feat(plugin): add request-scoped chat.model hook

[millsydotdev]

Issue for this PR

Closes:

• [FEATURE]: add chat.model plugin hook for pre-call model routing #18793

Addresses / partially addresses:

• [FEATURE]:Support runtime model switching via plugin API #24006

Type of change

• New feature

What does this PR do?

Adds a request-scoped chat.model plugin hook that fires before provider/model/auth resolution. A plugin can replace the model for a single request without changing the session's selected model. Enables per-task routing (local vs cloud, fast vs heavy) without session mutation.

This is a rework of the earlier #30457 implementation against the current dev API (provider.getModel with ProviderV2.ID / ModelV2.ID, SessionV1.User, current llm.ts structure). The chat.model hook type is added to @opencode-ai/plugin; the LLM run path triggers it and resolves the routed model via the provider service; chat.headers and downstream hooks see the routed model; logs include routedFrom + reason for observability.

No models are hardcoded anywhere. The hook output must reference models from the user.s provider configuration.

How did you verify your code works?

• 3 new integration tests in packages/opencode/test/session/llm.test.ts: noop route keeps the original model, routing replaces the model for the request, and the hook receives session/agent/message context. Full file: 31 pass / 0 fail.
• bun run typecheck: 30/30 packages pass.
• bun run lint: no new issues in changed files (the repo's single pre-existing lint error is in packages/app/test-browser/prompt-persistence.test.ts, untouched).
• Docs: chat.model example added to the plugins docs (config/env-driven, no hardcoded models).

Screenshots / recordings

n/a

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[millsydotdev]

This is a first-contributor PR from a fork, so the typecheck/test/nix-eval workflows are waiting for maintainer approval before they run (status: action_required). Everything they check has been verified locally:

• bun run typecheck: 30/30 packages pass
• GITHUB_ACTIONS=false bun turbo test: all packages pass (packages/opencode: 3231 pass / 0 fail, 247 files, includes the 3 new chat.model tests)
• bun run check:generated: clean
• bun run test:httpapi: 208/208 pass (coverage + auth modes; effect mode re-running now)
• lint: clean on changed files (the repo's pre-existing packages/app failures are untouched by this diff)

Could someone approve the workflow runs? Happy to adjust anything on review.