feat: prefer a cross-vendor advisor at fresh-eyes checkpoint sites fleet-wide#934
Conversation
…eet-wide PLUGIN-PHILOSOPHY "Fresh-eyes checkpoints" gains the independence-ladder escalation rule: at high-stakes checkpoint sites, prefer a cross-vendor advisor when one is installed (e.g. the OpenAI Codex plugin's /codex:review), with the fresh-context same-vendor subagent as the stated fallback. Eleven plugins ship one gate+fallback sentence at each primary delegation-instructing site, modeled on verification:confirm's already-shipped canonical phrasing; patch bump + CHANGELOG entry each. Closes #933 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3e1331a6bb
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".
Bare /codex:review can background and return only a status handle, so a verdict flow would silently complete without the preferred reviewer's findings (fanout already documents --wait as load-bearing). Applies to every command-bearing advisor site in this change and, same defect, to the canonical verification:confirm site (0.3.4). Addresses the Codex P2 review finding on quality-gate self mode. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dce9092006
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".
work-items CHANGELOG: both sides released 0.21.1 — main's #922 keeps it; this branch's advisor entry moves to 0.21.2 with the matching plugin.json bump.
|
Claude finished @kyle-sexton's task in 4m 28s —— View job PR Review
Overall the fleet-wide escalation pattern is clean and well-disciplined: the independence-ladder doctrine lands once in Finding 1 — Unresolved Codex P2: missing
|
|
Claude finished @kyle-sexton's task in 3m 18s —— View job Security Review
This PR is pure documentation/instruction text — Markdown skill files and IMPORTANT — Missing
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5e19ad8e8a
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".
…aming generator escalation Diff-review advisor sites (quality-gate self, implement-dispatch, fable-5 verification, workflow review stage, verification:confirm) now instruct passing the resolved diff base via --base, mirroring fanout's load-bearing flag — without it Codex auto-picks the working tree or default branch and reviews a different change set. naming reverted to 0.3.0: the codex plugin ships only diff-review surfaces, so a cross-vendor GENERATOR preference is uninstantiable today. Deferred with trigger: a generation-capable cross-vendor adapter ships. Addresses the two Codex P2 findings. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 18s —— View job PR Review — follow-up on commit
|
| Site | Fixed |
|---|---|
review/skills/quality-gate/context/self.md |
✓ |
implementation/skills/implement-dispatch/SKILL.md |
✓ |
playbooks/skills/fable-5/context/orchestration.md |
✓ |
session-flow/skills/workflow/context/steps.md |
✓ |
verification/skills/confirm/SKILL.md |
✓ |
The phrasings are consistent and mirror review:fanout's load-bearing flag pattern.
Finding 2 (low) — session-flow/steps.md without --base → CLOSED
The line now reads "passing the stage's resolved diff base via --base" — slightly vaguer than the SKILL.md sites (those say "the resolved diff base" with a known resolution path in Shared Inputs), but for a guide/checklist context this level of instruction is proportionate.
Naming generator escalation retraction → CLOSED
naming is reverted to 0.3.0. The SKILL.md cross-vendor generator instruction is removed, the CHANGELOG entry is removed, and the plugin.json version is back to 0.3.0. The Codex P2 finding (naming a command that doesn't exist in the Codex plugin's exposed surface) is correctly resolved. The commit message records the trigger condition for re-opening: "a generation-capable cross-vendor adapter ships."
Remaining sites without --base — scoping is correct
The prior security review flagged 13 sites; this commit addressed 5. The 8 remaining sites are not diff-review contexts, so --base is inapplicable to them:
| Site | Context — why --base doesn't apply |
|---|---|
planning/skills/plan/SKILL.md |
Codex reviews a plan artifact, not a PR diff |
planning/skills/devils-advocate/SKILL.md |
Same — plan artifact or incumbent identity |
claude-config/skills/audit-instructions/SKILL.md |
Codex refutes an instruction-culling proposal (no diff base) |
codebase-health/skills/audit/SKILL.md |
Codex re-verifies a codebase health finding set |
work-items/reference/pipeline-shape.md |
Abstract pipeline description; concrete --base lives in the pointed-to verification:* skills (now fixed) |
re-anchor/context/re-anchor-audit-correct.md |
Working-tree re-derivation; no PR diff base to resolve |
docs-hygiene/skills/audit-derivability/SKILL.md |
Spot-testing document derivability from exploration (noted as less applicable in the security review) |
docs-hygiene/skills/compress/SKILL.md |
Reviewing semantic drift in compressed content, not a PR diff |
The commit's scope decision ("diff-review advisor sites") is the correct categorization.
Finding 3 — Incidental commit-convention row removal → OBSERVATION, UNCHANGED
The row pointing to docs/conventions/commit-convention/ (a directory that doesn't exist in the repo) was removed in the merge of origin/main. This appears deliberate — the dead pointer is gone. Still not mentioned in the PR description; still worth confirming intentional if not already done.
Remaining open observation (from security review)
The namespace-collision risk (any plugin installed under the name codex receives these delegated calls) is still a latent gap, bounded by whatever plugin-system namespacing guarantees exist. It's unchanged by this commit and remains low-priority — only a consumer who installs a rogue plugin named codex is exposed.
Summary
Both Codex P2 findings are closed by dc3a8d9. The --base fix is consistent and correctly scoped to PR-diff contexts. The naming revert is clean and well-reasoned. The PR is in good shape for merge.
|
Claude finished @kyle-sexton's task in 4m 4s —— View job Security Review
This PR is pure documentation/instruction text — Markdown skill files and Previously flagged issues now resolved at HEAD (dc3a8d9):
SUGGESTION —
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dc3a8d9910
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".
…ation
Three review rounds showed per-site command examples drifting against
the surface the advisor plugin owns (--wait, --base, committed-only
diffs, non-diff artifacts). Sites now name the capability and the
fallback only — gate extended with artifact fit ("when its documented
surface can take this artifact, invoked per its own docs") — and the
philosophy records the rule: a checkpoint site never restates the
advisor's command flags. fanout remains the one concrete codex
integration; it resolves its own diff base, so its flags are
well-formed where they live.
Addresses the three round-3 Codex P2 findings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 7s —— View job PR Review — commit
|
| Finding | Resolution |
|---|---|
verification:confirm — no review base resolved before --base required |
Flags removed; "invoked per its own docs" delegates base selection to the advisor |
quality-gate:self — uncommitted edits dropped from preferred review |
"when its documented surface can take this artifact" gates capability-incompatible surfaces to fallback |
planning:plan — plan stress-test routed to a diff-review surface |
Same capability gate; advisor's docs govern what artifacts /codex:review accepts |
Non-diff-review sites (docs-hygiene, codebase-health, claude-config) |
Same gate resolves the wrong-artifact concern across all four |
The prior security finding — "non-diff-review sites auto-select wrong artifact" — is resolved by the new gating clause.
What's clean
Architectural principle established once. docs/PLUGIN-PHILOSOPHY.md now states:
Invocation mechanics — synchronous waiting, diff-base selection, which artifacts a surface can judge — are the advisor plugin's own documentation's concern: a checkpoint site names the capability and the fallback, never the advisor's command flags, which drift against the surface their owner evolves.
This prevents the flag-omission recurrence pattern that produced three rounds of P2 findings. The principle is correctly placed at the doctrine level, not replicated at sites.
"When its documented surface can take this artifact" gating. This clause is consistently applied at all new sites that are not guaranteed diff-review contexts (claude-config, codebase-health, docs-hygiene, planning, re-anchor, work-items). Diff-review sites that previously had --base (quality-gate, verification:confirm, implement-dispatch, fable-5, session-flow:steps) use the same gating, which is equally sound — the Codex plugin's own docs state when --base is required.
25 files updated consistently. Every site in the fleet carries the same capability-generic phrasing. The CHANGELOG entries are updated to match. No sites carry the old command-flag phrasing.
verification:confirm CHANGELOG is accurate. The 0.3.4 entry correctly notes that the site "no longer names advisor commands" and that "per-site command flags drift against the surface the advisor owns" — this matches the commit's stated design intent.
Observation — session-flow/steps.md prose wrapping
plugins/session-flow/skills/workflow/context/steps.md line 288–290
The new line is unwrapped compared to every other advisor-escalation sentence in the fleet:
- For a high-stakes diff, prefer a cross-vendor reviewer when one is installed — e.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs — with the fresh-context same-vendor subagent as the fallback,
never a route to a command that may not resolve
The prior revision had a natural break before "with the fresh-context…" The sentence is now a single 220-character line. This is a prose-consistency observation, not a correctness issue.
Remaining latent gap (unchanged from prior reviews)
The plugin namespace collision risk (any plugin installed under the codex namespace receives these delegated calls) is structurally unchanged — the sites still reference "the OpenAI Codex plugin" by name, not by install ID. The precondition for exploitation is control over the consumer's plugin installation. With the capability-generic phrasing this risk is no more or less exposed than it was in dc3a8d9; it's bounded by the plugin system's namespacing guarantees.
Summary
All open Codex P2 findings are resolved by 1ce76f0. The commit takes a principled architectural approach — delegating flag mechanics to the advisor plugin's own documentation — rather than patching individual omissions, which is the correct long-term fix. The session-flow/steps.md line-wrapping is a minor prose note. The PR is ready to merge on its technical merits.
|
Claude finished @kyle-sexton's task in 3m 29s —— View job Security Review
This PR is pure documentation/instruction text — Markdown skill files and Issues resolved since last security review (dc3a8d9 → 1ce76f0):
This closes all three open Codex P2 findings from prior reviews:
SUGGESTION — Plugin namespace collision is now fleet-wide across 11+ skill sitesConfidence: PLAUSIBLE Every checkpoint site now says "e.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs." The invocation resolves at runtime by reading the installed advisor's docs — whichever plugin the agent finds in the
The eleven checkpoint sites added by this PR do not include that qualifier. If a consumer has any plugin installed under the This was flagged in the prior security review (dc3a8d9). Commit 1ce76f0 removes the concrete command names, which reduces the direct exploitability (a rogue Precondition for exploitation: control over plugin installation in the consumer's environment (install a plugin under the expected advisor namespace). This limits practical severity. The only mitigation available today without a platform-level namespace-pinning mechanism is to note the intended install id in the doctrine at SUGGESTION — Fleet-wide data egress scope is not surfaced at advisor installationConfidence: PLAUSIBLE When a consumer installs the OpenAI Codex plugin (or any cross-vendor advisor), high-stakes artifacts flow to the advisor's vendor at 11+ sites across 10 plugins: diffs, plan artifacts, instruction-culling proposals, codebase health findings, document derivability spot-tests, compression semantic-diff audits, re-anchor re-derivation artifacts, and orchestration verification verdicts. A user who installs This is bounded by the user's own installation choice, so severity is low. But the intent-vs-scope gap is material: a user who audited the egress scope for No other security issues foundNo GitHub Actions workflow files were changed. No secrets, credentials, or tokens appear in the diff. No command injection surfaces exist (all changes are static instruction text). No path traversal, SQL injection, or XSS surfaces. No The |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1ce76f0481
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".
Sites gated on installation alone while the doctrine gates on "installed and set up" — an installed-but-unauthenticated advisor would be selected and prompt or fail instead of falling back. Addresses the round-4 Codex P2 finding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 44s —— View job Claude finished @kyle-sexton's task — PR Review — commit
|
| Site | Fixed |
|---|---|
review/skills/quality-gate/context/self.md |
✓ |
verification/skills/confirm/SKILL.md |
✓ |
implementation/skills/implement-dispatch/SKILL.md |
✓ |
planning/skills/plan/SKILL.md |
✓ |
planning/skills/devils-advocate/SKILL.md |
✓ |
playbooks/skills/fable-5/context/orchestration.md |
✓ |
session-flow/skills/workflow/context/steps.md |
✓ |
claude-config/skills/audit-instructions/SKILL.md |
✓ |
codebase-health/skills/audit/SKILL.md |
✓ |
docs-hygiene/skills/audit-derivability/SKILL.md |
✓ |
docs-hygiene/skills/compress/SKILL.md |
✓ |
re-anchor/context/re-anchor-audit-correct.md |
✓ |
work-items/reference/pipeline-shape.md |
✓ |
The doctrine in docs/PLUGIN-PHILOSOPHY.md:348 already read "when one is installed and set up"; every checkpoint site now matches it. A plugin that is installed but not authenticated or configured falls through to the fresh-context same-vendor fallback, which is the correct behavior.
session-flow:orchestrate — deliberately excluded, correctly
plugins/session-flow/skills/orchestrate/SKILL.md:52 reads "when one is on hand" rather than "when one is installed and set up." This site was excluded from 96c3ad6 for the same reason the prior review accepted it: imperative 3 is exported verbatim into a model-/tool-agnostic worker brief that leaves the Claude Code environment entirely — "installed" is not the right framing outside a Claude Code context. "On hand" is a looser, tool-agnostic reading of the same availability intent. The exclusion is correct.
All four rounds of P2 findings are now closed
| Round | Finding | Commit |
|---|---|---|
| 1 | Missing --wait for synchronous advisor invocation |
dce9092 |
| 2 | Missing --base at diff-review sites |
dc3a8d9 |
| 3 | Non-diff artifacts routed to git-diff-only surface; plan stress-test on diff-only surface; unresolvable --base at confirm |
1ce76f0 (capability-generic re-derivation) |
| 4 | Set-up readiness gate missing from 13 checkpoint sites | 96c3ad6 |
Latent gaps — structurally unchanged from prior reviews
Plugin namespace collision (PLAUSIBLE): Sites reference "the OpenAI Codex plugin" by command prefix (codex:), not by install id (codex@openai-codex). Any plugin installed under the codex namespace would receive delegation. Precondition: control over plugin installation. Mitigation available today: note the expected install id in the doctrine. No regression introduced by this commit.
Data egress scope (PLAUSIBLE): A consumer who installed codex@openai-codex for review:fanout now gets advisor routing at 11+ sites across 10 plugins. The expanded scope is not surfaced at the install boundary. Bounded by the user's own installation decision. A ## Data egress note in the plugin README or the MIGRATION-PLAYBOOK acceptance checklist would close it.
Both gaps are pre-existing; neither is introduced or worsened by 96c3ad6.
What's clean
- Four rounds of P2 findings closed. The architectural decision in
1ce76f0(capability-generic sites, invocation mechanics delegated to the advisor plugin's own docs) eliminated the recurring flag-omission failure mode.96c3ad6completes the last surface-level fix the approach requires. - Doctrine consistency. Every checkpoint site now matches the PLUGIN-PHILOSOPHY.md doctrine exactly.
review:fanoutuntouched. The one concrete Codex integration, which handles its own diff base, remains unchanged.- CHANGELOG entries. Touched plugins reflect the "installed and set up" alignment accurately.
Summary
All open Codex P2 findings across all four review rounds are resolved. The PR is ready to merge.
|
Claude finished @kyle-sexton's task in 2m 31s —— View job Security Review
This PR is pure documentation/instruction text — Markdown skill files and Issues resolved at HEAD (
SUGGESTION — "Invoked per its own docs" is explicit trust delegation to advisor plugin documentation, fleet-wideConfidence: PLAUSIBLE The architectural fix in
Every one of the 11+ new checkpoint sites carries this: "e.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs." This is explicit trust delegation to the advisor plugin's documentation. The invoking agent must read the advisor plugin's docs in two ways: (1) to determine artifact-fit eligibility, and (2) to determine invocation mechanics. If an advisor plugin's documentation files contain adversarial content — through supply-chain compromise of that plugin, or through a consumer installing a rogue plugin under the expected namespace — the invoking agent at each of the 11+ checkpoint sites is explicitly instructed to read and follow those docs. The prior approach (naming specific flags like Attack path: Supply-chain compromise of the advisor plugin (or installation of a rogue plugin under the Precondition: Requires control over a plugin installed in the consumer's environment. This limits practical severity. The SUGGESTION — Fleet-wide data egress scope is undisclosed at the advisor install boundaryConfidence: PLAUSIBLE When a consumer installs any cross-vendor advisor plugin, the new fleet-wide escalation routes high-stakes artifacts to that advisor's backend from 11+ sites across 10 plugins: implementation diffs, plan artifacts, instruction-culling proposals, codebase health finding sets, document derivability spot-tests, compression semantic-diff audits, re-anchor re-derivation artifacts, orchestration verification verdicts. A user who installs Concrete gap: The plugin-acceptance security review checklist at SUGGESTION — Plugin namespace collision is now fleet-wideConfidence: PLAUSIBLE Previously flagged in prior reviews; unchanged by this PR. Documenting for completeness since blast radius grew substantially.
One mitigating change from prior review: The capability-fit gate ("when its documented surface can take this artifact") means a rogue plugin would also need to document artifact support for each site type. This slightly raises the bar for a rogue plugin to successfully intercept all 11 sites. The risk is still latent and bounded by consumer plugin installation practices. No other security issues foundNo GitHub Actions workflow files were changed by this PR (the The |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 96c3ad6730
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".
…visor line The exported imperative gated on "on hand" alone; an unauthenticated or artifact-unfit advisor would be selected instead of the mandatory fresh-context fallback. Now "set up and able to judge this artifact", still naming no command (the brief is exported into tool-agnostic contexts). Addresses the round-5 Codex P2 finding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 54s —— View job Claude finished @kyle-sexton's task — PR Review — commit
|
| Round | Finding | Commit |
|---|---|---|
| 1 | Missing --wait for synchronous advisor invocation |
dce9092 |
| 2 | Missing --base at diff-review sites |
dc3a8d9 |
| 3 | Non-diff artifacts routed to git-diff-only surface; plan stress-test on diff-only surface; unresolvable --base at confirm |
1ce76f0 (capability-generic re-derivation) |
| 4 | Set-up readiness gate missing from 13 checkpoint sites | 96c3ad6 |
| 5 | orchestrate imperative 3 gated on "on hand" only — missing readiness + artifact-fit |
9218ec4 |
Observation — CHANGELOG for session-flow 0.13.1 predates the round-5 fix
The 0.13.1 CHANGELOG entry for orchestrate describes the tool-agnostic, no-command nature of the site but doesn't mention the "set up and able to judge this artifact" gate that 9218ec4 adds. The entry isn't incorrect — it just omits the gate condition. A one-phrase addendum ("presence-gated on readiness and artifact fit") would make the CHANGELOG accurately reflect HEAD, but this is cosmetic and not a merge blocker.
Observation — imperative 7 still uses "on hand" for conditions calibration
orchestrate/SKILL.md line 67: whether a capable advisor/verifier is on hand is a calibration condition (sizing input), not a delegation gate. "On hand" is appropriate here — it describes a factor to weigh in sizing decisions, not an instruction to select an unverified advisor. Correctly untouched.
Remaining latent observations (structurally unchanged)
Plugin namespace collision (PLAUSIBLE): Sites reference "the OpenAI Codex plugin" by command prefix, not by install id (codex@openai-codex). A plugin installed under the codex namespace that is not the genuine plugin would receive fleet-wide delegation. Precondition: control over consumer plugin installation. No regression from this commit.
Data egress scope (PLAUSIBLE): A consumer who installed codex@openai-codex for review:fanout now gets advisor routing at 11+ sites across 10 plugins. The expanded egress scope is not surfaced at the install boundary. Bounded by the user's own installation decision.
Neither gap is introduced or worsened by 9218ec4.
Summary
All five rounds of Codex P2 findings are closed. The final commit is a clean, minimal fix that correctly aligns the orchestrate export-bound site with the fleet-wide readiness + artifact-fit gate, preserving its tool-agnostic invariant. The PR is ready to merge.
|
Claude finished @kyle-sexton's task in 1m 51s —— View job Security Review — commit
|
…rine (#941) ## Summary PLUGIN-PHILOSOPHY's design boundary is the plugin specialization of the org-wide convention now owned by `melodic-software/standards` `conventions/engineering/shareable-artifact-design.md` (standards #237, merged in standards PR #238). The design boundary states the specialization relationship; the authoritative references gain the explicit entry. Completes the follow-up promised in #933's companion work. No linked issue: the tracked items (#933, standards#237) are already closed by their own PRs; this is their promised follow-up citation. ## Related - melodic-software/standards#237 — the convention doc this cites (closed by standards PR #238; not closed by this PR). - #933 — the advisor-escalation work this citation follow-up was sequenced after (already closed by #934). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 participant
Summary
Fleet-wide escalation of the fresh-eyes checkpoint rule: at sites whose verdict is high-stakes and where correlated blind spots are the risk, prefer a cross-vendor advisor when one is installed AND its documented surface can take the judged artifact, invoked per that plugin's own docs — with the fresh-context same-vendor subagent as the stated fallback. Sites never restate the advisor's command flags (three review rounds showed per-site --wait/--base/diff-semantics examples drifting against the surface the advisor owns);
review:fanoutremains the one concrete codex integration, since it resolves its own diff base.docs/PLUGIN-PHILOSOPHY.md"Fresh-eyes checkpoints" gains the independence-ladder escalation rule; sites carry only the gate+target+fallback sentence (rule stated once).verification:confirm's canonical site is aligned to the same capability-generic shape; fanout, implement→confirm, and debug env-hook untouched.### Changedentry per touched plugin.Touched plugins (all patch bumps)
planning 0.24.4, implementation 0.7.8, review 0.14.11, codebase-health 0.7.2, claude-config 0.9.2, work-items 0.21.2 (0.21.1 taken by #922 on main; resolved in merge), re-anchor 0.5.1, docs-hygiene 0.8.4, verification 0.3.4 (canonical confirm site gains --wait, same defect as the Codex P2 finding), session-flow 0.13.1, playbooks 0.3.2
Two site-level judgment calls worth review attention:
session-flow:orchestrate's sentence names no command — imperative 3 is exported verbatim into a model-/tool-agnostic worker brief where a plugin command cannot resolve (gate + cross-vendor target + fallback kept, example dropped);playbooks:fable-5was aligned at its SSOT (context/orchestration.md, which already named a cross-vendor reviewer) rather than touched at its two pointer sites.re-anchor gets ONE sentence in the shared method doc (
context/re-anchor-audit-correct.mdstep 3) — all eight skills read it at invocation; per-skill restatement would be the second copy the point-dont-copy discipline prohibits. The-deepfan-out skills (do-your-research-deep, recheck-against-upstream-deep) declare deltas replacing step 3 with throttled per-claim fan-outs the single-verdict escalation does not govern — skipped, no contradiction.Deliberate skips (audit rows excluded)
recheck-against-upstream— doc-source independence dominates model independence at that site.session-flow:running-retro— own-session transcript analysis; cross-model value weak.tidy,batch-simplify, planning design gates) — a script's pass/fail needs no advisor.Method caveat
Site list from a fleet sweep (single Explore agent, self-declared completeness caveat) over all plugins for author-verifier / plan-attacker / self-grade delegation sites. Sites the sweep missed are follow-up filings, not silent scope.
Related
Closes #933
🤖 Generated with Claude Code