| name | apm-review-panel |
| description | Use this skill to run a multi-persona expert advisory review on a labelled pull request in microsoft/apm. The panel fans out to five mandatory specialists plus a test-coverage specialist (active on every PR that touches src/) plus three conditional specialists (auth, doc-writer, performance-expert), all running in their own agent threads, and a CEO synthesizer. The orchestrator is the sole writer to the PR: ONE recommendation comment, no verdict labels, no merge gating. The panel is advisory -- it surfaces findings, prioritizes follow-ups, and renders a ship-recommendation that the maintainer and author weigh. Activate when a non-trivial PR needs a cross-cutting recommendation (architecture, CLI logging, DevX UX, supply-chain security, growth/positioning, optionally auth, docs, perf, and test coverage, with CEO arbitration). |
APM Review Panel - Fan-Out Advisory Review
The panel is FAN-OUT + SYNTHESIZER. Each persona runs in its own agent
thread (via the task tool) and returns JSON matching
assets/panelist-return-schema.json. The orchestrator schema-validates
each return, hands all returns to the apm-ceo synthesizer (also a task
thread, returns JSON matching assets/ceo-return-schema.json), then
renders ONE recommendation comment from assets/recommendation-template.md.
This skill is ADVISORY by design. It does not compute a binary verdict, it
does not apply verdict labels, and it does not gate merge. The panel
surfaces findings; the maintainer and the PR author decide ship.
Architecture invariants
- Advisory regime, not gate regime. There is no
APPROVE / REJECT,
no panel-approved / panel-rejected label, no deterministic verdict
computation. The CEO returns a ship_recommendation.stance (ship_now
/ ship_with_followups / needs_discussion / needs_rework); this is
prose for the human reviewer, never auto-applied as a label or status
check. This is the architectural fix for the previous regime's
over-strictness: removing the binary gate removes the incentive for
panelists to inflate required[] defensively.
- Three severity buckets, none of them gate. Findings carry
severity: blocking | recommended | nit. blocking is the highest
signal a panelist can send and renders prominently in the comment; it
still does not block merge. recommended is the default for substantive
feedback. nit is one-line polish. The orchestrator never reads
severity to gate anything.
- Single-writer interlock. Only the orchestrator writes to the PR:
exactly one
add-comment and one remove-labels call. The
remove-labels call always sweeps panel-review (trigger
idempotency) AND defensively removes panel-approved /
panel-rejected if present (legacy verdict labels from the
pre-advisory regime; they have no meaning here and would mislead
readers if left on a PR after a fresh advisory pass). NO add-labels
call -- there are no verdict labels to apply. Panelist subagents and
the CEO subagent return JSON only and MUST NOT call any gh write
command, post comments, apply labels, or touch the PR state.
- Single-emission discipline. Exactly one comment per panel run,
rendered from
assets/recommendation-template.md after all subagents
return.
- Non-empty turn exit (the run's hard contract). gh-aw decides
success by inspecting
agent_output AFTER your turn ends: a turn that
ends with zero safe outputs (agent_output = {"items":[]}) is detected
as a failure, the safe-output detection job is skipped, the
add-comment job never runs, and the workflow opens a "No Safe Outputs
Generated" issue. Therefore your turn MUST end with at least one safe
output -- the rendered comment on success (step 7), or an explicit
noop if the run genuinely cannot produce one. NEVER end the turn
empty.
- Synchronous fan-out -- never spawn-and-forget. Every
task spawn
(each panelist AND the CEO synthesizer) is BLOCKING: spawn it, WAIT for
its JSON return, then continue. Use the task tool's synchronous mode;
do NOT use its background/detached mode -- the variant that returns an
agent_id immediately and runs the subagent in the background -- for
any panelist or the CEO. Their returns are LOAD-BEARING: the comment
cannot be rendered without them. Spawning the CEO (or a panelist)
detached and then ending the turn while it is still running is the
documented cause of the empty-output failure above.
Agent roster
Topology
apm-review-panel SKILL (orchestrator thread)
|
FAN-OUT via task tool (panelists in parallel)
|
+-----+-------+-------+-----+-----+------+-----------+----------+
v v v v v v v v v (cond.)
py cli dx-ux sec grw auth doc-writer test-cov
| | | | | | | |
| each returns JSON per panelist-return-schema.json
+-----+-------+-------+-----+-----+------+-----------+----------+
|
v <-- S4 schema-validate
v <-- on malformed: re-spawn that persona
v
task: apm-ceo synthesizer
- aggregates findings across panelists
- resolves dissent
- emits headline + arbitration prose + principle alignment
- emits curated recommended_followups (prioritized)
- emits ship_recommendation (stance + prose)
- returns ceo-return-schema.json
|
v <-- S4 schema-validate
v
orchestrator (sole writer)
| |
v v
add-comment remove-labels
(max:2) [panel-review,
panel-approved,
panel-rejected]
(trigger reset +
legacy verdict sweep)
Conditional panelists
Three personas are conditional (auth, doc-writer, performance-expert). A
fourth (test-coverage) is mandatory on every PR that touches src/ and
only skipped on documentation-only PRs -- see its section below for why.
The orchestrator ALWAYS spawns ALL four tasks to keep the schema
return shape uniform; the prompt instructs the subagent to set
active: false with an inactive_reason if the condition does not
hold.
Auth Expert
Activate when the PR changes any of:
src/apm_cli/core/auth.py
src/apm_cli/core/token_manager.py
src/apm_cli/core/azure_cli.py
src/apm_cli/deps/github_downloader.py
src/apm_cli/marketplace/client.py
src/apm_cli/utils/github_host.py
src/apm_cli/install/validation.py
src/apm_cli/install/pipeline.py
src/apm_cli/deps/registry_proxy.py
Fallback self-check (when no fast-path file matched): "Does this PR
change authentication behavior, token management, credential resolution,
host classification used by AuthResolver, git or HTTP authorization
headers, or remote-host fallback semantics? If unsure, answer YES."
Doc Writer
Activate when the PR changes any of:
README.md
CHANGELOG.md
MANIFESTO.md
docs/src/content/docs/**
.apm/skills/**/*.md
.apm/agents/**/*.md
.github/skills/**/*.md
.github/agents/**/*.md
.github/instructions/**/*.md
.github/workflows/*.md (gh-aw natural-language workflows)
packages/apm-guide/**
Fallback self-check (when no fast-path file matched): "Does this PR
change user-facing documentation, agent or skill prose, instruction
files, CHANGELOG entries, README claims, or any natural-language
artifact a reader will rely on? If unsure, answer YES."
When the doc-writer is active and the PR includes documentation changes,
the persona reviews them for: (a) consistency with the existing voice
and structure, (b) accuracy against the code being changed, (c)
completeness for the typical reader (no orphan claims, no missing
prerequisites), (d) discoverability (cross-links, sidebar order if
Starlight content). When the doc-writer is active because of code
changes that SHOULD have updated docs but did not, the persona surfaces
that gap as a finding.
Performance Expert
Activate when the PR changes any of:
src/apm_cli/cache/**
src/apm_cli/deps/**
src/apm_cli/install/phases/**
src/apm_cli/install/pipeline.py
src/apm_cli/install/resolve.py
scripts/perf/**
src/apm_cli/core/command_logger.py (when the diff adds perf-instrumentation logs)
Also activate when the PR description claims a performance win
(speedup ratio, latency reduction, bytes-on-disk reduction, throughput
improvement) or attaches a perf-harness measurement table.
Fallback self-check (when no fast-path file matched): "Does this PR
change the hot path for dependency download, materialization, cache
layout, transport (git protocol, partial clone, sparse checkout),
parallelism, or any user-visible install/update wall-time? If unsure,
answer YES."
When active, the performance-expert reviews against the package-manager
performance playbook: transport minimization (depth, filter, sparse
scope), cache layering and dedup keys, parallelism and lock contention,
working-tree materialization cost, perf-harness methodology (cache
wipe, warm/cold separation, statistical noise), and pervasive
application of the chosen technique across install / update / run
surfaces (not just the one path the PR exercises).
Test Coverage Expert
Active by default on every PR that touches src/**/*.py. The only
condition that flips this persona to active: false is a
documentation-only PR -- the diff contains zero src/**/*.py files.
In that case set inactive_reason: "documentation-only PR -- no runtime code paths to defend".
The activation rule is intentionally narrow: under the advisory regime,
test outcomes are LOAD-BEARING for CEO arbitration (passed / failed /
missing test evidence outranks opinion-only findings -- see
apm-ceo.agent.md and panelist-return-schema.json evidence block).
A persona whose findings carry that weight cannot be silently skipped
on a heuristic. Better to spawn it on a pure refactor and have it
return a single nit-severity "no behavior surface touched -- no
coverage finding" line than to skip it and leave the CEO without
evidence to weigh. (Earlier revisions of this skill paired test-coverage
with auth and doc-writer as conditional for symmetry; that symmetry
broke when test evidence became load-bearing.)
The test-coverage-expert is paired with the devx-ux-expert lens and
defends the user-promise contracts the DevX persona enumerates (CLI
surface, error wording, install idempotency, lockfile determinism, auth
resolution). It MUST verify "no test exists" claims with view/grep
on the test tree before emitting a finding -- false-positive coverage
findings destroy trust in the field. It does NOT compute coverage
percentages, does NOT flag tests for pure refactors, and does NOT
duplicate python-architect on test-code design.
Routing matrix (CEO synthesis emphasis only)
These routes describe WHICH specialist's findings the CEO weights more
heavily for a given PR type. They do NOT change which personas run --
every mandatory persona always runs. Routing is a CEO synthesis hint.
- Architecture-heavy PR -> CEO weights Python Architect on
abstraction calls; CLI Logging on consistency.
- CLI UX PR -> CEO weights DevX UX on command surface; CLI Logging
on output paths; Growth Hacker on first-run conversion.
- Security PR -> CEO biases toward Supply Chain Security on default
behavior; DevX UX flags ergonomics regression from any mitigation.
- Auth PR (auth-expert active) -> CEO weights Auth Expert on
AuthResolver / token precedence; Supply Chain on token-scoping.
- Docs / release / comms PR (doc-writer active) -> CEO weights Doc
Writer on accuracy and voice; Growth Hacker on hook and story angle.
- Behavior-change PR (test-coverage active) -> CEO weights Test
Coverage Expert on regression-trap presence; DevX UX on which user
promises the change touches. A blocking-severity coverage finding on
a critical-promise surface (auth, lockfile, install, marketplace,
hooks) is the highest signal in this routing.
- Full panel (default) -> CEO synthesizes equally; calls out any
dissent in
dissent_notes.
Execution checklist
Work through these steps in order. Do not skip ahead. Do not emit any
output to the PR before step 6. Every task spawn below is BLOCKING:
wait for the subagent to return before continuing, and never end your
turn while a panelist or the CEO synthesizer is still running. The turn
ends only after the comment (step 7) and label sweep (step 8) -- or, if
no comment can be rendered, an explicit noop (step 9) -- are emitted.
-
Read PR context (the orchestrating workflow already fetched it
via gh pr view / gh pr diff). Identify changed files for the
conditional panelist routing decisions (auth-expert and doc-writer).
-
Resolve the conditional panelists using the rules above. Decide
for EACH conditional persona: spawn active OR spawn with
active: false + an inactive_reason. Either way, all three
conditional personas ARE spawned -- the schema requires uniform
return shape.
-
Fan out panelist tasks. Spawn the following tasks in PARALLEL
via the task tool, one task per persona:
python-architect (also asked to supply extras.diagrams:
class_diagram (mermaid classDiagram), component (mermaid
flowchart TD), and OPTIONAL sequence (mermaid
sequenceDiagram) blocks per the persona's section 1/2/3 contract)
cli-logging-expert
devx-ux-expert
supply-chain-security-expert
oss-growth-hacker
auth-expert (always - active per step 2)
doc-writer (always - active per step 2)
test-coverage-expert (always - active per step 2)
performance-expert (always - active per step 2)
Each task prompt MUST:
- Reference its persona file by relative path so the subagent loads
its own scope, lens, and anti-patterns.
- Include the PR number, title, body, and diff (passed inline).
- Cite
assets/panelist-return-schema.json and require the subagent
to emit JSON matching that schema as its FINAL message.
- State the calibrated severity contract: "Use
severity: blocking
ONLY for correctness regressions, security/auth bypasses, or
architectural faults that compound, with explicit rationale.
Default substantive feedback to recommended. Use nit for
one-line polish. The panel is advisory; nothing you return blocks
merge -- pick the severity that honestly matches your signal
strength."
- Restate the output contract: NO
gh write commands, NO posting
comments, NO label changes, NO touching PR state. JSON return only.
-
S4 schema gate. When each panelist task returns, parse the JSON
and validate against assets/panelist-return-schema.json. On
validation failure:
- Re-spawn that ONE panelist with an explicit error message pointing
at the violated rule.
- Maximum two re-spawn attempts per panelist. If still malformed,
synthesize a placeholder
{persona: "<slug>", active: true, summary: "Schema failure -- see extras.", findings: [], extras: {schema_failure: "<reason>"}}
and surface the failure in the CEO arbitration prompt.
-
Spawn the CEO synthesizer task. Pass the full set of validated
panelist JSON returns to a task invocation that loads
../../agents/apm-ceo.agent.md. Run it as a BLOCKING task and WAIT
for its JSON return -- do NOT spawn it detached (background mode that
returns an agent_id) and do NOT end your turn while it runs. Its
return is required to render the comment; ending the turn here is the
exact cause of the "No Safe Outputs Generated" failure. The prompt
MUST:
- Provide all panelist returns as structured input.
- Ask for: headline, arbitration prose, principle alignment (only
applicable principles), curated recommended_followups (prioritized
by signal, NOT a re-listing of every finding), ship_recommendation
(stance + prose).
- Cite
assets/ceo-return-schema.json and require JSON return.
- Restate the contract: the panel is advisory. The CEO does NOT pick
a verdict label. The
ship_recommendation.stance is prose for the
human reviewer, not a gate. NO gh write commands.
Validate the CEO return against assets/ceo-return-schema.json. On
failure, re-spawn once with the violation cited.
-
Resolve the notification audience. The advisory comment must
surface in the inboxes of the people who will act on it. Run:
gh pr view <PR_NUMBER> --json author,reviewRequests
Build notify_audience as the deduplicated list:
- the PR author's
@login (always included);
- every requested reviewer's
@login (these are the
CODEOWNERS-resolved reviewers GitHub auto-requested for the
touched paths, plus any explicitly-requested human reviewers);
- every requested team's
@org/team-slug (CODEOWNERS team
entries).
Filter out:
- bot logins (login ending in
[bot] or matching
dependabot|github-actions|copilot-pull-request-reviewer);
- the orchestrator's own identity (avoid self-ping).
Cap the final list at 6 handles to avoid notification noise (PR
author + up to 5 reviewers/teams). If the cap trims, prefer team
handles over individual logins. Pass the resulting list to the
template renderer as notify_audience.
This step replaces the maintainer-notification signal that the
pre-advisory verdict labels carried. It is the only mechanism by
which a fresh panel pass announces itself.
-
Render the comment. Load assets/recommendation-template.md,
fill the placeholders from the panelist + CEO JSON, and emit it as
exactly ONE comment.
Filling rules:
- The per-persona summary table renders ONLY active panelists, one
row per persona, with finding counts by severity and the persona's
summary field.
- The mermaid diagrams come from
python-architect.extras.diagrams.
If absent, render the placeholder lines from the template (do NOT
invent diagrams).
- The recommended follow-ups list renders the CEO's curated subset,
not every finding. Full per-persona findings collapse at the bottom.
- NEVER render the words "Verdict", "APPROVE", "REJECT", "blocked",
"merge gate", or any equivalent. The panel is advisory.
-
Sweep labels via safe-outputs.remove-labels. The list MUST be
[panel-review, panel-approved, panel-rejected] -- always all three,
regardless of which are currently on the PR. panel-review is the
re-run idempotency reset; the other two are LEGACY VERDICT LABELS
from the pre-advisory regime that have no meaning under the advisory
contract and would mislead readers if left on a freshly-reviewed PR.
safe-outputs.remove-labels is idempotent on missing labels, so
sweeping all three on every run is safe and self-healing. NO
verdict labels are applied.
-
Guarantee a non-empty exit. Your final action this turn MUST be a
safe output. In the normal path that is the single add-comment from
step 7 (the remove-labels sweep alone does NOT count -- it is not
the run's required output). Before ending the turn, confirm step 7
actually issued the add-comment call and it did not error. If, after
every subagent has returned, you genuinely cannot render a comment
(e.g. a fatal upstream error), call noop so the run records an
intentional no-action rather than an empty agent_output. Ending the
turn with zero safe outputs is a FAILURE, not a success -- see the
"Non-empty turn exit" architecture invariant.
Output contract (non-negotiable)
- Exactly ONE comment per panel run, rendered from
assets/recommendation-template.md. The safe-outputs.add-comment.max: 2 is a fail-soft ceiling; the discipline lives here.
- Exactly ONE
remove-labels call sweeping
[panel-review, panel-approved, panel-rejected].
- NO
add-labels call. The advisory regime has no verdict to encode.
- Subagents (panelists + CEO) NEVER write to PR state, NEVER call
gh pr comment, NEVER call gh pr edit --add-label. They return JSON.
The orchestrator is the sole writer.
- Never invent new top-level template sections or drop existing ones.
Gotchas
- Roster invariant. The frontmatter description, the roster table,
the conditional rules, the recommendation template, and the JSON
schema MUST agree on the persona set. If you change one, change all
in the same edit.
- Calibrated severity discipline. The advisory regime relies on
panelists honestly distinguishing
blocking from recommended. If a
panelist marks everything blocking, the comment becomes noisy and
the maintainer learns to ignore the field. The panelist prompts state
the contract explicitly; the CEO arbitration prose is the safety
valve when a panelist over-flags.
- Mermaid diagrams are template-required. The python-architect
persona is asked to supply
extras.diagrams.class_diagram,
extras.diagrams.component, and the OPTIONAL
extras.diagrams.sequence. The template renders nothing when they
are missing -- it does NOT invent diagrams. Real diagrams are
what makes the comment scannable for the human reviewer.
- Mermaid
classDiagram :::cssClass shorthand gotcha. GitHub's
mermaid renderer rejects :::cssClass appended to relationship
lines (e.g. A *-- B:::touched); use standalone
class Name:::cssClass declarations instead. Authority:
python-architect.agent.md:146-154.
- Doc-writer detects DRIFT, not just edits. When the PR changes
user-facing code that SHOULD have updated docs but did not, doc-writer
surfaces that as a finding. The conditional rule above is necessary
but not sufficient -- doc-writer reasons about doc consistency given
the diff, not just whether doc files were touched.
- False-negative auth gotcha. Auth regressions can be introduced
from non-auth files that change the inputs to auth -- host
classification, dependency parsing, clone URL construction, HTTP
authorization headers, or call sites that bypass
AuthResolver. If
a diff changes how a remote host, org, token source, or fallback path
is selected and you are not certain it is auth-neutral, activate
auth-expert as active: true.
- Test-coverage probe is mandatory. The test-coverage-expert MUST
verify "no test exists for X" via
view/grep on the tests/ tree
before emitting a finding. A false-positive coverage finding (test
exists but persona claimed it does not) destroys maintainer trust in
the field. The persona scope file enforces this; the orchestrator
passes the diff and trusts the persona to probe.
- Subagent write enforcement is contract-based, not sandbox-based.
Tool permissions are workflow-scoped, not subagent-scoped, so every
spawned task technically inherits the same
gh toolset. The
"subagents must not write" rule is enforced by the prompt contract in
each .agent.md plus the safe-outputs.add-comment.max: 2
fail-soft. If a subagent ever tries to post a comment, the cap
catches it.
- Empty-safe-output failure (background spawn-and-forget). The single
most common way this panel "succeeds" yet posts nothing is spawning the
CEO synthesizer (or a panelist) as a background/detached task and then
ending the turn while it is still running. The harness exits with
agent_output = {"items":[]}, gh-aw skips safe-output detection, the
add-comment job never runs, and the workflow opens a "No Safe Outputs
Generated" issue. Every task spawn MUST be awaited to completion, and
the turn MUST end with a safe output -- the comment, or an explicit
noop. See the "Synchronous fan-out" and "Non-empty turn exit"
architecture invariants and step 9.
- No verdict-label reset workflow. The previous regime had a
companion workflow
pr-panel-label-reset.yml that stripped verdict
labels on every push. The advisory regime has no verdict labels to
strip; that workflow is removed.