feat(playbooks): add a Claude Haiku 5.5 adaptation chapter

[kyle-sexton]

Refs: #4348
Refs: #6901

Summary

The fable-5 playbook sends each model to a per-model adaptation chapter but had none for Claude Haiku 5.5, and its meta-rule 3 still said "(e.g. Haiku today)". This adds a links-only Haiku 5.5 chapter tied to the multi-agent routing rubric and the docs/upstream/haiku-5-5-launch.md record. It builds on the chapter drafted in #6861.

Fix

• plugins/playbooks/reference/model-adaptation/haiku-5-5.md (new), in the AGENTS.md record shape, with each entry our decision plus pointer, as-of 2026-10-10 and recheck trigger:
  • "Where Haiku runs here" follows the rubric: no role routes to haiku until it passes multi-agent: paired model-routing eval for retrieval, review and mechanical-edit roles #6901.
  • Effort: high, never xhigh or max.
  • Caching and cost link the upstream record and state no figures.
  • Guide anchors were checked against the rendered page.
• skills/fable-5/SKILL.md: meta-rule 3 lists haiku-5-5.md among current models, and the example is now Haiku 4.5. The "Available models" as-of is re-dated after a re-read. The description is left unchanged, so no invocation probes are needed; the body routes the session.
• skills/fable-5/context/calibration.md: the fired record is re-dated after a re-read, and a broken Mythos availability pointer is moved to #fable-vs-mythos.
• skills/fable-5/evals/evals.json: one case where a Haiku 5.5 session loads the chapter.
• docs/plugin-philosophy.md:1250 and docs/conventions/loop-lane/README.md:419 go back to #latest-models-comparison. That is the rendered page's live id; docs: align repo guidance, loop lanes and the release schedule with Claude Haiku 5.5 #7100 had moved them to the raw heading slug, which does not resolve.
• Fragment: playbooks (minor).

Verification

• git grep "Haiku today" in SKILL.md: nothing. haiku-5-5.md is referenced.
• check-skill.sh plugins/playbooks/skills (5 pass), check-evals-quality.sh (0 warnings), check-changelog-fragments.sh --check, markdownlint, check-purged-em-dashes.sh, check-loop-lane-floor-drift.sh --check and check-conformance-registry.sh --check: all pass.
• Every file carrying a "next model release" or "any new model" trigger now has a 2026-10-1x as-of (15 files).

Related

• feat(playbooks): add a Claude Haiku 5.5 adaptation chapter #6861: the earlier draft of this chapter. This PR supersedes its chapter, fable-5 and loop-lane parts.
• docs: fix the false pinned-agents effort claim in plugin-philosophy #7099: follow-up for a false pinned-agents claim.

🤖 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 2026-10-11T08:42:34.404622Z	54ea21e	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 33s —— View job

---

Security review — PR #7108

• Gather context (diff via gh pr diff 7108, read all changed files)
• Classify request: invoked /review:security-review per custom instructions
• Check instruction-surface lens (deletions/narrowing of standing instructions) against the instruction exception register
• Report findings

Result: no security issues found.

This PR is documentation- and skill-content-only:

• A new reference chapter, plugins/playbooks/reference/model-adaptation/haiku-5-5.md, plus edits to SKILL.md, calibration.md, evals.json, a changelog fragment, and two URL-anchor fixes in docs/plugin-philosophy.md / docs/conventions/loop-lane/README.md.
• No executable code, scripts, workflow files, credentials, or data-handling paths are touched — nothing in scope for injection, authz, or secret-handling checks.
• Checked the diff against the instruction-surface lens: meta-rule 3 in SKILL.md swaps its "no chapter" example from "Haiku today" to "Haiku 4.5" but leaves the rule's substance unchanged (still: no chapter → apply the playbook generically). No Gate 0 class (irreversible-action, secret-handling, data-integrity, external-publication, legal-compliance, agent-authority) is removed, narrowed, or softened. If anything, the new chapter's "Where Haiku runs here" section adds a restriction (no role routes to haiku until it passes the routing eval), which tightens rather than loosens.
• GitHub Actions / zizmor-lane concerns: not applicable — no workflow files changed.

No findings to report, so no inline comments were posted.
· branch docs/playbooks-haiku-5-5-chapter

[claude]

Claude finished @kyle-sexton's task in 5m 1s —— View job

---

Code review complete

• Gather context (PR diff, CLAUDE.md/AGENTS.md/REVIEW.md, changed files)
• Check skip gate — PR is open and ready for review, not trivial; proceeded
• Review new chapter plugins/playbooks/reference/model-adaptation/haiku-5-5.md
• Review fable-5 skill changes (SKILL.md, calibration.md, evals.json)
• Review doc-anchor fixes (plugin-philosophy.md, loop-lane README.md)
• Review changelog fragment
• Adversarially validate candidate findings — none surfaced, nothing to verify
• Post inline comments / summary

Scope

Reviewed for correctness and alignment with CLAUDE.md/AGENTS.md, scoped to the 7 files this PR touches (git diff 9460fd7 HEAD, the PR's actual base per its commit parent). This is a docs-only PR, so REVIEW.md's code-review-lane checks (audit logging, multi-write atomicity, Actions dependency bumps) don't apply, and there's no security-lane overlap.

What I checked

• New chapter (haiku-5-5.md): follows this directory's AGENTS.md links-only convention (pointer / as-of / recheck trigger, no restated upstream text) and matches the sibling chapters' structure (sonnet-5-5.md).
• Cross-reference accuracy — fetched the live pages/sections this chapter and its siblings point at, and they resolve as claimed:
  • docs/plugin-philosophy.md#effort-floor / #effort-tiers do state the low-pin-never-on-haiku carve-out the chapter cites.
  • plugins/multi-agent/reference/routing-rubric.md#the-rule and docs/upstream/haiku-5-5-launch.md#decisions match the chapter's "no role routes to haiku until multi-agent: paired model-routing eval for retrieval, review and mechanical-edit roles #6901 passes, then high only" claim.
  • ecosystem-specialist.md#when-you-are-done and doc-drift-detector.md#when-you-are-done both exist.
  • code.claude.com/docs/en/model-config#haiku-5-5-context-window-and-pricing is a real heading on that page.
  • The calibration.md pointer swap to .../models/fable-5/overview#fable-vs-mythos lands on a real "Fable vs. Mythos" section that does cover Mythos availability — fixes the broken pointer as claimed.
  • No earlier haiku-*.md chapter exists, so "this is the first Haiku chapter" (line 15) holds.
• Stale-text sweep: confirmed repo-wide (not just SKILL.md) that no "Haiku today" or overview#compare-models references remain, so the fix in this PR is complete, not partial.
• evals.json: new case (id: 8) doesn't collide with existing ids 1–7, follows the existing case shape.
• Author-claimed verification (check-skill.sh, check-evals-quality.sh, check-changelog-fragments.sh, markdownlint, etc.) was not independently re-run — this lane's tool access didn't let me execute those scripts, so that's unverified by me, not confirmed.

Findings

None. This is a well-scoped, well-sourced docs change; I did not find a correctness or convention issue a careful reviewer would block on.
· branch

[github-actions]

claude-security-review has reviewed this pull request through 2725bca; a later push is reviewed from there.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 54ea21e031

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-review has reviewed this pull request through 2725bca; a later push is reviewed from there.

[claude]

Claude finished @kyle-sexton's task in 1m 20s —— View job

---

Incremental code review — PR #7108

Scope per lane instructions: only the change since 54ea21e in
plugins/playbooks/skills/fable-5/SKILL.md (the .claude-lane/incremental.diff), everything
else already reviewed at 54ea21e.

• Gather context (read .claude-lane/incremental.diff, full SKILL.md)
• Review the incremental change
• Adversarially verify candidate finding
• Report findings

[claude]

Claude finished @kyle-sexton's task in 30s —— View job

---

Security review — PR #7108 (incremental)

• Read review-lane scope: incremental since 54ea21e, limited to plugins/playbooks/skills/fable-5/SKILL.md
• Read .claude-lane/incremental.diff
• Apply skip gate / instruction-surface lens
• Report findings

Result: no security issues found.

The only change since the last review (54ea21e) is in
plugins/playbooks/skills/fable-5/SKILL.md,
and it's a revert: commit 2725bca ("keep the fable-5 description unchanged") removes the
, Haiku 5.5 text that 54ea21e had added to the skill's model-invoked description:
frontmatter field, per Codex's P2 comment asking for before/after invocation-measurement evidence
before extending that description.

This is a one-line edit to skill-trigger metadata (which phrases cause the skill to be
model-invoked), not a standing instruction, guardrail, or control:

• It doesn't delete, narrow, or soften anything in the instruction-surface sense (no
  irreversible-action, secret-handling, data-integrity, external-publication,
  legal-compliance, or agent-authority class is touched) — it only removes a model name from
  a trigger-phrase list.
• No code, workflow, permission, or data-handling path is touched.
• Net effect is narrower skill-invocation surface, not broader — the opposite direction from a
  risk-increasing change.

Skip gate item 2 (no security-relevant surface) applies; nothing further to check.
· branch docs/playbooks-haiku-5-5-chapter