🤖 feat: promote Claude Sonnet 5.1 as the sonnet model

[ThomasK33]

Summary

Prepares Xum for a hypothetical Claude Sonnet 5.1 drop: promotes Sonnet 5.1 (anthropic:claude-sonnet-5-1) to the SONNET known model, so the sonnet alias, the /sonnet command, tokenizer warming, and the first-time mux-gateway default model list all route to the new model. Sonnet 5 stays usable as the custom model string anthropic:claude-sonnet-5 and keeps its metadata entry and tokenizer approximation.

Stacked on #3988 (base branch mike/fable-5-1-drop), so this diff is Sonnet-only and does not touch the Fable 5.1 work.

Important

Merge gate: hold until Anthropic officially announces Sonnet 5.1, then verify the assumptions below against the announcement before merging. Merging early would break sonnet-alias sends (the API id would not exist yet).

Release-day checklist before merging:

1. Confirm the API id is exactly claude-sonnet-5-1.
2. Confirm pricing and envelope against the announcement (assumptions below).
3. Flip tests/ipc/streaming/sendMessage.reasoning.test.ts back to KNOWN_MODELS.SONNET.id (TODO(sonnet-5-1-release) marker) so live integration coverage exercises the promoted id.
4. If Sonnet 5.1 ships safeguard classifiers (unlike Sonnet 5), add a refusal-fallback chain + one-shot seed flag first (see Implementation).

Sonnet 5.1 is the most speculative of the three 5.1 prep PRs (Fable #3988, Opus, Sonnet): rumors point at Fable 5.1 with Opus 5.1 as likely and Sonnet 5.1 as possible. This model may never ship; if the announcement omits it, close this PR unmerged.

Background

Sonnet 5.1 is not announced; leak reports around the Fable 5.1 drop only possibly include it. This PR is prepared ahead of the drop (same play as #3988 for Fable 5.1 and #3750 for Opus 5), with assumptions to confirm on release day. All three are guesses extrapolated from Anthropic's current conventions and from Sonnet 5's published envelope — none is confirmed by any announcement:

• API id (guess, by convention): claude-sonnet-5-1 (dash form, dateless, matching the 5-generation convention of claude-sonnet-5 / claude-opus-5, the dash style of claude-haiku-4-5, and the assumed claude-fable-5-1 in 🤖 feat: promote Claude Fable 5.1 as the fable model #3988).
• Pricing (guess, no signal either way): unchanged from Sonnet 5's standard rate, $3/M input, $15/M output (cache write 1.25x input, cache read 0.1x input). Assumes no introductory discount like Sonnet 5's $2/$10 launch promo; if 5.1 launches with one, we list the standard rate anyway (same choice as the Sonnet 5 entry) so estimates stay correct once a promo ends.
• Envelope (guess, mirrors Sonnet 5): native 1M context, 128K max output, full 6-level effort ladder with native xhigh, adaptive thinking with thinking: {type: "disabled"} still accepted (so off stays available — Sonnet's ladder deliberately differs from Fable/Mythos, which reject disabled thinking), and no safeguard classifiers (so no default refusal-fallback chain; see below).

Implementation

Mirrors #3988's Fable promotion, adjusted for what the Sonnet 5 entry does today:

• knownModels.ts: SONNET -> claude-sonnet-5-1 (keeps the sonnet alias and tokenizer warming; tokenizer override stays on the Sonnet 4.5 approximation). The retired anthropic:claude-sonnet-5 id joins LEGACY_TOKENIZER_MODEL_OVERRIDES so exact-id lookup keeps its approximate tokenizer instead of warning and falling back to the generic per-provider tokenizer (the trap 🤖 feat: promote Claude Fable 5.1 as the fable model #3988 round 1 fixed for claude-fable-5/claude-opus-4-8).
• models-extra.ts: new claude-sonnet-5-1 pricing/limits entry; the Sonnet 5 entry is retained.
• models.ts: claude-sonnet-5-1 added to the native-1M patterns — the existing claude-sonnet-5 pattern only tolerates date suffixes, not -1, exactly the trap 🤖 feat: promote Claude Fable 5.1 as the fable model #3988 documents for Fable.
• No wire-format changes needed for thinking/effort: the wildcard matchers already cover the new id (anthropicSupportsNativeXhigh matches any claude-sonnet-5+; anthropicRejectsDisabledThinking correctly does NOT match Sonnet, which is Mythos-class-only) and native web-fetch id parsing handles the two-segment id. Tests now pin all of that rather than assuming.
• No fallback seed migration needed — deliberate divergence from 🤖 feat: promote Claude Fable 5.1 as the fable model #3988: DEFAULT_MODEL_FALLBACKS has no SONNET key (refusal fallback exists for Fable's safeguard classifiers; Sonnet 5 has none, and we assume 5.1 keeps none). With no default chain keyed by the moving id, there is no seed gap, so no defaultModelFallbacksSeededSonnet51 flag is introduced; 🤖 feat: promote Claude Fable 5.1 as the fable model #3988's defaultModelFallbacksSeededFable51 machinery and LEGACY_DEFAULT_MODEL_FALLBACKS are untouched, user-edited chains are never rewritten, and legacy chains stay byte-identical on disk. If the announcement reveals Sonnet 5.1 ships safeguard classifiers, this PR must grow the equivalent chain + one-shot flag before merge.
• First-time mux-gateway default models list moves anthropic/claude-sonnet-5 -> anthropic/claude-sonnet-5-1 (precedent: 🤖 feat: add support for Claude Opus 5 #3750 did the same for the Opus 4.8 -> 5 move). Existing configured model lists are untouched.
• Docs model table and built-in skill content regenerated.
• Live-API integration tests pinned: tests/ipc/streaming/sendMessage.reasoning.test.ts sends real Anthropic requests via KNOWN_MODELS.SONNET.id, which would 404 until the announcement (the merge-gate problem, but biting PR CI immediately). Those tests are pinned to the still-live anthropic:claude-sonnet-5 with a comment to flip back to KNOWN_MODELS.SONNET.id once Sonnet 5.1 is live — a release-day follow-up item.

Validation

• Behavioral coverage added for the new id: native-1M classification (including the gateway-prefixed form and the date-suffix-only pattern trap), display formatting (Sonnet 5.1), 6-level thinking policy with off preserved (pinning that Sonnet does NOT get the Mythos-class off-clamp), provider options (adaptive + summarized display, native xhigh/max effort, off -> thinking: {type: "disabled"}), no Xum-internal effort header, and native web-fetch support for the two-segment id.
• The alias flip was caught by (and updated in) the slash-command parser test, the task-tool alias resolution test, the mux-gateway first-time seed test, and the e2e slash-command scenario + mock router.
• Remaining claude-sonnet-5 fixtures were audited and stay valid as explicit custom model strings (story fixture, legacy-id assertions, retained metadata entry).

Risks

Low mechanically: additive registry/metadata changes with no config migration. The main user-visible effect is that the sonnet alias and /sonnet route to Sonnet 5.1 — until the model exists at the API, those sends would 404, which is why the merge gate above matters (and why it is stricter here: this model is not confirmed to exist at all). Existing Sonnet 5 selections keep working via the custom model string, and cost estimates for it keep resolving through the legacy tokenizer override. Stacked-base risk: if #3988 changes shape before merge, this branch must rebase onto it.

---

Generated with xum • Model: anthropic:claude-fable-5 • Thinking: xhigh

[ThomasK33]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. You're on a roll.

Reviewed commit: e01cbaf168

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".

[ThomasK33]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: da0aa90201

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".

[ThomasK33]

@codex review

Please take another look. Re: the round-2 P1 ("keep the sonnet alias on an available model"): the alias flip is the deliberate purpose of this prep PR and is guarded by the human-held merge gate in the PR description (HOLD until Anthropic officially announces Sonnet 5.1; close unmerged if it never ships) — the same accepted pattern as #3988 (Fable 5.1) and #3750 (Opus 5). Responded inline and resolved the thread; no code change made.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: da0aa90201

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".

[ThomasK33]

@codex review

Addressed the round-3 P2: the live-test flip-back to KNOWN_MODELS.SONNET.id is now an explicit release-day merge-gate item — TODO(sonnet-5-1-release) marker in sendMessage.reasoning.test.ts (4ca4353) plus a release-day checklist in the PR description. It cannot flip in this PR because the id does not exist at the API yet.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep it up!

Reviewed commit: 4ca4353ad4

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".