fix(provider): port upstream provider/model transform for Claude 5

[jack-champagne]

Why

Our base is sst/opencode v1.17.3 (2026-06-10). Sonnet 5 (2026-06-29) and Opus 5 (2026-07-24) both shipped after our fork point, and at our base their thinking plumbing is silently broken:

anthropicOpus47OrLater() = /opus-(\d+)[.-](\d+)…/   // requires BOTH major and minor

claude-opus-5 has no minor; claude-sonnet-5 isn't "opus". Neither matches → both fall through to the legacy thinking:{type:"enabled",budgetTokens:16000} path with only high/max tiers instead of the 5-tier type:"adaptive" path. Upstream fixed this in #38757.

This is not a "new models" problem. The catalog is fetched live from models.dev at build time (script/generate.ts; amicode-release.yml pins no snapshot), so Opus 5 already shows up in the picker — just hobbled, with no effort selector. QA-ing "can I select Opus 5?" gives a false negative on the need to sync.

What's ported

The end state of 4 files, not the 112 commits behind them. Cherry-picking that history would drag in the Effect layer-node → app-node refactor and the @opencode-ai/schema extraction (upstream gutted core/src/model.ts 127→41 and core/src/provider.ts 85→25 into a new package).

file	how	conflicts
provider/transform.ts	replaced wholesale w/ upstream HEAD	0 — fork never touched it
test/provider/transform.test.ts	replaced wholesale	0 — ours asserted 6 superseded behaviors
core/models-dev.ts	additive reasoning_options field	0 — without it the port is 4 type errors
session/llm/request.ts	git merge-file 3-way	0 — amicode patch #18 preserved

transform.ts also carries #38424 (prompt cache keys by SDK) and the Grok/GLM/Bedrock-Mantle variant fixes. request.ts brings #33392 (strict: false on OpenAI Responses-family tools).

Measured blast radius

Ran both transform versions over all 454 reasoning models in the live models.dev catalog: 418 identical, 36 changed, 0 errors.

• Claude 5 (opus + sonnet, across anthropic, all amazon-bedrock regions, and opencode zen): high,max → low,medium,high,xhigh,max with type:"adaptive".
• xAI Grok 4.3/4.5/4.20/build and GLM-5.2: gain effort variants where they previously had none.
• claude-opus-4-5: keeps its low,medium,high keys but now also sends budgetTokens:16000 alongside effort.
• Claude 4.8 / 4.6 / 4.1: unchanged.

Verification

• tsgo --noEmit green on packages/opencode and packages/core
• bun test test/provider/ → 503 pass / 0 fail
• bun test test/session/ test/server/ → identical before and after. The 3 llm-native-recorded failures and the Bun 1.3.14 teardown segfault both reproduce byte-for-byte on the unported tree.
• oxlint → 0 errors; prettier --check clean

Needs a QA pass before the hackathon build

CI cannot exercise these — they need live creds:

1. Whether the old legacy form actually 400s against Opus 5 or merely degrades. If it 400s, this was "Opus 5 unusable", not "Opus 5 hobbled".
2. claude-opus-4-5 — a model in use today — with the new budgetTokens+effort combination.
3. One xAI model, since Grok goes from no variants to effort variants.
4. Cache hit rate / spend, since #38424 changes prompt-cache keying across all providers.

Known trade

Porting file end-states makes a future clean upstream merge harder on these 4 files. Accepted deliberately. Worth noting the bug class recurs every model generation — version-regex parsing of model IDs will break again at Opus 6 — which argues for a standing narrow sync lane on transform.ts rather than one-off unfreezes.

Unrelated risk spotted (not fixed here)

The release build does a bare fetch(models.dev/api.json) with no fallback, so a models.dev outage hard-fails the build. Worth pinning MODELS_DEV_API_JSON for the hackathon build.