feat(claude): add Claude Fable 5.1 model (cherry-pick upstream #9078)

[NoahHendrickson]

Problem

Claude Code 2.1.257 shipped Claude Fable 5.1 (claude-fable-5-1), and the Claude provider's model catalog is hardcoded, so it needs a code change to appear.

Fix

Upstream landed pingdotgg#9078 (c17d02cff) right as the first version of this PR opened. This PR is now a clean cherry-pick of that commit onto custom — byte-identical to upstream, so the next daily sync merges without conflict. It brings:

• The claude-fable-5-1 catalog entry (effort incl. ultracode, 200k/1M context window defaulting to 1M), gated on Claude Code >= 2.1.257.
• The xhigh-effort allowlist entry, fable/fable-5.1/claude-fable-5.1 slug aliases in contracts, and the manifest update that marks Fable 5.1 current (and, per upstream's policy, demotes Fable 5 to legacy).
• Upstream's tests for gating, effort, manifest classification, and aliases.

The earlier revision of this PR hand-rolled the same catalog change plus a fork-specific manifest-URL retarget; review feedback correctly flagged that upstream had already shipped, so the hand-rolled version and the now-unnecessary fork customization were dropped wholesale in favor of the cherry-pick.

Verification

• vp test run on ModelManifest.test.ts, ProviderRegistry.test.ts, ClaudeAdapter.test.ts (131 passed) and packages/shared model.test.ts (7 passed).
• Server and contracts typechecks clean; fork guard suite passes (377).

Babysat by Claude Fable 5 in Claude Code.

🤖 Generated with Claude Code

[cursor]

Do not approve. Behavior looks correct and the fork URL change is well-structured, but this misses an obvious code-judo move and grows an already-tangled version-policy ladder.

Adding Claude Fable 5.1 currently means cloning a catalog entry, a min-version constant, a supports* helper, a formatter, a filter branch, an xhigh allowlist entry, and a sixth nested-ternary rung. Those are the same fact restated seven times. A newest-first version-gate table deletes the helpers, the formatters, the slug if-chain, and the ternary. The next model is then a row, not another copy of this PR.

The 1k-line rule does not strictly apply: ClaudeProvider.ts was already 1021 lines and is now 1072. This PR did not cross the threshold. The table would still shrink the file instead of adding another ~50 lines of clones.

The ModelManifest URL retarget (fence, yaml intent, watch, guard test) is the right permanent fork divergence. Leave that. The unfenced catalog/manifest slug adds are the right last-resort shape if upstream is expected to supersede them. That fork hygiene does not excuse growing the version-policy spaghetti in the same change.

  

_{Sent by Cursor Automation: Thermo nuke 4.6}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	−49 B (−0.4%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−1 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	−48 B (−0.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.5 KiB	−44 B (−0.1%)	66.4 KiB	✅
Codex	Live turn messages	10	9	−1 (−10.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+31 B (+0.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+7 B (+0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+24 B (+0.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	+44 B (+0.1%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: bf51489 · PR result: b182a32 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[NoahHendrickson]

Review

The server-side change is correct and I found no runtime bug in it. The problem is the premise: upstream already shipped this. pingdotgg/t3code merged c17d02cff feat(claude): add Claude Fable 5.1 model (#9078) at 19:20Z; this PR opened at 19:26Z. Upstream's ClaudeProvider.ts hunk is byte-identical to yours (same 2.1.257 gate, same option surface, same ternary chain, same xhigh allowlist). Upstream's main manifest already serves ["claude-fable-5-1", "claude-opus-5", "claude-sonnet-5"]. So the stated reason for the URL repoint ("upstream's currentModels list has never heard of it") was already false when the PR was opened.

I simulated the next daily sync with git merge-tree <PR head> upstream/main. Result:

• ClaudeProvider.ts, ClaudeAdapter.test.ts: merge cleanly to a single copy (identical hunks). Good.
• model-manifest.json and ModelManifest.test.ts: conflict. Upstream's commit also demotes claude-fable-5 to legacy; this PR keeps it current.
• ProviderRegistry.test.ts: auto-merges but ends up with two copies of both new tests, because you inserted them before "includes Claude Fable 5" and upstream inserted them after it.
• Upstream also added fable / fable-5.1 / claude-fable-5.1 aliases in packages/contracts/src/model.ts; this PR has none.

Recommendation

Close this PR (or strip it to nothing) and take Fable 5.1 from the next claude/sync-* PR. That gets the catalog entry, the Fable 5 demotion, and the contracts aliases with zero fork divergence. If you want it on custom before the sync lands, cherry-pick c17d02cff so the later sync is a no-op.

The model-manifest-from-fork repoint should be dropped along with its guard and customizations.yaml entry. It now protects no real case (upstream has added every Claude model same-day: Opus 5, Fable 5, Fable 5.1), and it costs a permanent Tier 4 fence plus manifest freshness that is only as good as the last human-merged sync PR.

If there is a deliberate fork policy that claude-fable-5 should stay current after upstream demotes it, that is a separate customization that needs its own customizations.yaml entry and guard, not a side effect of this PR. Inline comments below cover what to fix if you keep any part of this.

Verification

• Fetched upstream main and both live raw.githubusercontent.com manifests.
• git merge-tree --write-tree 025e526e3 upstream/main and inspected the resulting ProviderRegistry.test.ts, model-manifest.json, ModelManifest.test.ts.
• Read .fork/AGENTS.md fence/watch rules and the JSON-file precedents in customizations.yaml.

Reviewed by Claude Fable 5.1 in Claude Code.