docs: align repo guidance, loop lanes and the release schedule with Claude Haiku 5.5

[kyle-sexton]

Refs: #6901
Refs: #2019

Summary

Repo-level guidance still reasoned from Claude Haiku 4.5: a 200k window, a Feb 2025 cutoff, no effort support. This PR rewrites it for Haiku 5.5, links the multi-agent routing rubric, adds a vendor-reported upstream record, and rewords the model-release schedule row so any Claude release triggers an instruction audit.

Fix

• docs/upstream/haiku-5-5-launch.md (new): launch-page and system-card figures, each labeled vendor-reported, plus local measurements labeled by source. Recheck on a revised card or Haiku 4.5 retirement.
• docs/plugin-philosophy.md:
  • The effort carve-out now says Haiku 5.5 takes effort on the Anthropic API while haiku is Haiku 4.5 on other providers, as a pointer with an as-of date and a recheck trigger.
  • The tier-table "Bulk mechanical sweeps" row moves from haiku to sonnet, with a rubric link.
  • Eight fired model-release records were re-dated after checking each claim. The #latest-models-comparison anchor is repointed to #compare-models, and an effort-sources line is corrected.
• docs/conventions/loop-lane/README.md and both prompts/loops/ files: Haiku stays excluded from every lane on four linked grounds (reviewer floor, early stop and unverified completion, over-refusal, provider alias). The context-window and cutoff grounds are gone, and three fired triggers are re-dated. The loop-lane CHANGELOG gets a 10.3.0 entry that supersedes the 8.1.0 reason; history is untouched.
• .github/recurring-schedule.json: the frontier-release-instruction-ablation row (id unchanged) now covers any Claude model release. audit-instructions runs on any release and unhobble on frontier only. The trigger label is claude-model-release, last_checked is 2026-10-10, and the notes name [Maintenance] Model release: run audit-instructions (any release) + unhobble (frontier only) over this repo #2019.
• plugins/work-items/reference/standing-item-preconditions.md: the row's prompt sets the release class, and the id name is historical. Fragment: work-items (patch).

Verification

• git grep for 200k, cutoff and Feb 2025 over the loop-lane README and prompts/loops/: nothing. All four grounds are present.
• routing-rubric.md is linked twice from plugin-philosophy, and no | \haiku` |row is left.vendor-reported` appears 5 times in the upstream record. The row prompt no longer says "frontier Claude model release".
• check-loop-lane-floor-drift.sh --check, check-docs-naming.sh, check-purged-em-dashes.sh, check-conformance-registry.sh, check-line-citations.py, evaluate-schedule-precondition.test.sh and check-changelog-fragments.sh --check: all pass. markdownlint: 0 issues.
• Issue [Maintenance] Model release: run audit-instructions (any release) + unhobble (frontier only) over this repo #2019 is reopened and retitled for any model release. The Haiku 5.5 audit-instructions pass is posted there: 2 findings and 1 withheld. Unhobble was not run, because this is not a frontier release.

Related

• docs: fix the false pinned-agents effort claim in plugin-philosophy #7099: two stale records found during this sweep (a false pinned-agents claim, and a dead anchor in context-guard).
• feat(playbooks): add a Claude Haiku 5.5 adaptation chapter #6861 (draft) edits the same loop-lane and plugin-philosophy lines and will conflict with this PR.

🤖 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 2026-10-11T08:15:27.368595Z	0806184	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 1m 28s —— View job

---

Security review complete

• Gather context (PR diff, changed files)
• Check skip gate / instruction-surface deletion lens
• Read instruction-exception-register
• Review changed files for security-relevant issues
• Adversarial validation of candidate findings
• Post final findings

Scope: 9 files, all documentation/config (no application code, no GitHub Actions workflows, no scripts). Diffed origin/main...HEAD at 0806184.

Instruction-surface deletion lens: This diff narrows/rewords a standing instruction — the loop-lane "Alias binding" exclusion of haiku (docs/conventions/loop-lane/README.md) drops its old "smaller context window / older knowledge cutoff" grounds and replaces them with four new grounds (reviewer floor, early-stop/unverified-completion, over-refusal, provider alias). I checked this against the instruction exception register's six Gate 0 classes (irreversible-action, secret-handling, data-integrity, external-publication, legal-compliance, agent-authority).

The substantive rule — haiku is admissible in no loop lane — is preserved and, if anything, tightened: the "Bulk mechanical sweeps" tier row moves from haiku to sonnet (docs/plugin-philosophy.md#L1240), and the loop-lane prompt files only add clarifying text, never remove the "no haiku dispatch" prohibition. No agent-authority (or other Gate-0) rule is deleted, narrowed, or softened — only its justification/pointers were swapped for newer sources. Same pattern in .github/recurring-schedule.json and plugins/work-items/reference/standing-item-preconditions.md: the release-class trigger is broadened (any release vs. frontier-only) for the read-only audit-instructions pass, while unhobble (which does transiently strip protected rules) stays frontier-only, and requires_operator_confirmation: true is untouched.

Findings: None. No injection surfaces, authZ/authN logic, secret handling, or Actions trigger/permission changes are present in this diff — it's a documentation and routing-schedule rewrite. GitHub Actions hardening is out of scope for this lane (zizmor's), and no workflow files are touched here regardless.

[github-actions]

claude-security-review has reviewed this pull request through 7541691; a later push is reviewed from there.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 080618474a

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 6m 58s —— View job

---

CI code review complete

• Gather context (PR diff, changed files)
• Skip-gate check (open, not draft, non-trivial 9-file docs/config change — review proceeds)
• Review changed files for correctness/maintainability findings
• Adversarial validation of candidate findings
• Post final findings

Scope: all 9 changed files (git diff origin/main...HEAD at 7541691), all documentation/config — no application code.

Findings: none that clear this lane's high-signal bar. What I checked and ruled out:

• The Codex review comment's P2 (schedule row vs. plugin-philosophy.md:965-969 doctrine conflict) is already fixed. It was true against commit 0806184 (what Codex reviewed), but the latest commit 7541691 rewords the "Generation-triggered ablation" bullet to match the schedule row exactly: audit-instructions on any release, unhobble on a frontier release only. No action needed here.
• Trigger rename safety (.github/recurring-schedule.json:9, frontier-model-release → claude-model-release): confirmed via git grep that no script or doc matches on the trigger string itself (only .id is matched, in evaluate-schedule-precondition.sh:55 and its test fixture), so the rename doesn't break anything.
• standing-item-preconditions.md's claim that renaming the id would break "scripts that match it": verified true — evaluate-schedule-precondition.sh case-matches the literal frontier-release-since-last-checked string.
• Link/anchor integrity: spot-checked every new/changed relative link and anchor this PR touches (docs/upstream/haiku-5-5-launch.md's links to opus-5-5-task-cost.md and the upstream-drift convention; plugin-philosophy.md's and the loop-lane README's links to routing-rubric.md and the new upstream record; the #compare-models anchor rename; issue refs #6901/#2019) — all resolve to real files/sections, and the PR's own "Related" section already tracks the one pre-existing dead anchor elsewhere (context-guard) under docs: fix the false pinned-agents effort claim in plugin-philosophy #7099, so that's correctly scoped out rather than missed.
• A candidate I raised against myself and dropped: the new CHANGELOG [10.3.0] entry states Haiku 5.5's 1M-context/current-cutoff figures inline rather than only in the new docs/upstream/haiku-5-5-launch.md provenance record. I initially read this as conflicting with that record's "guidance elsewhere copies none of its figures" line, but that line cites .claude/rules/cost-claims.md as its basis (a cost-figure-specific rule), and the exact same pattern — stating context-window/cutoff figures directly in CHANGELOG prose — is established precedent in this same file's [8.1.0] entry. Not a real inconsistency.
• Model-tier table (plugin-philosophy.md:1240): "Bulk mechanical sweeps" now resolves to the same alias (sonnet) as the row above it. Not a defect — the new prose directly beneath explains why (no routing-eval pass yet for haiku).

Not independently verified (sandbox denied the shell commands needed): markdownlint-cli2, evaluate-schedule-precondition.test.sh, and the other shell-based checks the PR's Verification section lists are author-claimed, not confirmed by me.

[github-actions]

claude-review has reviewed this pull request through 7541691; a later push is reviewed from there.