docs: re-run the mods go/no-go after Claude Code 2.1.287 shipped mods

[kyle-sexton]

Claude Code 2.1.287 (changelog, 2026-10-01) adds Claude Mods, and code.claude.com now has a set of mods pages under plugins/mods/ (overview, reference, admin and others). Several first-mention detectors in ADR 0035 have fired:

• Criterion 2 (docs mention): holds now; the docs/plugin-philosophy.md mods row's "zero mentions in the official docs" basis, dated 2026-09-19, is superseded.
• Criterion 5: docs/upstream/claude-code-mods/ no longer matches the live pages, which drop the early-access warning.
• Criterion 1: the overview says v2.1.287 ignores CLAUDE_CODE_ENABLE_FUNCTION_HOOKS, which bears on it.
• Criterion 3: Any function-hook tool.call on Bash breaks Agent isolation: "worktree" — every Bash call refused with "isolation context for this agent was lost" anthropics/claude-code#92533 is still open, so the Defer verdict stands on criterion 3 alone.

What to do

Run the full go/no-go runbook on Claude Code 2.1.287 or later, with criterion 4 checked against the mods reference page. Then update ADR 0035, the mods row in docs/plugin-philosophy.md, and docs/upstream/claude-code-mods/sources.md with pointers to the docs pages, each with an as-of date and a recheck trigger.

Found during the Phase 7 drift triage of #5767 (2026-10-01). Related: #5733 (built-in cc-plugin-* inventory).

[kyle-sexton]

This was generated by AI during triage.

Decision defaulted: record the run and refresh the mods row and the source index, and leave ADR 0035 unamended while the verdict stays no-go — veto before merge
Basis: verified, docs/upstream/claude-code-mods/go-no-go.md:325-327 ("ADR 0035: leave it alone while the verdict holds ... a runbook run does not amend an accepted ADR by itself"). The issue body asks for an ADR 0035 update; the repository's own runbook says otherwise, so the runbook rule is the default. Alternative (reversible): also correct ADR 0035's Context sentence that says the feature is named nowhere in the official docs.

Routing: delegable, work class scoped (C3, docs evidence refresh). The verdict itself is not an open decision: go needs all five criteria, and criterion 3 fails today (anthropics/claude-code#92533 is OPEN, closedAt: null, checked 2026-10-04). Every criterion is agent-runnable headless; the Desktop, Team/Enterprise and interactive probes are optional and stay open if not run.

Related chain (no blocked-by edges added): #5777 (mod producer pilot) cannot ship a modules key until this runbook reports go; #5778 starts after #5777 reports.

Finding: the runbook's pin-bump quick check (criteria 1 to 3) did not run on the dependabot bumps to 2.1.287 (#5798) and 2.1.288 (#5998); the pin is now 2.1.288.

Agent Brief

Type: Task
Summary: Re-run the mods go/no-go runbook on the pinned Claude Code (2.1.287 or later), record the run, and refresh the records that cite the mods docs.

Current behavior:
The mods records carry a 2026-09-19 baseline at Claude Code 2.1.278: five of five criteria fail. Since then Claude Code 2.1.287 shipped mods with official docs pages under the plugins "mods" section (overview, create, reference, test, admin, troubleshoot). The mods row in the plugin philosophy doc's "Recorded gate runs" table still says the feature has zero mentions in the official docs, and the upstream mods source index has no rows for the new docs pages.

Desired behavior:

• The go/no-go runbook is run top to bottom on the installed Claude Code (version recorded three ways, as the runbook says), with the five criteria and the headless experiments (E1 to E6). Criterion 4 is checked against the mods reference page.
• The run is recorded with its date, versions and a per-criterion outcome table, in the form the runbook uses for its baseline.
• The mods row in the plugin philosophy doc gets a new Verified date and a basis that states the current reason for the verdict (for example: the docs now mention the feature; criterion 3 still fails).
• The upstream mods source index gains a row per newly relied-on docs page, each with an as-of date and a recheck trigger, and refreshed fetch dates for re-fetched rows.
• ADR 0035 stays unamended while the verdict is no-go (see the defaulted decision above).
• If the run finds all five criteria passing, do not apply a go verdict: stop, record the run, and escalate on this issue for a human decision, because a go needs a superseding ADR.

Key interfaces:

• The go/no-go runbook and its "After a run" rules are the procedure; follow them, do not restate them.
• CLAUDE_CODE_ENABLE_FUNCTION_HOOKS must be genuinely unset for criterion 1 (check the user and local settings files' env blocks first, read-only).
• The frozen research snapshot folder is never edited; a redone research pass gets a new dated folder.

Acceptance criteria:

• A recorded run exists with date, claude --version, npm latest and tag versions, and an outcome per criterion with its evidence.
• The philosophy doc's mods row has the new Verified date and a basis that no longer claims zero docs mentions.
• Every docs page the run relied on has a row in the source index with an as-of date and a recheck trigger.
• ADR 0035 is unchanged, unless the maintainer vetoes the defaulted decision.
• No probe output that contains instruction-file text, account details or other session content is committed; probe files live only in an OS temp directory.

Out of scope:

• Adding a modules key to any plugin, or any mod code under plugins/.
• Changing the @anthropic-ai/claude-code pin.
• The context-guard producer pilot (Pilot a mod producer for context-guard and rate-limit-guard #5777) and the mods adoption decision (Decide which plugin features should ship as mods #5778).
• Installing anything at user scope (the Desktop probe needs that; leave it open).

[kyle-sexton]

This was generated by AI during work-loop.

Ratify autonomous dispatch of this C3 item?

Classification: work-class: scoped (C3), task/bug-fix-shaped (re-run the go/no-go criteria and correct stale docs; triage defaulted "leave ADR 0035 unamended", veto before merge). Intended dispatch: the work-loop worker lane, through /work-items:work, to a draft PR the lane never merges. First drain is not complete and no item-level ratification exists, so it is queued here. Ratify through /work-items:attend-queue and it dispatches on a later cycle.

Lane: work-loop, instance melo-worker-1, cycle 1.

[kyle-sexton]

[kyle-sexton]

This was generated by AI during attend-queue.

Operator direction (2026-10-04): "go with the recommendations." This records that decision.

Ratification declined: draft PR #6152 already holds the work. The go/no-go was re-run on 2026-10-02 at Claude Code 2.1.288. It is recorded on the feat/5778-mod-conventions branch in draft PR #6152, which says Closes #5768. Adopting mods despite criteria 3 (#92533 still open) and 5 (the early-access phrase is still in the mods README) is an owner decision recorded as an ADR supersession, not a C3 bug fix. The item stays needs-human until #6152 merges, which closes it, or is abandoned.

Two blockers on #6152 as of 2026-10-04:

• It conflicts with main.
• Its ADR number 0049 now collides with 0049-run-ci-lanes-… on main (docs(conventions): add the pr-pipeline convention, config schema and lane ADRs #6139). The next free number is 0052.

#6154, #6156 and #6157 cite paths and lines on that branch, so their citations will go stale when the ADR is renumbered.