fix(apm-review-panel): restore in-context persona model (per agentskills.io)#908
Merged
danielmeppiel merged 1 commit intomainfrom Apr 24, 2026
Merged
Conversation
…lls.io) PR #882 introduced a sub-agent dispatch model: 'dispatch one sub-agent per persona via the task tool; do not open persona files in the orchestrator thread.' This broke the panel in two ways: 1. Sub-agents launched via the task tool inherit the parent's MCP gateway -- including safeoutputs. They will happily call safeoutputs-add_comment themselves, producing per-persona comment spam instead of one synthesized verdict (run 24899623301: 2 personas posted, no synthesized comment). 2. The orchestrator hangs waiting on background sub-agents, often times out before all 5 return, and the session ends without ever reaching CEO arbitration or single-comment emission (run 24897368527: zero comments). Per agentskills.io, the skill model is progressive disclosure inside ONE agent loop: the agent loads the SKILL.md, then loads referenced files (persona .agent.md) on demand, then produces output. There is no spawning of sub-agents -- a skill is instructions+resources, not multi-agent fan-out. Fix: restore the pre-#882 in-context persona model: - 'Notes' section: rewritten to say the panel is a sequence of reasoning passes inside one agent loop; persona files are loaded on demand when you take that lens. - 'Execution checklist' step 3: replaced 'Dispatch contract' with 'Persona pass procedure' -- read persona file, take its lens, write findings to working notes, drop the lens. - 'Execution checklist' step 5: CEO arbitration is also an in-context lens (load apm-ceo.agent.md), not a separate sub-agent. - 'Execution checklist' step 7: explicit single-emission rule. - Auth Expert routing: 'dispatch' -> 'activate the lens' wording. Keeps everything valuable from #882: Hybrid E auth routing, the completeness gate, the verdict-template asset, the cap-7 fail-soft ceiling. Workflow .md is unchanged -- it already just delegates to the skill, which is exactly what the user asked for. Closes #906 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
danielmeppiel
force-pushed
the
fix/panel-review-restore-in-context-model
branch
from
1eab7c2 to
81b80fc
Compare
Contributor
There was a problem hiding this comment.
Pull request overview
Restores the apm-review-panel skill to an in-context, single-agent progressive-disclosure model (no sub-agent dispatch) to prevent per-persona emissions and ensure a single synthesized verdict.
Changes:
- Rewrites the panel contract to run persona “lenses” sequentially in one agent loop, with a completeness gate and CEO arbitration in-context.
- Adds an explicit conditional activation rule for the Auth Expert (fast-path file list + fallback self-check).
- Updates the roster/routing/quality gates/output contract to reflect the lens-based approach and single-emission rule.
Show a summary per file
| File | Description |
|---|---|
| .github/skills/apm-review-panel/SKILL.md | Updates the bundled skill used by the workflow to enforce in-context persona passes, completeness gating, and single-comment emission. |
| .apm/skills/apm-review-panel/SKILL.md | Mirrors the same skill changes in the source-of-truth .apm/ location. |
Copilot's findings
Comments suppressed due to low confidence (2)
.github/skills/apm-review-panel/SKILL.md:225
safe-outputs.add-comment.maxis documented here as a fail-soft ceiling "currently 7", but the workflow frontmatter in.github/workflows/pr-review-panel.mdsetssafe-outputs.add-comment.max: 1. This mismatch can cause the panel to follow an outdated contract; please either update the workflow cap to 7 (if intended) or adjust this text to match the workflow (or avoid hard-coding the number).
- Produce **exactly one** comment per panel run. The
`safe-outputs.add-comment.max` cap in the workflow is a fail-soft
ceiling (currently 7, one per persona slot); the one-comment
discipline lives here.
.apm/skills/apm-review-panel/SKILL.md:225
safe-outputs.add-comment.maxis documented here as a fail-soft ceiling "currently 7", but the workflow frontmatter in.github/workflows/pr-review-panel.mdsetssafe-outputs.add-comment.max: 1. This mismatch can cause the panel to follow an outdated contract; please either update the workflow cap to 7 (if intended) or adjust this text to match the workflow (or avoid hard-coding the number).
- Produce **exactly one** comment per panel run. The
`safe-outputs.add-comment.max` cap in the workflow is a fail-soft
ceiling (currently 7, one per persona slot); the one-comment
discipline lives here.
- Files reviewed: 2/2 changed files
- Comments generated: 2
Comment on lines
194
to
+199
| 6. Now (and only now) load `assets/verdict-template.md` and fill it | ||
| in with the collected findings + arbitration. | ||
| 7. Emit the filled template as exactly ONE comment via the workflow's | ||
| `safe-outputs.add-comment` channel. You (the orchestrator) write | ||
| the comment; never delegate emission to a sub-agent and never call | ||
| the GitHub API directly. | ||
|
|
||
| ### Dispatch contract | ||
|
|
||
| For each persona being dispatched, run this exact procedure: | ||
|
|
||
| 1. Dispatch one sub-agent for that persona only -- never chain | ||
| multiple personas inside a single sub-agent invocation. | ||
| 2. Pass only: | ||
| - the PR title and body summary, | ||
| - the relevant diff context for that persona's scope, | ||
| - why this persona is in scope (or, for Auth Expert, the rule | ||
| that activated it), | ||
| - the required return shape (findings only; never the final | ||
| comment text and never top-level verdict headings). | ||
| 3. Capture the raw return in working notes under | ||
| `<persona-name>: <findings>` or, for an inactive Auth Expert, | ||
| `Auth Expert inactive reason: <one sentence>`. | ||
| 4. Do not summarise unopened persona files yourself; do not paste | ||
| persona file contents into the orchestrator context. | ||
| `safe-outputs.add-comment` channel. Never call the GitHub API | ||
| directly. This is the ONLY output emission for the entire panel | ||
| run -- no per-persona comments, no progress comments. |
There was a problem hiding this comment.
This skill requires assets/verdict-template.md (referenced in the routing matrix and steps 6/Output contract), but there is no assets/ directory under .github/skills/apm-review-panel/ (and none under .apm/skills/apm-review-panel/ either). As written, step 6 will fail when the agent tries to load the missing template; please add the template file to the skill bundle (and ensure it is mirrored into both .apm/ and .github/) or remove the dependency and keep the verdict structure entirely in SKILL.md.
This issue also appears on line 222 of the same file.
Suggested change
| 6. Now (and only now) load `assets/verdict-template.md` and fill it | |
| in with the collected findings + arbitration. | |
| 7. Emit the filled template as exactly ONE comment via the workflow's | |
| `safe-outputs.add-comment` channel. You (the orchestrator) write | |
| the comment; never delegate emission to a sub-agent and never call | |
| the GitHub API directly. | |
| ### Dispatch contract | |
| For each persona being dispatched, run this exact procedure: | |
| 1. Dispatch one sub-agent for that persona only -- never chain | |
| multiple personas inside a single sub-agent invocation. | |
| 2. Pass only: | |
| - the PR title and body summary, | |
| - the relevant diff context for that persona's scope, | |
| - why this persona is in scope (or, for Auth Expert, the rule | |
| that activated it), | |
| - the required return shape (findings only; never the final | |
| comment text and never top-level verdict headings). | |
| 3. Capture the raw return in working notes under | |
| `<persona-name>: <findings>` or, for an inactive Auth Expert, | |
| `Auth Expert inactive reason: <one sentence>`. | |
| 4. Do not summarise unopened persona files yourself; do not paste | |
| persona file contents into the orchestrator context. | |
| `safe-outputs.add-comment` channel. Never call the GitHub API | |
| directly. This is the ONLY output emission for the entire panel | |
| run -- no per-persona comments, no progress comments. | |
| 6. Now (and only now) assemble the final verdict using the | |
| `## Output contract` section in this file. Do not load any | |
| external template file. | |
| 7. Emit the completed verdict as exactly ONE comment via the | |
| workflow's `safe-outputs.add-comment` channel. Never call the | |
| GitHub API directly. This is the ONLY output emission for the | |
| entire panel run -- no per-persona comments, no progress | |
| comments. |
Comment on lines
194
to
196
| 6. Now (and only now) load `assets/verdict-template.md` and fill it | ||
| in with the collected findings + arbitration. | ||
| 7. Emit the filled template as exactly ONE comment via the workflow's |
There was a problem hiding this comment.
This skill requires assets/verdict-template.md (referenced in the routing matrix and steps 6/Output contract), but there is no assets/ directory under .apm/skills/apm-review-panel/ (and none under .github/skills/apm-review-panel/ either). As written, step 6 will fail when the agent tries to load the missing template; please add the template file to the skill bundle (and ensure it is mirrored into both .apm/ and .github/) or remove the dependency and keep the verdict structure entirely in SKILL.md.
This issue also appears on line 222 of the same file.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
PR #907 added a 'do not delegate' clause for emission, but the panel still fails. Run 24899623301: 2 personas (Growth Hacker, Supply Chain) posted their findings as separate comments via
safeoutputs-add_comment, no synthesized CEO verdict. Earlier run 24897368527: zero comments.Root cause
The user noted: the panel was working way better before #882. Confirmed via git history.
PR #882 changed the skill from a clean progressive-disclosure model into a sub-agent dispatch model:
This breaks two ways:
Sub-agents launched via the
tasktool inherit the parent's MCP gateway -- includingsafeoutputs. They will happily callsafeoutputs-add_commentthemselves. The 'instruct each sub-agent: do not post' clause from chore(panel-review): tighten one-comment contract; safety cap 1 -> 7 #905 was lossy guidance, not enforcement. Result: per-persona comment spam.The orchestrator hangs waiting on background sub-agents. It frequently times out before all 5 return, and the session ends without ever reaching CEO arbitration or single-comment emission.
Per agentskills.io
A skill is instructions + resources loaded into ONE agent loop. There is no spawning of sub-agents. The pre-#882 SKILL.md got this right.
Fix
Restore the in-context persona model:
apm-ceo.agent.md), not a separate sub-agent.What's preserved from #882
assets/verdict-template.mdas the verdict shapeWorkflow .md is unchanged
It already just says 'use the apm-review-panel skill on this PR' -- exactly what the user asked for. All review behavior lives in the skill.
Verification
Will be verified live by re-applying the
panel-reviewlabel on PR #889 after merge.Closes #906.