feat(autopilot): maintainer canvas, panel-mode, and comment writer

.agents/skills/autopilot-comment/SKILL.md

@@ -0,0 +1,112 @@
+---
+name: autopilot-comment
+activation_card: on
+description: >-
+  Use this skill to post or patch ONE GitHub comment for a microsoft/apm
+  autopilot run. It owns public vs debug body assembly and the AI
+  disclaimer footer. Never labels, assigns, requests reviewers, or
+  merges. Autopilot skills that write comments must load this skill
+  and must not call gh comment APIs themselves.
+---
+
+# autopilot-comment -- Sole GitHub comment writer
+
+Any autopilot skill that would write a GitHub comment loads this
+skill in the same session and stops if the card is missing. Do not
+spawn a nested session. Do not comment from the caller.
+
+## Activation card
+
+`activation_card: on`. Before any GitHub comment write, emit this
+Enter card with every field filled. Missing field -> stop.
+
+```text
+skill: autopilot-comment
+skill_path: <resolved directory of this SKILL.md>
+mode: run
+subject: microsoft/apm#<number>
+path: comment
+intent: post one GitHub comment | patch one GitHub comment
+origin: unattended | actor-session
+write: on | off
+debug: off | on
+repo: microsoft/apm
+target_kind: issue | pull
+target: <positive integer>
+source_skill: <canonical autopilot skill name>
+comment_id: none | <positive integer>
+invocation: agentic-workflow | actor-session
+receipt: <HTML comment watermark line>
+public_body: <author-facing prose, no JSON, no footer>
+debug_body: <full internal template without JSON fence, or none>
+```
+
+Rules:
+
+- `write` defaults to `on` when the caller omitted it.
+- `write: off` returns the assembled body only. Do not post or
+  patch.
+- `write: on` posts or patches exactly one comment. Never labels,
+  assigns, requests reviewers, or merges.
+- `debug` defaults to `off` when omitted or unknown. Omitted
+  `debug` is not a missing-field stop.
+- `debug: off` -> HTML `receipt` plus `public_body`. Do not post
+  internal headings from `debug_body`.
+- `debug: on` -> first line `[i] Skill debug is on.` then `receipt`
+  plus `debug_body` (or `public_body` if `debug_body` is `none`).
+- `comment_id: none` with intent `post`. A numeric `comment_id`
+  with intent `patch`. Mismatch -> stop.
+- `source_skill` is the caller (for example
+  `autopilot-issue-triage-worker`). Never `autopilot-comment`.
+- `origin` fail-closed unknown -> `unattended`.
+- Strip any JSON fence from bodies before assembly. Never post
+  JSON on GitHub.
+
+After the write (or dry-run), emit this Exit receipt:
+
+```text
+skill: autopilot-comment
+subject: microsoft/apm#<number>
+path: comment
+write: on | off
+debug: off | on
+posted: yes | no
+patched: yes | no
+comment_id: <id or none>
+source_skill: <canonical autopilot skill name>
+approved: n/a
+```
+
+## Assemble
+
+1. Drop JSON fences and `triage-recommendation` blocks.
+2. Build the body from `debug` as above. ASCII only.
+3. Append this exact footer, substituting `source_skill`:
+
+```text
+---
+
+Generated by <source_skill>. This comment is AI-generated and may contain errors.
+```
+
+4. If the footer is already the last lines, do not duplicate it.
+
+## Write
+
+Actor-session / Copilot App / Cloud / Remote:
+
+- issue + post: `gh issue comment <target> --body-file ...`
+- pull + post: `gh pr comment <target> --body-file ...`
+- patch: `gh api -X PATCH /repos/microsoft/apm/issues/comments/<comment_id>`
+
+Agentic Workflow: `safe-outputs.add-comment` for post. Patch is
+not available; stop if intent is `patch`.
+
+One comment per activation. No progress comments.
+
+## Hard nos
+
+- Do not assign, request reviewers, merge, or edit labels.
+- Do not implement.
+- Do not compose schedulers or other workers.
+- ASCII only.

.agents/skills/autopilot-comment/apm.yml

@@ -0,0 +1,10 @@
+name: autopilot-comment
+version: 0.1.0
+description: Sole GitHub comment writer for microsoft/apm autopilot skills. Assembles public vs debug body and appends the AI disclaimer.
+author: Microsoft
+license: MIT
+type: skill
+
+dependencies:
+  apm: []
+  mcp: []

.agents/skills/autopilot-issue-triage-scheduler/SKILL.md

@@ -163,7 +163,11 @@ selected list; when a slot returns, fill it with the next item.
    else run the worker in this thread for that one issue, then
    the next. When a slot returns, dispatch the next queued
    issue. Do not stop because the pool was full. Each slot reads
-   the complete comment history before advising.
+   the complete comment history before advising. Pass `debug: off`
+   unless this run's caller set `debug: on`. Pass
+   `comment_via: autopilot-comment`. Worker kickoff must load
+   `autopilot-comment` before any GitHub comment. Workers must
+   not call `gh issue comment`.
 4. Optionally present ONE consolidated triage digest. Wait if the
    caller asked for a checkpoint. Never treat silence as accept.
 5. Print a final report from the table.

.agents/skills/autopilot-issue-triage-scheduler/assets/fan-out-pool.md

@@ -19,6 +19,9 @@ is empty.
 1. Prefer a new session (Copilot App / Cloud / Remote) named
    `Issue triage #<issue-number>` whose kickoff runs
    `autopilot-issue-triage-worker` on exactly one issue.
+   Kickoff must also load `autopilot-comment` (`comment_via:
+   autopilot-comment`, `debug: off` unless the caller set on).
+   The worker must not call `gh issue comment`.
 2. Else spawn a sub-agent (`task`) with the worker skill.
 3. Else run the worker sequentially in this session.
 

.agents/skills/autopilot-issue-triage-scheduler/scripts/fetch_queue.py

@@ -119,6 +119,9 @@ def is_bot_author(author: str, author_type: str) -> bool:
 def skip_reason(record: dict[str, Any], mode: str) -> str | None:
     """Return a skip token, or None when state/body filters pass.
 
+    Bot-authored issues skip. Bot-authored PRs stay eligible so they
+    can be triaged, accepted, reviewed, and merged.
+
     Completed-advice markers stay eligible; `triage_state.plan_batch` owns
     that sweep skip plus the per-author quota.
     """
@@ -128,7 +131,9 @@ def skip_reason(record: dict[str, Any], mode: str) -> str | None:
         return "merged"
     if record.get("locked"):
         return "locked"
-    if is_bot_author(str(record["author"]), str(record.get("author_type", "User"))):
+    if record.get("kind") != "pr" and is_bot_author(
+        str(record["author"]), str(record.get("author_type", "User"))
+    ):
         return "bot-authored"
     body = record.get("body")
     if not isinstance(body, str) or is_empty_body(body):

.agents/skills/autopilot-issue-triage-worker/SKILL.md

@@ -48,6 +48,7 @@ intent: advise one already-selected issue
 origin: unattended | actor-session
 write: on | off
 json: off | on
+debug: off | on
 repo: microsoft/apm
 issue: <positive integer>
 invocation: agentic-workflow | actor-session
@@ -58,13 +59,24 @@ Rules:
 - `write` defaults to `on` when the caller omitted it.
 - `write: off` returns the filled template only. Do not comment,
   add labels, or remove labels.
-- `write: on` posts the one advisory comment and processing /
-  classification labels. Never human decision labels. Never assign.
+- `write: on` posts one GitHub comment via
+  `autopilot-comment` and processing / classification labels.
+  Never human decision labels. Never assign. Do not call
+  `gh issue comment` yourself.
 - `json` defaults to `off` when omitted or unknown. Omitted `json`
   is not a missing-field stop.
 - `json: off` -> no `triage-recommendation` JSON receipt.
 - `json: on` -> fill that JSON only as an internal payload
   (session file or parent Exit). Never post it on GitHub.
+- `debug` defaults to `off` when omitted or unknown. Omitted `debug`
+  is not a missing-field stop.
+- `debug: off` -> pass `public_body` as the Suggested issue
+  comment (unwrap the markdown fence) to `autopilot-comment`.
+  Do not pass recommendation, classification, brief, next
+  action, or persona details as `public_body`.
+- `debug: on` -> pass `debug_body` as the filled template
+  without the JSON fence. `autopilot-comment` prefixes
+  `[i] Skill debug is on.`
 - `origin` fail-closed unknown -> `unattended`.
 - One issue. Do not nest a scheduler path.
 
@@ -76,17 +88,19 @@ subject: microsoft/apm#<issue-number>
 path: triage
 write: on | off
 json: off | on
+debug: off | on
 posted: yes | no
 labels_applied: <comma list or none>
 approved: n/a
 ```
 
 The panel is fixed at **2 mandatory specialist lenses + up to 3
 conditional lenses + 1 always-active arbiter = 6 persona sections in
-one triage comment**. You play each
+the filled template**. You play each
 lens in turn from inside a single agent loop (progressive-disclosure
 skill model -- no sub-agent dispatch). Routing chooses *which* lenses
-execute; it never changes which headings appear in the final comment.
+execute; it never changes which headings appear in the filled
+template. The public GitHub body still follows `debug`.
 
 This skill mirrors the `autopilot-pr-review-worker` orchestration shape on
 purpose. Same single-comment discipline, same completeness gate, same
@@ -374,26 +388,36 @@ write/receipt step.
 7. Now (and only now) load `assets/triage-template.md` and fill it
    in with the collected findings, recommendation, classification,
    proposed brief, and suggested comment body.
-8. Verify the rendered comment contains every top-level heading from
-   the template, all six persona `<details>` sections, and the closing
-   `triage-recommendation` JSON block. If any element is missing, re-render
-   from the template instead of posting a hand-composed substitute.
+8. Verify the INTERNAL fill contains every top-level heading from
+   the template, all six persona `<details>` sections, and (only
+   when `json: on`) the closing `triage-recommendation` JSON.
+   If any required element is missing, re-render from the
+   template. That filled template is not the GitHub body when
+   `debug: off`.
 9. If `write: on`, apply the advisory writes yourself. The scheduler
    never comments or labels. A worker summoned without a scheduler
    still writes when `write` is on. No-op when target plus
-   conversation watermark already match. Post exactly one template
-   comment (`gh issue comment` in actor-session / Copilot App /
-   Cloud / Remote; Agentic Workflow `safe-outputs.add-comment`).
-   Add `triage/recommended` plus useful classification from the
-   contract allowlist (`gh issue edit --add-label` or
-   `safe-outputs.add-labels`). Never write human decision labels.
-   Remove only `triage/requested` after successful advice. If
-   `write: off`, return the filled template only. Also return the
-   `triage-recommendation` JSON to the caller. Emit the Exit
-   receipt (`posted: yes` only when a comment was written). Never
-   authorize implementation. This is the ONLY triage-comment
-   emission for the entire panel run -- no per-persona comments,
-   no progress comments.
+   conversation watermark already match. Probe
+   `$REPO_ROOT/.agents/skills/autopilot-comment/SKILL.md` (or the
+   sibling package). Missing -> stop. Do not post. Post exactly
+   one comment by activating `autopilot-comment` (`source_skill:
+   autopilot-issue-triage-worker`). Do not call `gh issue comment`
+   or `safe-outputs.add-comment` yourself. Missing footer after
+   that write is a failed post -- patch via `autopilot-comment`
+   (`intent: patch`) rather than leaving the comment bare.
+   `debug: off`: Suggested issue comment body only as
+   `public_body`. `debug: on`: filled template as `debug_body`.
+   Add `triage/recommended` plus useful
+   classification from the contract allowlist
+   (`gh issue edit --add-label` or `safe-outputs.add-labels`).
+   Never write human decision labels. Remove only
+   `triage/requested` after successful advice. If `write: off`,
+   return the filled template only. Also return the
+   `triage-recommendation` JSON to the caller when `json: on`.
+   Emit the Exit receipt (`posted: yes` only when a comment was
+   written). Never authorize implementation. This is the ONLY
+   triage-comment emission for the entire panel run -- no
+   per-persona comments, no progress comments.
 
 ### Persona pass procedure
 
@@ -419,15 +443,21 @@ This contract is non-negotiable -- it is the difference between a
 triage that lands as one cohesive comment and one that fragments into
 per-persona noise.
 
-- Produce **exactly one** comment per triage run.
-- Use `assets/triage-template.md` as the comment body. Keep its
+- Produce **exactly one** GitHub comment per triage run.
+- Always fill `assets/triage-template.md` internally. Keep its
   section headings exactly as written. Adapt the body of each
   section to the issue. Do not invent new top-level sections or drop
   existing ones.
-- The GitHub issue comment is human prose only: HTML receipt,
-  headings, classification, brief, next action, suggested reply,
-  and persona details. Do not post JSON, machine fences, or
-  `comment_markdown` dumps on the issue.
+- Load `autopilot-comment` before any GitHub comment write.
+  Fill its card (`source_skill: autopilot-issue-triage-worker`,
+  `debug` copied from this card, `receipt` = the HTML watermark
+  line). Missing skill or missing card -> stop. Never post
+  with `gh issue comment`.
+- `debug: off` `public_body`: Suggested issue comment prose
+  (no wrapping fence, no other headings).
+- `debug: on` `debug_body`: the filled template without JSON.
+- Do not post JSON, machine fences, or `comment_markdown` dumps
+  on the issue. `autopilot-comment` appends the AI disclaimer.
 - Emit the template's `triage-recommendation` JSON only when
   `json: on`. Keep `schema_version: 2` and `advisory_only: true`.
   Never attach that JSON to the GitHub comment. Consumers must

.agents/skills/autopilot-issue-triage-worker/apm.yml

@@ -10,5 +10,6 @@ type: skill
 # against the consumer repo's installed .apm/agents/** or
 # .github/agents/** -- identical to the autopilot-pr-review-worker precedent.
 dependencies:
-  apm: []
+  apm:
+    - ../autopilot-comment
   mcp: []

.agents/skills/autopilot-issue-triage-worker/assets/triage-template.md

@@ -5,9 +5,13 @@ conditional lenses say "Not activated -- <reason>". The caller is the
 sole writer; direct invocation returns this without posting.
 Fill the v2 JSON block only when the activation card has `json: on`.
 Default is `json: off`: omit the JSON receipt. Never post that JSON
-(or any machine fence) on the GitHub issue. The public comment is
-prose plus the HTML receipt. The JSON is advisory data, not a
-label/milestone write instruction.
+(or any machine fence) on the GitHub issue. Default `debug: off`
+posts only the HTML receipt plus the Suggested issue comment body.
+`debug: on` posts this filled template prefixed with
+`[i] Skill debug is on.` Both paths go through
+`autopilot-comment`, which appends the AI disclaimer.
+Never post with `gh issue comment`. The JSON is advisory
+data, not a label/milestone write instruction.
 -->
 
 <!-- apm-triage-advisory:v2 target=issue#<N> watermark=<latest-comment-id-or-updated-at> -->