[kyle-sexton]

Closes #4346

Summary

The playbooks:fable-5 pack is now authored by Claude Fable 5.1 (model id claude-fable-5-1), regenerated in a live Fable 5.1 session per the issue's Handoff procedure and the owner's Option A decisions of 2026-09-29 and 2026-10-01. The doctrine transfers: as the author I rejected no core line or chapter, which matches the Fable 5.1 prompting guide's own statement that Fable 5 prompts carry to 5.1 without changes. So the regeneration is a fold and re-authoring, not a rewrite: the guide's handful of deltas become model-neutral first-person doctrine in their owning chapters, fable-5-1.md is retired, the six sibling chapters are re-baselined to the new author, the three stale verification records are re-derived with 2026-11-01 four-part records, and the evals assert the new routing.

Where the issue's two acceptance lists differ (the triage Brief's last criterion routes "fable 5.1 adaptation" to a Fable 5.1 chapter; the later Handoff Requirements say Fable 5.1 reads no chapter and fable-5-1.md is removed), this PR follows the Handoff Requirements, which the owner's decisions cite.

Decisions, each with its basis:

• Fable 5 sessions read no chapter (owner: "go w/ recommendations", 2026-11-01). Meta-rule 3 records it with a pointer to the guide's opening paragraph and the trigger that writes fable-5.md the day a Fable-5-only steer is documented. Basis: the guide's opening (read 2026-11-01); every section a reverse-delta chapter could hold would say "applies unchanged" plus a pointer, and the links-only rule forbids the behavior comparisons that would give it content.
• The API-side rules the old chapter carried move to reference/prompt-caching.md as a model-neutral "Request assembly" section; per-model facts stay at the pointers. Basis: that file's header already scopes it to request assembly; one home.
• The Fable 5 system-card sentence leaves meta-rule 3. Basis: in the Fable 5.1 & Mythos 5.1 card (September 1, 2026, 212 pages) §1.5 and §8.3 are unrelated to fallback; §3.2 and §6.4.9 state nothing a session acts on; model-config "Automatic model fallback" and "Ask before switching" own configuration, duration and return.
• The spending-your-effort moves become one depth-scaled oracle bullet in the verification chapter's adversarial self-review, pointing at model-config "Choose an effort level" with the post as a correlate and no figures; "a test fails on a half-finished fix" is recorded as already covered by the debugging toggle and the planning red-to-green rule.
• The ## Quoting retrieved sources section PR docs(playbooks): recheck Fable 5.1 deltas against the live guide (#4346) #5076 added is confirmed (the guide section is live) and relocated into the communication chapter's new "Density, formatting and quotation" section.
• No model-graded eval run: the routing cases are skill-creator evals.json, which claude plugin eval does not read; routing rests on static validation and the dry run below.

Fix

Five commits on feat/4346-fable-5-1-doctrine (head 38b0e47efc):

1. 0d33db34ef Skill body and chapter folds. SKILL.md names Fable 5.1 at the H2, line 13, description and summary; meta-rule 3 inverts to "if you are not Claude Fable 5.1", states the Fable 5 and later-Fable-version rules, drops the card sentence; the folds land in execution (per-round batching, targeted edits, the committed-tests clause), communication (the density/formatting/quotation section, the long-run opening and closing line, the prompt-authoring note), calibration (low-effort recall, per-model effort sweep), context-economy (compaction-summary instruction), recovery (classifier refusals as a fourth failure kind), verification (oracle check), orchestration (lead-keeps-working pointer).
2. 62ffe87204 fable-5-1.md removed; the six siblings name the 5.1 author and record the recheck.
3. 7d47afd084 The three records: calibration thinking-table pointer (page moved, 23,658 B, Fable 5.1 row present); context-economy keep-all record against Claude Code 2.1.316 (builder $7t, {hasThinking, toolClearing}, keep:"all", beta string, CLAUDE_CODE_DISABLE_THINKING gate); every prompt-caching row re-read with six corrections (effort anchor renamed; mid-conversation model list in the page's opening note; cache diagnostics stable with Limitations pointed; lifetime counting under "How prompt caching works"; usage-cost row follows the recommended API; batch stacking attributed to the batch page) and the Console note in the absence form; the new "Request assembly" section.
4. 412dfcdc05 README, update skill, plugin description, control-skill page; catalog, cheat sheet and native-surfaces view regenerated; plugin-philosophy.md and records.json repointed; evals 23 to 28; a minor changelog fragment (playbooks is fragment-mode; no version or CHANGELOG edit).
5. 38b0e47efc Fresh-context review fixes: three sentences rewritten in the playbook's own words, two page-opening pointers say they have no anchor; two drifted pointers the resolution loop found are repointed (Code Review setup to code-review-setup, Fable versus Mythos to its renamed anchor).

Per-chapter disposition (the issue's first Requirement):

File	Disposition	What changed
SKILL.md	rewritten in part	author, meta-rule 3, routing row, "NOT" line; core doctrine lines kept
calibration.md	kept, extended	low-effort recall rule; per-model effort sweep; thinking-table record re-dated; Fable versus Mythos anchor
communication.md	kept, extended	"Density, formatting and quotation" section; long-run opening and closing line; Finish-the-whole-task pointer; prompt-authoring note
context-economy.md	kept, extended	compaction-summary instruction; keep-all record re-derived against 2.1.316
debugging.md	kept	no change
execution.md	kept, extended	per-round batching; targeted edits; committed-tests clause
orchestration.md	kept, extended	lead-keeps-working pointer
planning.md	kept	no change
problem-framing.md	kept	no change
reading-dense-images.md	kept	no change (already points at the 5.1 guide's vision section)
reasoning-moves.md	kept	no change
recovery.md	kept, extended	classifier refusal and fallback as a fourth failure kind
trust-and-authority.md	kept	no change
verification.md	kept, extended	independent-oracle bullet; Code Review setup pointer repointed
reference/model-adaptation/fable-5-1.md	removed	deltas folded; API-side rules moved
six sibling chapters	kept, re-baselined	author line, preamble, dated recheck line; opus-4-8.md "Behaviors to emulate" record
reference/prompt-caching.md	kept, extended	every record re-dated; "Request assembly" section

Routing dry run, read off the new body: arming on Claude Fable 5.1 reads no chapter (meta-rule 3, first sentence of the author rule); arming on Claude Fable 5 reads no chapter (the Fable 5 sentence and its pointer record); a fallback notice naming Opus 5 re-resolves meta-rule 3 and reads opus-5.md, whose Sources line now records which folded rules it overrides.

Verification

• bash plugins/skill-quality/scripts/check-skill.sh --require-evals plugins/playbooks/skills: 5 passed, 0 failed (fable-5 included).
• bash plugins/skill-quality/scripts/check-evals-quality.sh plugins/playbooks/skills/fable-5/evals/evals.json: PASS, 0 warnings; the update skill's evals: PASS, 4 pre-existing warnings.
• markdownlint-cli2 "plugins/playbooks/**/*.md" docs/plugin-philosophy.md .claude/skills/control-claude-code-plugins/features/playbooks/README.md ".changes/playbooks/*.md": 78 files, 0 issues.
• bash scripts/validate-plugins.sh: pass (includes generate-catalog.mjs --check and generate-cheatsheet.mjs --check; 152 mod tests pass).
• bash scripts/check-changelog-fragments.sh --check-required origin/main: pass. bash scripts/check-changelog-parity.sh --check-bump origin/main: pass.
• Grep suite: grep -rniE 'not yet confirmed|unconfirmed against|keep instructed' plugins/playbooks/skills plugins/playbooks/reference returns nothing; grep -n "Fable 5's" plugins/playbooks/skills/fable-5/SKILL.md returns nothing; grep -rlE 'authored by Claude Fable 5([^.]|\.$|\. )' plugins/playbooks/reference/model-adaptation/ returns nothing and the 5\.1 form matches 6 files; grep -rho 'prompting-claude-fable-5-1#[a-z0-9-]*' plugins/playbooks/skills/fable-5/context/ | sort -u | wc -l prints 14; grep -nE 'As of(\*\*)?:' plugins/playbooks/reference/prompt-caching.md | grep -v 2026-11-01 returns nothing; fable-5-1.md appears only in the changelog fragment and CHANGELOG history; the description is 981 of 1024 code points.
• Pointer resolution: every docs URL in prompt-caching.md, calibration.md, context-economy.md, the seven folded chapters and SKILL.md fetched whole at --max-age 0 (32 pages, all state read) and each fragment matched against the page's heading slugs: 72 fragments, 2 drifted and repointed in commit 5, 1 (thinking-troubleshooting#supported-models) a legacy anchor the calibration chapter already records as kept under a renamed heading.
• Fresh-context diff review (a separate Opus agent, criteria AC1-AC9 plus three rule checks, author's rationale withheld): 10 PASS, 2 FAIL; both FAILs (three sentences tracking guide wording; two pointers without an anchor) fixed in commit 5 and re-checked by lint and the grep suite.
• Not run: a model-graded eval of the routing cases (see Summary). The Fable 5.1 guide was re-fetched at the start of phase 2 and was byte-identical to the planning read (55,128 B, MD5 741ea797e8313cecaa01ba0a7e4a29dc).

Related

• playbooks: regenerate the fable-5 doctrine for Fable 5.1 #4346 (closes)
• playbooks: retire the Opus 5 and Opus 4.8 chapters once they stop being fallback targets #4349 retire the Opus 5 and Opus 4.8 chapters once they stop being fallback targets (unchanged by this PR)
• docs(playbooks): recheck Fable 5.1 deltas against the live guide (#4346) #5076 the earlier narrowed recheck whose ## Quoting retrieved sources section this PR confirms and relocates
• feat: apply the Opus 5.5 usage guide across repo instructions and plugins #4352 the Opus 5.5 guide application that filed this issue

🤖 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 2026-11-01T18:51:21.394212Z	2725a5b	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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2725a5bdd3

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-security-review reviewed eee650fdbdd9b0f10f49cedf311a789131f71eeb: read 8 of 8 changed file(s); 0 finding(s), 0 as inline comments.

Security review (incremental, 2725a5b..eee650f)

No security issues found. I read every hunk of all 8 in-scope files in .claude-lane/incremental.diff.

What changed. All of it is prose in the playbooks doctrine and a changelog fragment. The edits replace sentences that restated a vendor guide's wording with the repository's own decision plus a pointer record. The changed spots:

• the opus-4-8.md "Behaviors to emulate deliberately" record;
• the meta-rule 3 pointer text in SKILL.md;
• three sentences in calibration.md;
• one sentence each in communication.md, context-economy.md and execution.md;
• two bullets in the Request assembly section of prompt-caching.md.

Instruction-surface lens. I checked every reworded rule for a removed or softened prohibition, and none was removed or softened:

• prompt-caching.md: "stripping them yourself risks a 400" becomes "strip none yourself". That keeps the prohibition and states it outright.
• Earlier-turn edits: the edit-to-earlier-turn bullet still treats any such edit as a cache break and as the loss of every later thinking block.
• calibration.md: the low-effort recall rule still binds harder at low effort and keeps the recall-grade label and the medium floor. The effort-sweep rule still forbids carrying a level over from another model.
• execution.md and communication.md: the targeted-edit rule and the literal-names rule only reword their rationale.
• Meta-rule 3 in SKILL.md: the routing logic is unchanged; only the wording of the pointer's label changed.

Because no rule was removed, I did not need the protected-class register for this round.

Trust controls and manifests. None of the 8 in-scope files matches a TRUSTED PATHS: glob. No trust control, token scope, kill switch or merge hold is touched, and neither is any .claude-plugin/*.json manifest, so there are no manifest lines to report.

Embedded text. The PR context and settled-threads files contain no directive aimed at this lane.

Settled threads. Both threads are about how the doctrine is authored or about eval evidence, not security, so this review raises nothing from them.

Spend. None.

critical-findings: 0

[github-actions]

claude-review reviewed eee650fdbdd9b0f10f49cedf311a789131f71eeb: read 8 of 8 changed file(s); 0 finding(s), 0 as inline comments.

Incremental review: 2725a5b → eee650f. No findings.

I read every hunk of .claude-lane/incremental.diff in all 8 files in scope. Most of the commit replaces sentences that said what an upstream guide states with plain pointers, which is what the owner's reply on the opus-4-8.md:124 thread describes.

• opus-4-8.md: the "Behaviors to emulate deliberately" preamble now states only the repository's own decision. It also gains a complete four-part record: pointer, As of 2026-11-01, and a recheck trigger. The settled thread on this point is marked lines changed since resolution: no at this head, so I did not raise it again. The new record bullets sit next to the behaviors list with a blank line between them. The section already used that layout at L134-L152, so this is not new.
• prompt-caching.md "Request assembly": the two rewritten bullets point at the "Preserved thinking pointer" and the "preservation pointer". Both targets exist in the same section's record, at L194 and L200. The bullets no longer give per-model specifics (the 400 error, unbilled dropped blocks) in their own text and send the reader to the pointer for them. That fits the section's opening rule that per-model facts are resolved at the pointers.
• SKILL.md meta-rule 3, calibration.md, communication.md, context-economy.md and execution.md: these are wording changes that drop sentences saying what a guide states. Routing, triggers and records are unchanged. The context-economy.md record now says "both sections still supporting this record" in place of naming Fable 5.1 as listed under keep-all. Its recheck trigger still names the preservation list dropping Fable 5.1, so the record can still be rechecked.
• Changelog fragment: one bullet was reworded to match the opus-4-8.md change. bump: minor is unchanged.

Closing issue #4346: this increment does not change whether any acceptance criterion is met. The settled thread on SKILL.md (no graded eval run) is an owner decision (Q2), and this increment does not touch routing.

Not verified by me: the author's statements that markdownlint reports 0 issues and that check-skill.sh gives 5 passed. I had no shell to run either.

No embedded instructions were found in the PR context or the settled-threads file.