feat: Claude Opus 5.5 support — catalog, request compat, prompt preset, recommended default at max

[code-yeongyu]

Summary

Claude Opus 5.5 (claude-opus-5-5, released 2026-09-22) becomes a first-class model and the recommended Opus.

This is a fix release, not only a feature. 2026.9.22-3 already exposes claude-opus-5-5 — the release-bot catalog regen pulled it from models.dev — but with a partial row: no off: null, no supportsDisabledThinking, no per-message effort, no refusal fallback, the inference-profile-only bare id on Bedrock, and no prompt preset. A user on 22-3 who selects Opus 5.5 and turns thinking off gets a 400, and otherwise runs on the generic system prompt. See the tarball verification comment.

• Catalog (1ca1a143ae): anthropic / amazon-bedrock (global/us/eu/jp/au profiles; bare id profile-only like Opus 5) / openrouter / vercel-ai-gateway rows via one strict generate-models run. 1M ctx, 128k out, $4/$20 with $0.20 cache reads, xhigh+max, adaptive-only (off: null, supportsDisabledThinking: false), per-message effort, allowedFallbackModels: [claude-opus-5] (live Models API allowlist is [opus-4-8, opus-5]; the generator keeps only managed-effort targets).
• Runtime compat (80e450b3f5): thinking.type: disabled and forced tool_choice (any / named) are never sent to 5.5, on Messages and Bedrock, including rows with no generated compat (models.json / gateways). A thinking-off turn pins effort low.
• Prompt preset (5cdb43d11c): claude-opus-5-5.ts — the dieted Opus 5 core plus the coding-agent deltas from Anthropic's Opus 5.5 guide (the four text-only turn endings that count as stopping early while work is owed and the two legitimate stops; explore-before-acting on loosely specified tasks; delegation framed around elapsed time). Matcher resolves opus-5-5 / opus-5.5 before the generic opus-5 substring. o200k probe: claude-opus-5 1899 → claude-opus-5-5 2002 tokens.
• Defaults (fc3847bf93): RECOMMENDED_DEFAULT_MODELS carries ["claude-opus-5-5", "max"] where ["claude-opus-5", "xhigh"] was; fable-5-1 / fable-5 ladders become opus-5-5:max → opus-5:max → opus-4-8:max → opus-4-6:max; new claude-opus-5-5 ladder opus-5:max → opus-4-8:max → opus-4-6:max. Opus 5 stays selectable. Filed as a Breaking Change in packages/coding-agent/CHANGELOG.md so a reader asking why their default moved finds it.

Product decision (LO, this session, verbatim): "opus 5 사용하도록 되어있는거 ... 다 오푸스 5.5 로", "xhigh 쓰던 애들 다 max 리즈닝으로", "기본 추천 모델에 있어야함 ... 추천리즈닝 맥스".

xhigh → max migration (shipped defaults)

git grep xhigh ∩ opus-5 (non-5.5) over packages/*/src + packages/*/docs, tests/trackers/catalog excluded — before 8, after 6. The two that moved were the shipped defaults: recommended-models/index.ts:20 (["claude-opus-5", "xhigh"] → ["claude-opus-5-5", "max"]) and docs/settings.md:86. The 6 that remain are Cursor's own catalog facts (cursor-variant-aliases.json claude-opus-5-thinking-xhigh*, cursor/model-capabilities.ts Opus 5 tiers) — Cursor's model list, not a senpi default; Cursor does not serve 5.5 yet. No shipped default carries claude-opus-5:xhigh after this PR. User-config sites (defaultModel, recommendedModels, retry.fallbackChains) are untouched and get the migration note in the changelog.

Docs naming Opus 5 as the pick

git grep claude-opus-5|Opus 5 (non-5.5) over *.md, changelogs/trackers/pico excluded — before 7, after 1. The remaining line is prompt-preset/AGENTS.md:25, the file-tree row describing claude-opus-5.ts itself (Opus 5 stays supported).

Live API evidence (model id as accepted: claude-opus-5-5)

Direct api.anthropic.com with the org key returned billing 400 on every request, so the raw shapes went through the Anthropic-Messages gateway this host uses; the negatives come from the live Models API instead.

request	result
GET /v1/models/claude-opus-5-5 (server-side-fallback beta)	200: thinking.types.enabled.supported: false, adaptive: true; effort low..max; allowed_fallback_models: ["claude-opus-4-8","claude-opus-5"]
output_config.effort: low, no thinking field	200 end_turn → "pong", thinking_tokens 0
tools + tool_choice: auto	200 tool_use → get_weather({"city":"Paris"})
thinking: adaptive + effort: max	200 end_turn → thinking block + "pong", thinking_tokens 49
thinking: {type: disabled}	refused (gateway 403 "Request not allowed"; documented Anthropic 400 "thinking.type.disabled" is not supported for this model)
tool_choice: {type: any}	refused (gateway 403; documented Anthropic 400 tool_choice: type "tool" and "any" are not supported)

Headless senpi from this branch (node + tsx, sandbox agent dir, provider re-pointed at the gateway):

• --model anthropic/claude-opus-5-5 --thinking max "reply with exactly: pong" → exit 0, pong
• --thinking off same prompt → exit 0, pong (no thinking.type.disabled 400)
• --thinking low "use the read tool ..." → exit 0, read tool_use in the session, correct answer

RED → GREEN

• packages/ai: 12 failing (getModel undefined) → 144 passing across 8 suites.
• preset: 12 failing (ids resolved to claude-opus-5) → 102 passing (incl. Opus 5 exclusion, brand-identity).
• defaults: 84 passing across 8 pinned suites.
• bun run check exit 0 and scripts/check-pr-changelog.mjs --base origin/main PASS, both run from the worktree's canonical (symlink-resolved) path rather than through a symlinked parent, which on this machine produces false Type X is not assignable to type X errors from double-resolved package paths. Re-run them from the real path, not through a ~/local-workspaces symlink: behind the symlink TypeScript resolves every module twice and emits a wall of Type 'X' is not assignable to type 'X' that is a path artifact, not a defect.

Verifying the published artifact

.omo/evidence/.../verify-published-opus-5-5.mjs (kept with this run's evidence) checks a published version end to end: preconditions first — dist.tarball resolves, curl -I returns 200, size above 1 MB, tar -tzf entry count above 1000, nested @earendil-works/pi-ai present — and only then the row assertions. npm metadata leads the blob by minutes, and against a 4 KB JSON error body every "X is absent" assertion passes vacuously.

Control run against the known-bad 2026.9.22-3: preconditions 5/5 PASS, row assertions 6/7 FAIL, exit 1. The assertions can fail, so a pass on 22-4 will mean something.

Base: 9fff11a3c (post-22-3 stamp). Ships as senpi 2026.9.22-4 once the release lane clears.

[code-yeongyu]

What 2026.9.22-3 already shipped for Opus 5.5 (verified against the published tarball)

The 22-3 release-bot catalog regen picked claude-opus-5-5 up from models.dev on its own, so the published package already exposes the model — with an incomplete row. Verified by downloading senpi-2026.9.22-3.tgz (HTTP 200, 44,210,616 bytes, 27,066 entries — presence precondition before reading) and reading node_modules/@earendil-works/pi-ai/dist/providers/data/:

"claude-opus-5-5": {
  "thinkingLevelMap": { "xhigh": "xhigh", "max": "max" },
  "compat": { "forceAdaptiveThinking": true, "supportsTemperature": false, "supportsStrictTools": true }
}

Missing there, added here: thinkingLevelMap.off: null and compat.supportsDisabledThinking: false (on 22-3 off is selectable, so a thinking-off turn sends thinking.type: "disabled" and the model 400s), compat.supportsMidConvoEffort: true (no per-message effort), and compat.allowedFallbackModels (no server-side refusal fallback). amazon-bedrock.json on 22-3 also carries the bare anthropic.claude-opus-5-5 id, which is inference-profile-only — this PR emits only the five profiles.

So 22-4 is not only "add Opus 5.5": it repairs a row users can already select on 22-3.