refactor(runtime): move chunk interpretation into the agent adapter

[FreshlyBrewedCode]

Part of #32 · Closes #35

runtime/agent-step.ts string-matched CUSTOM + opencode.session-id to populate
AgentStepFinished.sessionId, and string-matched structured-output.complete — which isn't a
provider capability at all, but @tanstack/ai-opencode's text adapter simulating structured
output by re-parsing its own final message (ADR 0001 §1). That left opencode's event vocabulary
leaking into the runtime, so a second adapter (claude-code, codex, a raw provider loop) would have
had to emit opencode's exact CUSTOM event names just to surface a session id or a structured
output. ADR 0012 §1/§2 draws the seam: the adapter interprets its own stream, the runtime consumes
a normalized signal.

What changed

• AgentAdapter.stream now yields AgentAdapterYield ({ chunk, signal? }) instead of a bare
  unknown chunk. AgentSignal is a small closed union — sessionId, structuredOutput,
  runError — documented against the exact AgentStepFinished fields each one populates.
• opencode-adapter.ts gains extractOpencodeSignal, the only place under src/runtime/ left
  that string-matches opencode.session-id and structured-output.complete; it also folds in
  RUN_ERROR detection, which agent-step.ts used to do itself.
• agent-step.ts drops all vendor string-matching: it now just switches on yieldItem.signal._tag
  to populate partial.sessionId / structuredOutput / runError, and keeps interpreting only
  the standard AG-UI TEXT_MESSAGE_*/RUN_FINISHED types for finalText accumulation and usage.
• replay/adapter.ts's corpus replayer and createSlowFakeAdapter both updated to the new yield
  shape; the corpus replayer reuses extractOpencodeSignal so recorded opencode traces keep
  producing the same signals without re-deriving the logic.
• New tests: runtime/agent-step.test.ts (signal-to-field wiring, chunk passthrough, no-signal
  case), runtime/opencode-adapter.test.ts (extractOpencodeSignal unit coverage including the
  guarded/negative cases), and a corpus round-trip assertion in replay/adapter.test.ts that
  checks signal extraction end-to-end against a real recorded trace.

Notes for reviewers

• RUN_ERROR is a standard AG-UI type, not opencode-specific, but its detection now lives inside
  extractOpencodeSignal alongside the two vendor-specific signals — every adapter owns detecting
  all three signal kinds for its own stream, per ADR 0012 §2, rather than the runtime special-casing
  one AG-UI type generically.
• The corpus replay adapter imports extractOpencodeSignal from opencode-adapter.ts rather than
  duplicating the matching logic. That's a deliberate coupling: the recorded corpora are opencode
  chunks, so replaying them faithfully means reusing the real interpretation, not reimplementing it.
  A future non-opencode adapter is not obligated to imitate opencode chunk shapes at all — it would
  supply signals through its own stream implementation.
• resolveOutput's two-tier structured-output extraction (event value, then finalText re-parse)
  in runtime/run.ts is untouched — this ticket only changes where the tier-1 value comes from.
• Vendor event-name strings (opencode.session-id, structured-output.complete) still appear in
  agent-step.test.ts and opencode-adapter.test.ts as literal test fixtures for the extraction
  logic — that's expected; no interpretation logic outside opencode-adapter.ts matches on them.

Verification

• Added unit tests cover signal extraction (extractOpencodeSignal), signal-to-field wiring in
  buildAgentStepEffect, and a real-corpus round trip confirming the recorded session-id trace
  still yields the expected signal and chunk count.
• CI's check job (format, lint, typecheck, bun test) passed on this PR; I did not additionally
  run it locally. There's no CI coverage of a live opencode run, so AgentStepFinished.sessionId/
  .output/.error parity for a live run is established by code inspection (the moved logic in
  extractOpencodeSignal is byte-for-byte equivalent to what agent-step.ts used to do) rather
  than an end-to-end run.

Stack

1. refactor(cli): replace hand-rolled parsers with effect/unstable/cli #52 — 33-parse-cli-with-effect-cli (issue Parse CLI arguments with effect/unstable/cli #33)
2. refactor(errors): represent domain failures as Schema.TaggedError #53 — 34-domain-errors-as-tagged-errors (issue Represent domain failures as Schema.TaggedError #34)
3. refactor(runtime): move chunk interpretation into the agent adapter #54 — 35-move-chunk-interpretation-into-adapter (issue Move chunk interpretation into the agent adapter #35) — this PR
4. refactor(runtime): move headless permission setup into the agent adapter #55 — 37-move-headless-permissions (issue Move headless permission setup out of the workspace allocator #37)
5. feat(runtime): add Effect composition root and agent runtime service #56 — 36-agent-runtime-service (issue Add an Effect composition root and make the agent runtime a service #36)
6. refactor(daemon): move singletons into per-daemon layers #57 — 38-move-singletons-into-layers (issue Move the daemon's remaining singletons into layers #38)

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}