fix: point at upstream for doctor prompt-audit write posture and relay turn cost

[kyle-sexton]

No related issue: follow-ups found while landing #5827. The two findings that need a design decision are filed as #5832 and #5833.

Summary

• harness-config:audit-instructions said that /doctor prompt-audit's write posture was undocumented. The upstream memory page now documents it.
• fleet:reach's relay probe table stated a per-turn dollar figure. It now links the official cost-tracking docs instead.

Fix

• plugins/harness-config/skills/audit-instructions/SKILL.md, doctor Boundary section:
  • The mutation gate states our decision: applying the proposed edits is the person's call, and the skill never chains into /doctor.
  • The availability line notes that the subcommand needs the bundled claude-api skill to be on. SKILL.md stays at 499 lines.
• reference/native-doctor.md: the "undocumented" row is replaced by a links-only record. It points to https://code.claude.com/docs/en/memory#audit-your-instruction-files, carries the as-of date 2026-10-02, and gives a recheck trigger.
• plugins/fleet/skills/reach/reference/relay.md: the Cost row keeps what the cost went to (SessionStart hooks and context loading) and links https://code.claude.com/docs/en/costs#track-your-costs.
• Version bumps with CHANGELOG entries: harness-config to 1.3.2 (docs(upstream): answer "What a task costs on Opus 5.5" with routing, observability and pointer updates #5827 holds 1.3.1) and fleet to 0.2.2.

Verification

• scripts/check-changelog-parity.sh --check, --check-bump origin/main, --check-order and --check-preserved origin/main all pass.
• scripts/check-changed-skills.sh origin/main: 2 skills, 0 errors. The one warning is pre-existing, in the untouched reference/criteria.md.
• markdownlint-cli2 reports 0 issues on the 5 touched markdown files. typos is clean.
• run-plugin-tests.sh passes on the 6 audit-instructions suites plus check-doc-citations and check-structure.

Related

• docs(upstream): answer "What a task costs on Opus 5.5" with routing, observability and pointer updates #5827: the session that found these.
• mod-authoring rule glob plugins/*/types/** matches no file, failing the instruction-placement gate #5832: the mod-authoring rule glob matches no file.
• discovery research gate passes an artifact with zero accepted claims without saying so #5833: the discovery research gate passes an artifact with zero accepted claims.

🤖 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 2026-10-02T16:46:10.422475Z	26ae2d4	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.

[claude]

Claude finished @kyle-sexton's task in 2m 35s —— View job

---

Code review complete

• Load /review:code-review instructions
• Gather PR diff and context
• Review changed files for correctness/maintainability issues
• Post findings (inline + summary)

Scope: git diff origin/main...HEAD (7 files, docs/changelog/manifest only — no scripts or
executable code changed).

Findings: None that clear the high-signal bar.

Notes from the review (not blocking):

• plugins/harness-config/skills/audit-instructions/SKILL.md (doctor Boundary section) drops the
  old "it survives disableBundledSkills" clause while adding the new claude-api dependency
  clause. At first pass this reads like silently-dropped information, but it's actually consistent:
  if prompt-audit now depends on the bundled claude-api skill (itself subject to
  disableBundledSkills), restating an unqualified "survives disableBundledSkills" would be
  misleading for that subcommand. The full detail is still preserved in
  reference/native-doctor.md,
  so this isn't a defect — just worth a sentence in the PR description for the next reader.
• The new reference/native-doctor.md row (pointer + 2026-10-02 as-of + recheck trigger) conforms
  to the upstream-drift convention's required parts and to
  .claude/rules/skill-bodies-state-current-rules.md.
• Version sequencing (harness-config 1.3.0 → 1.3.2, intentionally skipping 1.3.1 reserved by
  docs(upstream): answer "What a task costs on Opus 5.5" with routing, observability and pointer updates #5827) is fine under check-changelog-parity.sh: --check-bump only requires strictly-greater
  and newly-added, --check-order only requires no gaps/repeats in heading order, so the skip
  needs no special annotation.
• plugins/fleet/skills/reach/reference/relay.md and the two CHANGELOG.md/plugin.json pairs
  are straightforward, internally consistent, and match their stated versions.

Author-claimed verification (markdownlint, typos, changelog-parity, check-changed-skills,
run-plugin-tests.sh) was not independently re-run in this review.

[claude]

Claude finished @kyle-sexton's task in 39s —— View job

---

Security review complete

• Gather context (PR diff, changed files)
• Apply skip gate
• Review diff for security-relevant findings (including instruction-surface deletion lens on SKILL.md changes)
• Post final results

No security issues found.

This PR is documentation/metadata-only: two version bumps (plugin.json), two CHANGELOG.md entries, and three reference/skill doc edits that replace a stated dollar figure and an "undocumented" uncertainty note with links-only pointers to upstream docs (consistent with this repo's upstream-drift convention). No code paths, input handling, auth/access logic, or CI workflow files are touched.

I checked the one text change that could plausibly weaken a guardrail — the doctor Boundary "Mutation gate" line in plugins/harness-config/skills/audit-instructions/SKILL.md:

Old: "Its write posture is undocumented; this skill never chains into /doctor."
New: "Applying its proposed edits is the person's call; never chain into /doctor."

The substantive control — the skill never invokes /doctor itself, only offers it to the person — is unchanged in both versions. The edit only replaces the reasoning (uncertain write posture) with a stated decision now that upstream documents the write posture, per reference/native-doctor.md. No guardrail was removed or narrowed.

No findings to post inline.

Branch ·

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 26ae2d41a1

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".

[claude]

Claude finished @kyle-sexton's task in 3m 3s —— View job

---

Code review complete

• Load /review:code-review instructions
• Gather PR diff and context (git diff origin/main...HEAD, head 8492bdf)
• Apply skip gate (not a repeat review at this head — a new commit landed since the prior review at 26ae2d4)
• Review changed files for correctness/maintainability issues
• Adversarially verify candidate findings
• Post findings (inline + summary)

Scope: 7 files, docs/changelog/manifest only (no scripts or executable code).

Findings: None that clear the high-signal bar. No inline comments posted.

This commit (8492bdf) is itself a follow-up to the Codex and self-review findings posted
against 26ae2d4 (dropping the hardcoded claude-api gate list from the SKILL.md availability
line, and stamping the relay.md Cost row with an as-of date and recheck trigger). I checked that
the fix is complete and internally consistent rather than re-raising what it already addresses:

• Verified against the live upstream page
  (code.claude.com/docs/en/memory#audit-your-instruction-files):
  it does state "The audit runs through the bundled /claude-api skill" and that nothing changes
  until the user applies proposed edits, so the claim this PR adds in
  reference/native-doctor.md line 14
  is accurate, not a hallucinated dependency.
• SKILL.md lines 127-130
  now states the decision and points at the records without re-listing the specific gates, matching
  skill-bodies-state-current-rules.md and the reviewer's "Fix:" comment on 26ae2d4.
• The harness-config CHANGELOG.md 1.3.2 entry was itself edited in this second commit to drop
  the stale "needs the bundled claude-api skill on" wording, so the changelog and the skill body
  stay in sync.
• relay.md line 210's
  Cost row now carries its own as-of date and recheck trigger, independent of the section-level
  stamp above the table (which only covers the cli-reference/agent-view pages) — checked against
  the live costs page's "Using the /usage command" section, which is a plausible target for that
  recheck trigger.
• Version sequencing (harness-config 1.3.0 → 1.3.2, fleet 0.2.1 → 0.2.2) and SKILL.md's line
  count (499, matching the PR description) check out.

Author-claimed verification (markdownlint, typos, changelog-parity, check-changed-skills,
run-plugin-tests.sh) was not independently re-run in this review.
·
Branch

[claude]

Claude finished @kyle-sexton's task in 1m 2s —— View job

---

Security review complete

• Gather context (PR diff, changed files, HEAD SHA 8492bdf)
• Invoke /review:security-review and apply its skip gate / instruction-surface lens
• Post final results

No security issues found.

Scope

git diff origin/main...HEAD (7 files): two plugin.json version bumps, two CHANGELOG.md
entries, and three reference/skill doc edits replacing a stated dollar figure and an
"undocumented" uncertainty note with links-only pointers to upstream docs, per this repo's
upstream-drift convention.
No code paths, input handling, auth/access logic, secrets, or CI workflow files are touched — the
skip gate's "no security-relevant surface" condition applies, except for one instruction-surface
edit checked explicitly below.

Instruction-surface deletion lens

The Boundary section in
plugins/harness-config/skills/audit-instructions/SKILL.md
rewrites the doctor "Mutation gate" and "Availability is never assumed" lines. I checked this
against the instruction exception register's
six protected classes (irreversible-action, secret-handling, data-integrity,
external-publication, legal-compliance, agent-authority):

• The actual control — this skill is report-only and never invokes /doctor itself, only offers
  it to the person — is preserved word-for-word in substance across old and new text ("this
  skill never chains into /doctor" → "never chain into /doctor").
• What changed is reasoning, not the rule: the old text hedged on an "undocumented" write posture;
  the new text states a decision now that upstream documents it
  (reference/native-doctor.md#L14),
  and the hard-coded gate list (DISABLE_DOCTOR_COMMAND, skillOverrides, version, bundled
  claude-api dependency) is replaced with "read live from the pointers in the records" — the
  correct shape under .claude/rules/skill-bodies-state-current-rules.md, and the fix the author
  already applied in response to Codex's P2 comment on this same line.
• No Gate 0 class is weakened; if anything the agent-authority constraint (never chain into
  /doctor) reads more definite than before, not less.

plugins/fleet/skills/reach/reference/relay.md's Cost row (also previously flagged by Codex) now
carries both an as-of date and a recheck trigger, satisfying the pointer convention. The two
CHANGELOG.md/plugin.json pairs are consistent, metadata-only changes.
·
Branch

[claude]

Claude finished @kyle-sexton's task in 51s —— View job

---

Code review complete

• Load /review:code-review instructions
• Apply skip gate
• Gather PR diff and context (git diff origin/main...HEAD, head c42ffad)
• Review changed files for correctness/maintainability issues
• Post findings (inline + summary)

Skip gate hit: c42ffad is a merge of main into this branch with no content changes —
git diff 8492bdf..c42ffad -- plugins/fleet plugins/harness-config is empty, confirming the
tree for all 7 changed files is byte-identical to 8492bdf. That commit already received a full
code review from this lane (no findings) plus a security review (no findings), both posted above.
Re-reviewing unchanged content would just restate those results, so per the skip gate ("this head
already has a successful review from this lane that still applies") I'm stopping here rather than
re-litigating.

No new findings. No inline comments posted.

For reference, the substance already covered at 8492bdf:

• plugins/harness-config/skills/audit-instructions/SKILL.md's
  doctor Boundary section states the mutation-gate decision and reads gates live from the
  pointer in reference/native-doctor.md, addressing both Codex P2 comments on 26ae2d4.
• plugins/harness-config/skills/audit-instructions/reference/native-doctor.md#L14
  carries the pointer, the 2026-10-02 as-of date, and a recheck trigger.
• plugins/fleet/skills/reach/reference/relay.md#L210's
  Cost row now carries its own as-of date and recheck trigger, independent of the section-level
  stamp above the table.
• Version bumps (harness-config 1.3.0 → 1.3.2, fleet 0.2.1 → 0.2.2) and their CHANGELOG.md
  entries stay consistent with the body text.

Author-claimed verification (markdownlint, typos, changelog-parity, check-changed-skills,
run-plugin-tests.sh) was not independently re-run in this review.
·
Branch

[claude]

Claude finished @kyle-sexton's task in 58s —— View job

---

Security review complete

• Gather context (PR diff, changed files, HEAD SHA c42ffad)
• Load /review:security-review and apply its skip gate
• Compare current diff against the last-reviewed commit
• Post final results

No security issues found — skip gate hit (condition 3: this head already has a successful security review that still applies).

c42ffad is a merge of main into this branch. git diff origin/main...HEAD for this PR is
byte-for-byte identical to git diff origin/main...8492bdf, the commit the security review already
completed and reported clean on (see the two "Security review complete" comments above, at
2026-10-02T16:43:19Z and 2026-10-02T16:48:27Z). The merge brought in unrelated upstream commits
(e.g. a287fa4, fb2a38b, 28c0493, …) but introduced no new changes to this PR's own 7 files:

• plugins/fleet/.claude-plugin/plugin.json, plugins/fleet/CHANGELOG.md,
  plugins/fleet/skills/reach/reference/relay.md
• plugins/harness-config/.claude-plugin/plugin.json, plugins/harness-config/CHANGELOG.md,
  plugins/harness-config/skills/audit-instructions/SKILL.md,
  plugins/harness-config/skills/audit-instructions/reference/native-doctor.md

The one instruction-surface edit (the doctor Boundary "Mutation gate" / "Availability is never
assumed" lines in
SKILL.md)
was already checked against the instruction exception register's protected classes in the prior
review and found to preserve the substantive control (never chain into /doctor) while only
replacing hedged reasoning with a stated decision now that upstream documents it — see
reference/native-doctor.md#L14.
No new text to re-check here.

Still a docs/changelog/manifest-only PR: two version bumps, two changelog entries, three
reference/skill doc edits. No code paths, input handling, auth/access logic, secrets, or CI
workflow files touched.
·
Branch