lidge-jun · Sep 28, 2026 · Sep 28, 2026 · Sep 28, 2026 · Sep 28, 2026 · Sep 28, 2026
@@ -0,0 +1,123 @@
+# 260929 Claude Sonnet 5.5 catalog, pricing and adapter contract
+
+## Problem
+
+Anthropic released Claude Sonnet 5.5 (`claude-sonnet-5-5`) on 2026-09-28. Live Anthropic discovery on
+the running proxy already lists `anthropic/claude-sonnet-5-5` (1M context) but with no
+`max_output_tokens`, `supports_reasoning: false`, no effort ladder and no price, because nothing in
+the static catalog knows the id. Every other provider that carries `claude-sonnet-5` has no row at all.
+
+The model also changes the request contract. Sonnet 5 accepts `thinking: {type: "disabled"}`, and the
+Anthropic adapter sends exactly that for reasoning `none` on every Sonnet >= 5.0. Sonnet 5.5 rejects it
+with a 400 and adds `thinking: {type: "between_tools"}` as its lowest setting. It also rejects forced
+`tool_choice` (`any`/`tool`) like Opus 5.5, and non-default temperature/top_p/top_k.
+
+## Evidence (collected 2026-09-29)
+
+| Source | Surface | Facts |
+|---|---|---|
+| platform.claude.com/docs/en/about-claude/pricing | Aside exec | Sonnet 5.5: $2 in, $2.50 5m write, $4 1h write, $0.20 hit (standard 0.1x), $10 out; batch $1/$5; no fast mode; no long-context premium; `inference_geo: us` 1.1x |
+| platform.claude.com/docs/en/models/overview | Aside exec | id and alias `claude-sonnet-5-5`; Bedrock `anthropic.claude-sonnet-5-5`; Google Cloud / Foundry `claude-sonnet-5-5`; 1M context; 128K output; adaptive thinking; default effort high |
+| .../models/sonnet-5-5/whats-new-sonnet-5-5 and migration-guide | Aside exec | `thinking.type` accepts only `adaptive` and `between_tools`; `disabled` and `enabled` 400; `between_tools` 400 at xhigh/max; forced tool choice 400; non-default sampling params 400 |
+| .../build-with-claude/effort | Aside exec | low, medium, high, xhigh, max; default high |
+| .../build-with-claude/fast-mode | Aside exec, kimi | fast mode is Opus-only; no Sonnet 5.5 fast anywhere |
+| models.dev/api.json | curl, kimi | anthropic, azure, vertex, kilo, openrouter, vercel, bedrock `global.anthropic.claude-sonnet-5-5`; all 1M / 128K, 2 / 10 / 0.2 / 2.5 |
+| openrouter.ai/api/v1/models | curl | `anthropic/claude-sonnet-5.5` 1M / 128K, 2 / 10 / 0.2 / 2.5 |
+| ai-gateway.vercel.sh/v1/models | curl | `anthropic/claude-sonnet-5.5` 1M / 128K, 2 / 10 / 0.2 / 2.5; no -fast |
+| api.kilo.ai gateway models | kimi | `anthropic/claude-sonnet-5.5` 2 / 10 / 0.2 / 2.5 |
+| api.venice.ai/api/v1/models | kimi | `claude-sonnet-5-5` 1M / 128K, 3.75 / 18.75 / 0.375 / 4.6875 |
+| cursor.com/docs/models/claude-sonnet-5-5 | kimi | id `claude-sonnet-5-5`, 200K default / 1M max, thinking supported, 2 / 2.5 / 0.2 / 10; no fast; not yet in the live Cursor roster |
+| github.blog changelog 2026-09-28, docs models-and-pricing | kimi | Copilot GA; $2 / $0.20 cached / $2.50 write / $10; 1M context; no API id published |
+| docs.devin.ai/desktop/models, running proxy devin/* | kimi | no Sonnet 5.5 |
+| kiro.dev/docs/models, changelog | kimi | no Sonnet 5.5 |
+| api.commandcode.ai/provider/v1/models | kimi | no Sonnet 5.5 |
+| opencode.ai/zen/v1/models, zen/go | kimi | no Sonnet 5.5 (go has no Claude at all) |
+| api.opper.ai/v3/models | kimi | no Sonnet 5.5 |
+| zenmux.ai/api/v1/models, Cloudflare catalog | kimi | no Sonnet 5.5 |
+| installed Claude Code 2.1.283 binary | strings | `sonnet` alias still resolves to `claude-sonnet-5`; no `claude-sonnet-5-5` string |
+
+Raw captures: `.tmp/sonnet55/` (scratch, not committed).
+
+## Provider classification
+
+| Provider | Decision | Id | Numbers |
+|---|---|---|---|
+| anthropic, anthropic-apikey | ADD seed, context, snapshot row, overlays (verified) | `claude-sonnet-5-5` | 2 / 10 / 0.2 / 2.5, 1M / 128K |
+| amazon-bedrock | ADD 6 snapshot rows; `global.` published, `anthropic.` from Anthropic docs, us/eu/jp/au preemptive | `*.anthropic.claude-sonnet-5-5` | base 2 / 10 / 0.2 / 2.5; regional 1.1x = 2.2 / 11 / 0.22 / 2.75 (same rule as Opus 5.5 rows) |
+| openrouter, vercel-ai-gateway, kilo | ADD snapshot rows (published) | `anthropic/claude-sonnet-5.5` | 2 / 10 / 0.2 / 2.5 |
+| venice | ADD snapshot row (published) | `claude-sonnet-5-5` | 3.75 / 18.75 / 0.375 / 4.6875 |
+| github-copilot | ADD snapshot row (published price; hyphen id convention of its Sonnet 5 / Opus 5.5 rows) | `claude-sonnet-5-5` | 2 / 10 / 0.2 / 2.5 |
+| cursor | ADD capability, effort tiers, picker family, price overlay (published price; roster shape preemptive) | `claude-sonnet-5-5` | 2 / 10 / 0.2 / 2.5; 1M window |
+| devin, devin-cli | ADD preemptive seed, context and derived overlays | `claude-sonnet-5-5` | 1M; Anthropic list price |
+| kiro | ADD preemptive model and 1M context (dot spelling like its other rows) | `claude-sonnet-5.5` | price via vendor fallback |
+| opencode-zen | ADD preemptive snapshot row | `claude-sonnet-5-5` | 2 / 10 / 0.2 / 2.5 |
+| zenmux | ADD preemptive snapshot row (its dot convention) | `anthropic/claude-sonnet-5.5` | 2 / 10 / 0.2 / 2.5 (its Sonnet 5 row shape) |
+| cloudflare-ai-gateway | ADD preemptive snapshot row (its hyphen convention) | `anthropic/claude-sonnet-5-5` | 2 / 10 / 0.2 / 2.5 |
+| Claude Desktop picker suggestions | ADD `claude-sonnet-5-5` (suggestion list only) | | |
+| Command Code, Opper | NO static site: neither seeds `claude-sonnet-5` today; live discovery owns the roster and price resolves through the Anthropic vendor row | | |
+| opencode-go | NO: carries no Claude model | | |
+| any fast tier | NO: Anthropic has no Sonnet fast mode | | |
+| Claude Code native tier map, web-search / vision sidecar defaults | UNCHANGED: Claude Code 2.1.283 still sends `claude-sonnet-5`; the sidecars send `thinking: disabled`, which Sonnet 5.5 rejects | | |
+
+## Diff-level plan (wp2)
+
+1. `src/adapters/anthropic.ts`
+   - `supportsExplicitThinkingDisable`: Sonnet 5.0 <= v < 5.5 only (5.5 rejects `disabled`).
+   - New `usesBetweenToolsFloor`: Sonnet >= 5.5. Reasoning `none` sends `thinking: {type: "between_tools"}`
+     with no effort (the API default high is inside the accepted low..high range) and drops temperature/top_p.
+   - `rejectsForcedToolChoice`: Opus 5.5 and Sonnet >= 5.5.
+   - `rejectsSamplingParameters`: Sonnet >= 5.5 drops temperature/top_p on every path.
+2. `src/usage/expected-prices.ts`: `CLAUDE_SONNET_55` = 2 / 10 / 0.2 / 2.5; overlays anthropic,
+   anthropic-apikey (verified), cursor (verified, Cursor model page), devin and devin-cli
+   (verified-derived, preemptive).
+3. `scripts/model-metadata.source.json` rows listed above, cloned from each provider's claude-sonnet-5
+   row with the new id/name/cost; regenerate `src/generated/model-metadata.ts`.
+4. `src/providers/registry/model-seeds.ts`: `claude-sonnet-5-5` before `claude-sonnet-5` in
+   `ANTHROPIC_MODELS`, 1M in the context map, provenance comment.
+5. Devin: `entries-core.ts` seed list and `DEVIN_MODEL_CONTEXT_WINDOWS`.
+6. Kiro: `KIRO_MODELS` + `KIRO_MODEL_CONTEXT_WINDOWS` (`claude-sonnet-5.5`).
+7. Cursor: `CURSOR_CAPABILITIES["claude-sonnet-5-5"]` shaped like Opus 5.5 (flat effort ids, regular
+   FULL ladder, no fast/thinking variant until the live roster is measured); effort-map tier row;
+   `models-capabilities.ts` picker family regex.
+8. `src/claude/intercept/model-bindings.ts` suggestion list.
+9. `structure/providers/chat-compat.md`: Sonnet 5.5 thinking floor, forced tool choice and sampling.
+10. Tests next to existing ones, within ratchet caps: adapter wire shape (none -> between_tools, forced
+    choice -> auto, sampling stripped, Sonnet 5 unchanged), usage-cost resolver across providers,
+    kiro/cursor rosters if pinned.
+
+## Verification
+
+- `bun run generate:model-metadata`, model-metadata sync test
+- focused: tests/adapters/anthropic, tests/usage, tests/providers/cursor, tests/providers/kiro,
+  devin tests, registry parity; `bun run test:changed`
+- `bun run typecheck`, file-size ratchet + test-layout, `bun run structure:check`, `bun run privacy:scan`
+- Live: adapter-built requests against api.anthropic.com for Sonnet 5.5 with reasoning none, medium,
+  required tool choice; expect 200 after the fix (token read in memory, never printed).
+- Fresh-process resolver probe for every provider row.
+
+## Bounds
+
+Write scope: files above plus this unit. PR to `dev`, exact-head CI, merge, then release via
+`scripts/release.ts` (user asked to deploy on 2026-09-29).
+
+
+## Plan after audit (020)
+
+- Item 5 is two files: `src/providers/registry/entries-core.ts` devin seed list and
+  `DEVIN_MODEL_CONTEXT_WINDOWS` in `src/adapters/devin/live-models.ts`. `DEVIN_STATIC_MODELS` stays
+  untouched; live discovery owns the roster.
+- Item 6: no `KIRO_NATIVE_EFFORT_FIELDS` entry. Kiro's `claude-sonnet-5` uses emulated effort and only
+  Opus has a measured native field.
+- Item 7: no `models-capabilities.ts` regex change, matching what Opus 5.5 actually shipped. The Cursor
+  row is regular-only until the id appears in the live GetUsableModels roster; that listing is the
+  trigger to re-shape it (thinking variant or flat effort ids, whichever the roster shows).
+- Sidecars: add a shared `anthropicThinkingOff(modelId)` in the adapter and use it in the web-search
+  and vision sidecars, so a sidecar pointed at Sonnet 5.5 sends `between_tools` instead of a
+  rejected `disabled`. Defaults stay `claude-sonnet-5`.
+- Tests: `kiro-adapter.test.ts` gets only the in-line roster element (3 lines of ratchet headroom);
+  new assertions go to sibling files registered in `scripts/test-layout/layout.json` and
+  `tests/fixtures/test-layout-expected.json`. Update the usage-cost surface loops, the cursor catalog
+  flat-wire case, the devin context pin and add anthropic wire-shape cases.
+- Residuals recorded, not fixed here: dotted Bedrock ids do not parse in `claudeFamilyVersion` (all
+  families), and the Messages-native passthrough forwards caller `thinking`/sampling verbatim.
+
@@ -0,0 +1,20 @@
+# 020 Audit (wp1): Sonnet 5.5 plan
+
+Auditor: kimi/kimi-for-coding (max), read-only, 2026-09-29. Verdict: NEAR-PASS. The four adapter
+predicates, the between_tools-without-effort choice and the provider classification were confirmed
+against source. Dispositions:
+
+| # | Finding | Disposition |
+|---|---|---|
+| 1 | `claudeFamilyVersion` matches only `^` or `/` before `claude-`, so dotted Bedrock ids (`us.anthropic.claude-sonnet-5-5`) never parse | RESIDUAL. Pre-existing for every family including Opus 5.5; widening the boundary changes wire shape for all Bedrock-style Claude ids and needs its own live evidence. Recorded in 030 |
+| 2 | Messages-native passthrough forwards caller `thinking` and sampling fields verbatim | RESIDUAL. The passthrough contract is caller-owned bytes; a Messages client addressing Sonnet 5.5 gets Anthropic's own 400. Recorded |
+| 3 | Web-search and vision sidecars hard-code `thinking: disabled`; a user who points them at Sonnet 5.5 gets a 400 | RESIDUAL. Defaults stay `claude-sonnet-5`. Recorded as an override hazard |
+| 4 | Kiro native effort field for `claude-sonnet-5.5` | DECIDED: no entry. Kiro's `claude-sonnet-5` uses emulated effort today and only Opus has a measured native field |
+| 5 | Cursor local picker family regex was promised for Opus 5.5 and never landed | DECIDED: skip for Sonnet 5.5 too; the live bundle table owns that prediction |
+| 6 | claude-cli and native Anthropic catalog spread `ANTHROPIC_MODELS` | ACCEPTED: user-visible roster addition, intended |
+| 7 | Devin context map lives in `src/adapters/devin/live-models.ts`; `DEVIN_STATIC_MODELS` stays untouched | FOLDED into wp2 diff |
+| 8 | Cursor row shape is regular-only while the docs say thinking is supported | FOLDED: follow-up trigger is the model appearing in the live GetUsableModels roster; shape then follows the measured ids |
+| 9 | `kiro-adapter.test.ts` has 3 lines of ratchet headroom | FOLDED: only the in-line roster element there; any new assertion goes in a sibling file registered in both layout maps |
+| 10 | Pinned tests: kiro roster, usage-cost surface loops, cursor catalog flat-wire case, devin context pin, anthropic reasoning wire cases | FOLDED into wp2 test list |
+| 11 | No docs-site locale diff needed (no Sonnet fast tier) | ACCEPTED |
+
@@ -0,0 +1,39 @@
+# 030 Done (wp2): Sonnet 5.5 catalog, pricing and adapter contract
+
+## Outcome
+
+- Snapshot rows (15), regenerated `src/generated/model-metadata.ts`: anthropic `claude-sonnet-5-5`;
+  Bedrock `anthropic.`, `global.` (2 / 10 / 0.2 / 2.5) and `us.`, `eu.`, `jp.`, `au.` (1.1x:
+  2.2 / 11 / 0.22 / 2.75); openrouter, vercel-ai-gateway, kilo `anthropic/claude-sonnet-5.5`; venice
+  `claude-sonnet-5-5` (3.75 / 18.75 / 0.375 / 4.6875); github-copilot `claude-sonnet-5-5`; preemptive
+  opencode-zen `claude-sonnet-5-5`, zenmux `anthropic/claude-sonnet-5.5`, cloudflare-ai-gateway
+  `anthropic/claude-sonnet-5-5`. All 1M / 128K.
+- Seeds: `ANTHROPIC_MODELS` and context map (also reaches claude-cli and the native Anthropic catalog),
+  Devin seed and context (preemptive), Kiro `claude-sonnet-5.5` (preemptive), Cursor capability and
+  effort tier (regular-only flat ids until the live roster lists it), Claude Desktop picker suggestion.
+- Price overlays: anthropic, anthropic-apikey, cursor (verified); devin, devin-cli (verified-derived,
+  preemptive). Venice, Kiro, Command Code and Opper resolve through the Anthropic vendor row, as Opus
+  5.5 already does.
+- Adapter: new `src/adapters/anthropic-model-contract.ts` owns the per-family rules. Sonnet 5.5
+  reasoning `none` sends `between_tools`; forced tool choice degrades to `auto`; temperature/top_p
+  are dropped. The web-search and vision sidecars use the same module for their thinking-off field.
+
+## Live evidence (2026-09-29, api.anthropic.com, OAuth, adapter-built bodies)
+
+| Request | Before | After |
+|---|---:|---:|
+| `claude-sonnet-5-5`, reasoning none | 400 (`thinking.type.disabled` not supported) | 200 (`between_tools`) |
+| `claude-sonnet-5-5`, temperature 0.2 | 400 (`temperature` is deprecated for this model) | 200 (stripped) |
+| `claude-sonnet-5-5`, tool choice required | 400 (`any` not supported) | 200 (`auto`) |
+| `claude-sonnet-5-5`, effort medium and max | 200 | 200 |
+| `claude-sonnet-5`, reasoning none (control) | 200 (`disabled`) | 200 (unchanged) |
+
+## Residuals
+
+- Dotted Bedrock ids (`us.anthropic.claude-sonnet-5-5`) do not parse as a Claude family, so the rules
+  above do not apply on that id shape; true for every family today.
+- Messages-native passthrough forwards the caller's `thinking` and sampling fields unchanged.
+- Rows marked preemptive (Devin, Kiro, opencode-zen, ZenMux, Cloudflare, Bedrock regional) are inert
+  until each provider lists the model; Kiro and Devin show the id in their static pickers, so a call can
+  fail upstream until they ship it.
+- The Cursor row is re-shaped once the live GetUsableModels roster lists `claude-sonnet-5-5`.