Support CC 2.1.170 (Fable 5 release)

[skrabe]

Summary

Support for CC 2.1.170 (the Fable 5 release).

• data/prompts/prompts-2.1.170.json: extracted with our extractor, upstream identifierMaps adopted for 306 shared prompts. Corpus is structurally identical to 2.1.169 (389 named, 0 new/removed); content changed only in 12 data-* API-doc prompts picking up Fable 5 / Mythos 5 (new tier above Opus, $10/$50 per MTok, thinking:{type:"disabled"} → 400). We name 389 vs Piebald's 346.
• autoModeClassifierModel patch: new match method for the 2.1.170 resolver shape — Anthropic added a Fable branch that falls back to the default Opus model when the main model is Fable/Mythos. Regression test added with the real minified fixture.
• Extractor: named the three partial-map code-review slots (VERIFY_VOTE_DEFINITIONS, RECALL_BIASED_RUBRIC, SWEEP_MISS_CATEGORIES) whose pristine stubs rendered ${UNKNOWN_N}.
• README: target version 2.1.170, refreshed counts.

Test plan

• pnpm lint, pnpm test (336 passing incl. 4 new patch tests), pnpm build
• --apply on a pristine 2.1.170 native install: clean, four zeros, auditMisbinds 0, fresh extraction matches committed JSON
• Smoke: patched binary returns READY

[BenIsLegit]

Hey @skrabe sorry for going off topic from the PR, but I mentioned you in this issue incase you want to take over the npm package
BenIsLegit#3

[dividedby]

Heads-up from downstream, evidence only — not asking you to block or rework this PR on it. While gating this branch on our side we ran a duplicate-id scan across the prompts JSONs and found a chronic carryover class that's present in this PR's data/prompts/prompts-2.1.170.json (verified at head e096008). Since you said in the #6 review you'd rather root-cause this class than lock in snapshots, here's the full picture so one pass can cover all of it.

1. Three byte-identical duplicate pairs, carried for ~30 releases. Each id appears twice with entries that hash identically (sha256 over the sorted JSON entry):

id	carryover version stamp
system-prompt-memory-description-of-user-details	2.1.94
system-prompt-description-part-of-memory-instructions	2.1.69
system-prompt-claude-in-chrome-browser-automation	2.1.20

The same three pairs are already in prompts-2.1.141.json on main (which also has a fourth, system-prompt-version-banner-detailed ×2, that later resolved), and they persist unchanged through every JSON we checked up to this PR's 2.1.170 — a dedup miss in carryover that re-propagates each bump.

2. The same class compounded: system-reminder-cross-session-peer-message-authority-warning ×4. Trajectory on main: 0 (2.1.165) → 1 (2.1.167) → 1 (2.1.168) → 4 (2.1.169), and still 4 in this PR. The four entries are: one live 4-piece template stamped 2.1.169 (identifiers: [0,1,2]) plus three byte-identical stale copies of the superseded single-piece 2.1.167 literal. So this one is the dedup miss stacked on a fuzzy-carryover miss — the old literal shape wasn't matched to the new template shape, and then the stale copies multiplied.

3. Why we're flagging it rather than just shrugging: the ×4 case has two divergent shapes (different piece-sets/identifierMaps) under one id, which makes override binding ambiguous — the same mis-bind/orphan surface auditMisbinds exists to guard. And none of the current gates see it: at this branch's head, tools/ and skills/showtime/driver.mjs have no duplicate-id check (grep for "duplicat" comes back empty), so the Four-zeros + auditMisbinds=0 run stays green while the class recurs every bump.

Reproduction recipe (any prompts JSON):

import json, hashlib, collections
d = json.load(open("data/prompts/prompts-2.1.170.json"))
c = collections.Counter(p["id"] for p in d["prompts"])
for i, n in c.items():
    if n > 1:
        entries = [p for p in d["prompts"] if p["id"] == i]
        hashes = {hashlib.sha256(json.dumps(e, sort_keys=True).encode()).hexdigest()[:12] for e in entries}
        print(i, "x", n, "distinct-shapes:", len(hashes))

Cost on your side if you pick it up: a dedup/supersede pass in the extractor's carryover (and optionally a dup-id assert in the gate so it can't silently return). Entirely your call on whether/when — the fix is yours; we just had the multi-version diff view handy and didn't want the evidence to go stale after this merges. Happy to share the full per-version scan if useful.

[dividedby]

Downstream feature ask — nothing for this PR to do; posting here because issues are disabled. (Separate topic from the dup-id evidence above; keeping pings one-at-a-time.)

Ask: make the escape-or-not decision mechanical for override authors

Whether a ${VAR} in an override needs escaping depends on the prompt's delimiter in cli.js — backtick template literals interpolate live (literal examples must be escaped), quoted strings are inert (escaping them corrupts the render via the patcher's backslash-doubling). Your patcher already reads the delimiter at apply time (systemPrompts.ts, the content[matchIndex - 1] check), but nothing surfaces it to an author before apply + review — the rule is folklore. Evidence it bites:

• We shipped 2 wrong escapes against the quote-delimited data-anthropic-cli / skill-cowork-plugin-authoring-mcp-discovery in Align opus-4-7 overrides to upstream vocab (CC 2.1.169) + escape literal examples lobotomized-claude-code#5; you caught and reverted them by manual review (cb08b73).
• Fresh, from gating this PR pair this week: our static orphan scan flagged JSON as an orphan variable — it's actually ${JSON.stringify(...)} JS-literal examples inside the workflow-script-* / skill-design-sync-* overrides (both fable-5 and opus-4-8 sets). Without delimiter context, neither authors nor tooling can classify those mechanically.
• prompts-2.1.170.json carries no delimiter context (per-prompt fields today: name/id/description/pieces/identifiers/identifierMap/version).

Two possible shapes — your architecture call

1. Extractor emits a per-prompt delimiter field (backtick | quote) in the prompts JSON. The data exists at extraction time. Honest cost: a schema addition that rides every future extraction/regeneration, and a one-time churn for anything diffing prompts JSONs.
2. An auditMisbinds-style escape audit: flags escaped placeholders under a quote delimiter and unescaped literal ${...} examples under a backtick delimiter, per override. Honest cost: zero until run, one more tool surface to maintain; catches mistakes at audit time rather than guiding authors at writing time.

Happy to prepare a draft PR for whichever you'd pick — or neither; if you'd rather keep the rule in review, that's a fine answer and we'll keep validating on our side. No urgency, nothing here blocks 2.1.170.