[kyle-sexton]

Refs: #6916

Summary

The plugins/discovery findings from wave E of the repo-mode /multi-agent:audit-defaults sweep. Six claims had drifted from the upstream pages they rest on, or restated upstream facts with no pointer.

Fix

• reference/parent-contract.md: the deny-rule pointer now goes to sub-agents#disable-specific-subagents, and the type-list pointer stays on #restrict-which-subagents-can-be-spawned. The record is re-dated. The researcher's high pin now says it is not absolute and points at sub-agents#choose-an-effort-level for what overrides it.
• agents/explorer.md: the agent is told not to assume a scout runs on its own model, and a pointer to sub-agents#built-in-subagents says which model built-in Explore uses.
• skills/research/context/deep-tier.md: the 1-16 maxConcurrent clamp is labeled as the workflow script's own cap (workflows/research-sweep.js:85-87), with a pointer to workflows#behavior-and-limits for the runtime limit.
• skills/setup/SKILL.md (two places): "env vars are read at session start" now applies to shell variables only. Whether a settings-file env value reaches a running session is left to env-vars#set-variables-in-settings-files, after review (ce1707a307).

Not in this PR: the unset fork-mode default in setup (SKILL.md:63) is stale, but correcting it changes what setup reports. It needs eval evidence first.

Verification

• Each quoted upstream line was re-read raw (fetch-docs.sh) on 2026-11-01 and still holds.
• scripts/check-affected.sh --tests on the committed head: 14 passed, 2 skipped.
• No dispatch or report changes, so no eval ran.

Remaining on #6916

This pull request delivers one plugin's share of wave 5. Still open on the issue:

• 44 files not yet swept: plugins/harness-config (8), plugins/plugin-quality (4), plugins/playbooks (30) and plugins/mutation-testing (2). They wait until the concise rewrites in flight in those plugins merge, so the sweep reads the new lines.
• Wave 5's other pull requests: docs(discovery): point drifted upstream claims at their sections #11771, docs: point repository docs and lane prompts at current upstream sections #11772, docs(docs-hygiene): correct audit-derivability's subagent-loading claims #11774, docs(review): narrow the opus pin rationale and add upstream pointers #11775, docs(knowledge): narrow drifted upstream claims and drop a model pin #11776, docs(autonomy): point the routines terms and stop gate at upstream sections #11777, docs(harness-ops): date and point three drifted upstream claims #11783, docs(songwriting): qualify object-writing's model and cost claims #11784, docs(code-tidying): label batch-simplify routing fallbacks as local choices #11785, docs(instruction-placement): align MS-13 cells with the re-pinned quote #11786, docs(session-flow): point depth-limit and task-list claims upstream #11790.
• Edits that would change what a skill does, held for eval evidence and to be filed as one slice under docs(multi-agent,session-flow): fix repo-mode audit pilot findings and schedule a full sweep #6916: discovery:setup's fork-default report, docs-hygiene:audit-derivability's spot-test agent, a write-for-agents cut, and the two planning:draft-goal-condition cuts.
• The final sweep count comment on docs(multi-agent,session-flow): fix repo-mode audit pilot findings and schedule a full sweep #6916.

Related

• Refs docs(multi-agent,session-flow): fix repo-mode audit pilot findings and schedule a full sweep #6916: the sweep issue, with one PR per plugin.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T01:01:16.199722Z	7330b1f	PR opened

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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7330b1f1f1

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

claude-security-review reviewed ce1707a3074eda2fd1959b9bca78fbd08af6369f: read 3 of 3 changed file(s); 0 finding(s), 0 as inline comments.

Security review: incremental (7330b1f → ce1707a)

I found no security issues in the incremental change.

I read every hunk of .claude-lane/incremental.diff in all three files in scope, and checked them against the PR context.

• .changes/discovery/fix-6916-w5-upstream-pointers-3a7d91c2.md: only the changelog wording changed.
• plugins/discovery/agents/explorer.md: the line "A scout does not run on this agent's model" now reads "Do not assume a scout runs on this agent's model", and the pointer to the upstream built-in subagents section stays. Under the instruction-surface lens, this tells the agent less about which model a scout runs on. It doesn't remove a rule about tools, permissions, posting, credentials or what untrusted text may reach. Nothing in a protected class is lost.
• plugins/discovery/skills/setup/SKILL.md: two passages now point to the upstream env-vars section for whether a settings-file env value reaches a running session, instead of saying it can. setup is a read-only report ("Do not write anything"), and this change leaves its gates, allow rules and tool grants as they were.

Nothing in scope is a trust control, merge hold, plugin manifest, workflow or token request. None of the changed files matches a listed trusted path. The PR context contains no embedded instructions to the lane. I didn't read docs/contracts/instruction-exception-register/README.md; I judged the softened rule by what breaks without it. Nothing in the change spends credits or can bill.

critical-findings: 0

[github-actions]

claude-review reviewed ce1707a3074eda2fd1959b9bca78fbd08af6369f: read 3 of 3 changed file(s); 0 finding(s), 0 as inline comments.

Incremental review (7330b1f → ce1707a): I read every hunk of .claude-lane/incremental.diff in all 3 files in scope. No findings.

What changed since the last review:

• plugins/discovery/agents/explorer.md: "A scout does not run on this agent's model" is now "Do not assume a scout runs on this agent's model". The line still points at sub-agents#built-in-subagents for the Explore model. The page now points to upstream for that fact and doesn't state it itself, which is what docs(multi-agent,session-flow): fix repo-mode audit pilot findings and schedule a full sweep #6916 asks for.
• plugins/discovery/skills/setup/SKILL.md (two places): the claim "a settings-file env value can reach a running session, with exceptions" is gone. Both places now leave that question to env-vars#set-variables-in-settings-files. The shell-variable statement (read at session start, so it takes effect next session) is still there, and the as-of date and recheck trigger are kept.
• .changes/discovery/fix-6916-w5-upstream-pointers-3a7d91c2.md: the wording now describes the new behavior, and the bump: patch level still fits a docs-only fix.

Non-blocking notes (not counted as findings):

• The PR description is now out of date. Its Fix section still says a scout "does not run on this agent's model" and that a settings-file env value "can reach a running session". The head commit replaced both assertions with pointers to upstream. Consider updating the body so the squash message matches what ships.
• Line 9 of the changelog fragment is now about 120 characters, longer than the lines around it. That line also chains two "and" clauses ("no longer says … for the next session, and points at …"). It's cosmetic only.

No .claude-plugin manifest, closing-issue acceptance criteria, or trusted-path file is touched by this increment. The PR says "Refs #6916" and closes no issue. I didn't run anything, so the PR's verification claims (fetch-docs.sh re-reads, check-affected.sh --tests: 14 passed, 2 skipped) are the author's and unverified.