chore(llm): replace vendored lib/llm with the published @opencode-ai/ai

[Makisuo]

Why

lib/llm was a vendored copy of opencode's packages/llm, pinned by SHA, kept in sync by scripts/sync-llm-upstream.ts and guarded by a drift workflow. The justification in MAPLE.md was that upstream is private: true — "no npm release to depend on and no semver contract".

That is no longer true. opencode v2 publishes the same core as @opencode-ai/ai: identical dependency set (@smithy/eventstream-codec, @smithy/util-utf8, aws4fetch, effect@4.0.0-rc.111 — our pinned line), the same export tree, and a superset of features. Both of our local deltas also dissolved upstream: it now uses Schema.TaggedError, and it maps the SSE decoder failure into its public error itself.

So the vendor, the SHA pin, the sync script, the drift workflow, and the oxfmt/oxlint/knip ignore entries that existed only for lib/llm all go away. Only apps/api ever depended on it.

What a reviewer should know

Version channels are dev/beta/next with no semver, so the version is pinned exactly (0.0.0-beta-18050). Note bunfig.toml sets minimumReleaseAge = 86400, so a bump has to pick a build older than 24h.

Upstream had moved on in ways that are more than renames. Each of these is a real behavioural surface:

• Model → LanguageModel; LLMError → AIError, tag "LLM.Error" → "AI.Error".
• LLM.updateRequest → LLMRequest.update, whose patch is canonical, so tool choice goes through ToolChoice.make("none") / ToolChoice.named(name).
• finishReason on LLMResponse and on the finish event is now { normalized, raw }, not a bare string.
• retryable is gone from AIError and from the provider-error event. Classification moves to the Maple seam as RETRYABLE_REASONS = {RateLimit, ProviderInternal} in platform/Llm.ts — exactly what the old per-reason getters returned. The in-band provider-error retry branch in turn.ts is deleted: only the Bedrock protocol ever set that flag, and Llm.ts deliberately does not import Bedrock, so the branch was already dead for both live providers. Its test is rewritten to pin what remains (terminal, and the provider's message never reaches the client).
• defaults.limits is gone. Maple's context/output windows now live in a WeakMap side table keyed by the model withLimits returns. providerOptions is not usable for this — everything in it is sent to the provider.
• providerOptions is no longer namespaced per provider. The adapter lowers the whole record into the request body, so our { openrouter: { reasoning, usage } } was shipping a literal openrouter body key and the reasoning-effort default never reached OpenRouter. This one typechecked cleanly; only the tests caught it. Worth a second pair of eyes.
• LLM.stream/generate now honestly declare R = LLMClient.Service — the vendored copy erased that requirement with an as cast. It now threads through runChatTurn, and buildTaskTool takes the resolved client as a value and provideServices it, because a tool handler's effect must require nothing.

Two things left alone deliberately: the wire-visible error tag "@maple/llm/LlmCallError" in packages/domain/src/llm.ts (renaming it changes error identifiers), and apps/api's typecheck:test, which is red in ~15 pre-existing files unrelated to this change.

Verification

• bun typecheck — 39/39 tasks
• bun run lint
• bun run --cwd apps/api test — 2439 passed, 277 skipped
• wrangler deploy --dry-run — 13.6 MB raw / 2.7 MB gzip, with no google-auth-library and no node builtins pulled into the Worker bundle

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

🍁 Maple PR preview

Warning

Preview cleanup could not be confirmed. The Alchemy teardown outcome was skipped.

Final commit fe80377 · View workflow run