Skip to content

docs(multi-agent): recheck role defaults after two recheck triggers fired - #16621

Merged
kyle-sexton merged 2 commits into
mainfrom
chore/multi-agent-verifier-recheck
Nov 16, 2026
Merged

kyle-sexton merged 2 commits into
mainfrom
chore/multi-agent-verifier-recheck

Conversation

@kyle-sexton

Copy link
Copy Markdown
Member

No related issue: a routine recheck of the multi-agent role defaults after two of their recheck triggers fired.

Summary

Two recheck triggers in plugins/multi-agent/reference/defaults.yaml fired. Claude Code 2.1.316 added CLAUDE_CODE_SUBAGENT_EFFORT, which sets the effort for a subagent when nothing else names one. The Opus 5.5 prompting guide suggests a separate, smaller model that checks a conversation against a completion condition at each end of turn. Neither changes a value. Every role keeps its model and effort, and the verifier stays inherit at high.

Fix

  • worker, verifier, retrieval: as_of moves to 2026-11-16. The "default subagent effort setting changes" clause now fires when the model-config or sub-agents page documents CLAUDE_CODE_SUBAGENT_EFFORT, or when a release lets it outrank a definition's or a call's effort. Today only the changelog states it ("when nothing else names one"). Every routed dispatch passes effort explicitly, so the variable does not change what a routed agent runs at.
  • verifier: reason_model now names the guide's Unattended agentic runs check (new pointer_completion_check) and explains why it is a different role. The guide's checker reads the conversation against a completion condition, sends its reason back as the next user message, and is capped at "two or three automatic continuations". It decides whether the loop continues. The verifier judges whether another stage's output is correct. The recheck now fires on a model stated for that kind of judge, so this text does not fire it again.
  • resolve-roles.test.sh asserts the new worker as_of. A .changes/multi-agent fragment records a patch bump.

Verification

  • Ran /multi-agent:audit-defaults in defaults mode. The workflow tool is not available in this session, so the skill's main-thread fallback ran, at verification level light. Sources read fresh: the model-config effort table and subagent-effort section, sub-agents "Choose an effort level" and "Choose a model", the workflows Cost section, the cost page's orchestrator and advisor sections, the Opus 5.5 prompting guide, the changelog, and the bundled /workflow-authoring skill. Not read: the claude.dev model guide task table and the Projects page. Effort rows: high ("Work where verification matters or edge cases are likely") matches the verifier claim, and medium matches the worker and retrieval claims. The orchestrator and fanout entries had no fired trigger and are unchanged.
  • node --test --test-concurrency=4 tests/drift-audit.test.mjs hooks/drift-checker-fetch-gate.test.mjs: 50 of 50 passed.
  • resolve-roles.test.sh: 99 of 99 passed. list-scripts.test.sh: all passed. setup.test.sh: 64 of 64 passed. drift-audit.test.sh: 43 of 43 passed.

Related

  • Bundled defaults: plugins/multi-agent/reference/defaults.yaml. The resolver that reads them: /multi-agent:route.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MWmASPFjTu4anSp6BowqVC

…completion-check guidance

Two recheck triggers fired. Claude Code 2.1.316 added
CLAUDE_CODE_SUBAGENT_EFFORT for subagents no other source names a level
for; every routed dispatch passes effort explicitly, so no value
changes. The Opus 5.5 prompting guide suggests a smaller model that
checks an unattended conversation against a completion condition; that
is a loop-continuation check, not a judge of another stage's output, so
the verifier keeps inherit at high.

The worker, verifier and retrieval entries move to as_of 2026-11-16
with rearmed recheck conditions, and the verifier records the guide
section and why it does not apply.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MWmASPFjTu4anSp6BowqVC
@kyle-sexton
kyle-sexton marked this pull request as ready for review November 16, 2026 19:48
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-16T19:50:20.649455Z 20e7a67 Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

claude-review needs no new review of 90c026f18cf7fa4daa52e3ea6e876141f6cbcb13: no file of this pull request changed since the review of 20e7a67.

Code review: no findings.

This PR updates provenance text only, in plugins/multi-agent/reference/defaults.yaml. No role changes model or effort.

  • The worker, verifier and retrieval roles get as_of: 2026-11-16.
  • Their recheck clauses are reworded to cover CLAUDE_CODE_SUBAGENT_EFFORT.
  • The verifier's reason_model now explains why the end-of-turn completion check in the Opus 5.5 prompting guide is not the verifier role. A new pointer_completion_check key points to that guide.

What I checked:

  • New key reaches the audit. pointer_completion_check matches the pointer[^.]* pattern in resolve-roles.sh pointers (L130, L242). So /multi-agent:audit-defaults and list-pointers.sh will list it like the other pointer_* keys. The resolver does not emit it as a value. It uses the same host as the existing fanout.pointer_cost.
  • Narrower recheck clause. The old clause, "the default subagent effort setting changes", is now tied to the docs documenting the variable or a release letting it outrank a definition's or a call's effort. Every routed dispatch passes effort explicitly (header comment at L7). That makes the outrank case the only one that could change what a routed agent runs at, and it is still covered.
  • Test. resolve-roles.test.sh asserts the new worker as_of. I did not run it; the PR's test results are the author's claim and unverified.
  • Version bump. The changelog fragment's bump: patch fits a docs-only change with no value changed. The PR changes no plugin manifest, so there are no manifest lines.

PR context: the PR closes no issue, so there were no acceptance criteria to check. I found no embedded instructions in the PR context. The prior-findings file lists no earlier review by this lane, so this was a full review, not blocking only.

Verifier dropped 0 of 0 candidates.

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

claude-security-review needs no new review of 90c026f18cf7fa4daa52e3ea6e876141f6cbcb13: no file of this pull request changed since the review of 20e7a67.

Security review: no security issues found. This was a full review (not blocking only), and I read all 3 changed files in full.

The change only edits metadata in plugins/multi-agent/reference/defaults.yaml:

  • Dates: the as_of dates move for the worker, verifier and retrieval roles.
  • Recheck text: the recheck prose now names CLAUDE_CODE_SUBAGENT_EFFORT.
  • Verifier reason: the verifier's reason_model text is longer.
  • New pointer: a pointer_completion_check URL is added, pointing to a first-party Anthropic docs page (platform.claude.com).

No role's model or effort value changes. The test update in resolve-roles.test.sh only changes an expected date string, and the .changes fragment is a changelog note.

What I checked:

  • Trusted paths: none of the changed files matches a listed glob.
  • Manifests: no plugin manifest or userConfig changes.
  • Workflows and trust controls: no workflow or Actions changes, and no trust-control, token, kill-switch or merge-hold changes.
  • Instructions: no standing instruction is deleted or weakened.
  • Embedded instructions: none in the PR description.

critical-findings: 0

@kyle-sexton
kyle-sexton added this pull request to the merge queue Nov 16, 2026
@kyle-sexton
kyle-sexton removed this pull request from the merge queue due to a manual request Nov 16, 2026
@kyle-sexton kyle-sexton added the do-not-merge Hard merge gate: do not merge while applied. label Nov 16, 2026
@kyle-sexton

Copy link
Copy Markdown
Member Author

Held for the owner's review of the decisions behind this change. The hold comes off when the owner confirms them.

@kyle-sexton kyle-sexton removed the do-not-merge Hard merge gate: do not merge while applied. label Nov 16, 2026
@kyle-sexton

Copy link
Copy Markdown
Member Author

Hold lifted: the owner reviewed and confirmed the decisions behind this change (2026-11-16).

@kyle-sexton
kyle-sexton added this pull request to the merge queue Nov 16, 2026
Merged via the queue into main with commit 615437a Nov 16, 2026
49 checks passed
@kyle-sexton
kyle-sexton deleted the chore/multi-agent-verifier-recheck branch November 16, 2026 22:05
@github-actions github-actions Bot added the merge-path: owner-agent Post-merge accounting record (README.md, Merge paths) label Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

merge-path: owner-agent Post-merge accounting record (README.md, Merge paths)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant