references safe-outputs.add-comment.max: 1, so these instructions are contradictory and may lead to inconsistent agent behavior. To achieve the PR goal (one comment per persona), update the apm-review-panel skill output contract/verdict template to explicitly support multi-comment emission (or keep single-comment emission here and treat max: 7 only as a safety cap).

Two coupled changes to the panel review workflow: 1. Output contract (Step 3): make it explicit that the single allowed comment MUST be the CEO's final synthesized verdict, not any sub-agent's intermediate output. The orchestrator is now instructed to tell each sub-agent: 'do not post any comment; return your findings to the orchestrator.' 2. Safe-outputs cap (frontmatter): bump add-comment max from 1 to 7. With the prompt-level contract above, the cap is just a fail-soft ceiling: when an LLM occasionally drifts and tries to emit an extra comment, the run no longer aborts at the safe-outputs gate and lose the verdict. Bloat removal: dropped the workflow-only ASCII rule from the agent prompt (it's a project source-code rule that doesn't apply to PR-comment prose) and the meta references to safe-outputs internals (the agent doesn't need to know it runs in gh-aw). Recompiled pr-review-panel.lock.yml via 'gh aw compile pr-review-panel'. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…ntract The previous wording in #905 framed the CEO as a separate emitter and told the orchestrator to instruct each sub-agent 'do not post any comment'. This produced two regressions: 1. Sub-agents launched via the task tool do not have safe-outputs MCP access to begin with -- the no-op instruction added orchestration bloat without protecting anything. 2. The 'CEO synthesizes the single verdict' phrasing made the orchestrator wait for a CEO sub-agent that was never dispatched. Result: agent_output.json={"items":[]} -- zero comments posted. Fix lives in the skill (the authoritative source), not the workflow: - SKILL.md execution checklist step 5: orchestrator runs CEO arbitration as itself, never delegates to a sub-agent. - SKILL.md execution checklist step 7: orchestrator writes the comment to safe-outputs.add-comment, never calls GitHub API. - SKILL.md output contract: cap reference updated to fail-soft ceiling of 7 (matches workflow frontmatter). Workflow .md is now a thin shell: gather PR context, then defer to the skill for routing, dispatch, arbitration, emission. No more duplicated orchestration logic at the workflow boundary. Closes#906 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…ntract (#907) The previous wording in #905 framed the CEO as a separate emitter and told the orchestrator to instruct each sub-agent 'do not post any comment'. This produced two regressions: 1. Sub-agents launched via the task tool do not have safe-outputs MCP access to begin with -- the no-op instruction added orchestration bloat without protecting anything. 2. The 'CEO synthesizes the single verdict' phrasing made the orchestrator wait for a CEO sub-agent that was never dispatched. Result: agent_output.json={"items":[]} -- zero comments posted. Fix lives in the skill (the authoritative source), not the workflow: - SKILL.md execution checklist step 5: orchestrator runs CEO arbitration as itself, never delegates to a sub-agent. - SKILL.md execution checklist step 7: orchestrator writes the comment to safe-outputs.add-comment, never calls GitHub API. - SKILL.md output contract: cap reference updated to fail-soft ceiling of 7 (matches workflow frontmatter). Workflow .md is now a thin shell: gather PR context, then defer to the skill for routing, dispatch, arbitration, emission. No more duplicated orchestration logic at the workflow boundary. Closes#906 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…905) Two coupled changes to the panel review workflow: 1. Output contract (Step 3): make it explicit that the single allowed comment MUST be the CEO's final synthesized verdict, not any sub-agent's intermediate output. The orchestrator is now instructed to tell each sub-agent: 'do not post any comment; return your findings to the orchestrator.' 2. Safe-outputs cap (frontmatter): bump add-comment max from 1 to 7. With the prompt-level contract above, the cap is just a fail-soft ceiling: when an LLM occasionally drifts and tries to emit an extra comment, the run no longer aborts at the safe-outputs gate and lose the verdict. Bloat removal: dropped the workflow-only ASCII rule from the agent prompt (it's a project source-code rule that doesn't apply to PR-comment prose) and the meta references to safe-outputs internals (the agent doesn't need to know it runs in gh-aw). Recompiled pr-review-panel.lock.yml via 'gh aw compile pr-review-panel'. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…ntract (#907) The previous wording in #905 framed the CEO as a separate emitter and told the orchestrator to instruct each sub-agent 'do not post any comment'. This produced two regressions: 1. Sub-agents launched via the task tool do not have safe-outputs MCP access to begin with -- the no-op instruction added orchestration bloat without protecting anything. 2. The 'CEO synthesizes the single verdict' phrasing made the orchestrator wait for a CEO sub-agent that was never dispatched. Result: agent_output.json={"items":[]} -- zero comments posted. Fix lives in the skill (the authoritative source), not the workflow: - SKILL.md execution checklist step 5: orchestrator runs CEO arbitration as itself, never delegates to a sub-agent. - SKILL.md execution checklist step 7: orchestrator writes the comment to safe-outputs.add-comment, never calls GitHub API. - SKILL.md output contract: cap reference updated to fail-soft ceiling of 7 (matches workflow frontmatter). Workflow .md is now a thin shell: gather PR context, then defer to the skill for routing, dispatch, arbitration, emission. No more duplicated orchestration logic at the workflow boundary. Closes#906 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

CopilotAI review requested due to automatic review settings April 24, 2026 14:54

Copilotstarted reviewing on behalf of Daniel Meppiel (danielmeppiel)April 24, 2026 14:54 View session

CopilotAI reviewed Apr 24, 2026
View reviewed changes

Daniel Meppiel (danielmeppiel)force-pushed the chore/panel-review-max-7-comments branch 3 times, most recently from a66f98e to e799f18Compare April 24, 2026 15:02

Daniel Meppiel (danielmeppiel) changed the title ~~chore(panel-review): raise safe-outputs add-comment cap from 1 to 7~~chore(panel-review): tighten one-comment contract; safety cap 1 -> 7Apr 24, 2026

Daniel Meppiel (danielmeppiel)force-pushed the chore/panel-review-max-7-comments branch from e799f18 to 570a57fCompare April 24, 2026 15:12

Daniel Meppiel (danielmeppiel) merged commit 1e6a621 into mainApr 24, 2026
7 checks passed

Daniel Meppiel (danielmeppiel) deleted the chore/panel-review-max-7-comments branch April 24, 2026 15:19

Daniel Meppiel (danielmeppiel) mentioned this pull request Apr 24, 2026
fix(panel-review): orchestrator self-arbitrates and emits in skill contract #907
Merged

This was referenced Apr 24, 2026
fix(apm-review-panel): restore in-context persona model (per agentskills.io) #908
Merged
docs(changelog): resolve 0.9.2 merge conflict, dedupe sections, add missing post-0.9.2 entries #932
Merged
chore(release): 0.9.3 #950
Merged

Navigation Menu

Uh oh!

chore(panel-review): tighten one-comment contract; safety cap 1 -> 7 - #905

chore(panel-review): tighten one-comment contract; safety cap 1 -> 7#905
Daniel Meppiel (danielmeppiel) merged 1 commit into
mainfrom
chore/panel-review-max-7-comments

Daniel Meppiel (danielmeppiel) commented Apr 24, 2026•
edited
Loading

Uh oh!

CopilotAI left a comment

Uh oh!

CopilotAIApr 24, 2026

Uh oh!

CopilotAIApr 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		# - the only write surface is safe-outputs.add-comment (max 7,
		# one per persona)

Uh oh!

Conversation

Daniel Meppiel (danielmeppiel) commented Apr 24, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Prompt (Step 3 — Output contract)

Bloat removal

Safe-outputs (frontmatter)

Lockfile

Security surface

Uh oh!

CopilotAI left a comment

Choose a reason for hiding this comment

Pull request overview

Copilot's findings

Uh oh!

CopilotAIApr 24, 2026

Choose a reason for hiding this comment

Uh oh!

CopilotAIApr 24, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Daniel Meppiel (danielmeppiel) commented Apr 24, 2026•
edited
Loading