fix(laminar): drain OTel processors before process.exit() to stop losing final agent spans

[Alezander9]

Summary

Laminar traces were missing the final agent LLM span on ~27% of overnight eval runs. Eddie's measurements:

eval	n	missing-final	rate
gemini-3-flash-preview	99	75	75.8%
glm-5.1 (3 seeds pooled)	300	102	34.0%
mimo-v2.5-pro	100	14	14.0%
gpt-5.5	100	1	1.0%
claude-opus-4-7	100	0	0.0%
all	699	192	27.5%

Score impact is nil (the runner reads agent steps from --format json stdout, not from spans), but it makes the Laminar UI useless for any human dive into a glm/gemini trace — the closing text-emit LLM call is just not there.

Root cause

Three problems compose:

1. Plugin event hooks are fire-and-forget. packages/opencode/src/plugin/index.ts:249 does void hook["event"]?.({event}). The bcode-laminar session.idle handler ends the turn span and await processor.forceFlush(), but its Promise is discarded by the bus subscriber.
2. process.exit() is unconditional. packages/opencode/src/index.ts:252 calls it in the top-level finally. Node's event loop is killed mid-flight; any in-flight gRPC export from the BatchSpanProcessor is severed.
3. BSP defaults. maxExportBatchSize=512, scheduledDelayMillis=5000ms. The very last spans of a run sit in the queue waiting either for the 5 s timer or our forceFlush — and forceFlush loses the race to process.exit().

Why it's model-dependent

The bug only ever bites the very last LLM span. Models split into two camps:

• Tool-call+text-in-same-step (claude-opus, gpt-5.5): they reason aloud while choosing the next action, so the final answer text rides on a step that also had a tool call earlier in the run. That doStream span ended many seconds before idle and was already flushed in an earlier batch. → 0–1% miss.
• Tool-only-then-final-text (glm-5.1, gemini-3-flash): every middle turn is pure tool calls, then ONE extra LLM round emits only text with finishReason=stop. That lone final doStream span ends 50–200 ms before idle, so it's the freshest unflushed thing in the BSP queue when process.exit() fires. → 34–76% miss. Gemini-3-flash is the worst because its TTFB is fastest, so the gap between span-end and idle is the smallest.
• Mimo at 14% sits in between (mixed behavior).

It's a clean continuum on "fraction of runs where the closing text is in a separate, late, lone span."

What this PR does

In the top-level finally of packages/opencode/src/index.ts, before process.exit():

1. Fetch the global OTel TracerProvider via @opentelemetry/api (already a dep of packages/opencode).
2. Duck-check for forceFlush (the API-level type doesn't declare it; BasicTracerProvider from @opentelemetry/sdk-trace-base, which NodeSDK.start() registers, implements it).
3. Promise.race against a 3 s timeout so a wedged exporter cannot hang bcode on exit.

Generic OTel-API approach with no laminar import — benefits any future tracing plugin equally (DataDog, New Relic, OTLP collector, etc.).

What this PR does NOT do

• Does not change scheduledDelayMillis. The default 5 s exists to batch many spans into one gRPC call. Reducing it to 1 s would multiply gRPC traffic during long runs (~5×) without addressing the root race; once we drain at exit, the queue is fully flushed regardless of timer. Leaving the BSP defaults alone.
• Does not fix the underlying fire-and-forget event subscriber (src/plugin/index.ts:249's void hook["event"]?.(...)). That's the proper upstream fix; should be a PR to anomalyco/opencode. This patch is a tactical drain at the only place we control unconditionally — the exit point.

Risk / blast radius

+12 lines in index.ts. One Yellow-zone touch (logged in EXCEPTIONS.md § "Drain OTel processors before process.exit()"). Behavior when no OTel provider is registered: the duck-check fails and the exit path is unchanged. Behavior when LMNR_PROJECT_API_KEY is unset (laminar plugin no-ops): the global TracerProvider is the API noop provider with no forceFlush, duck-check fails, exit path is unchanged. Behavior when exporter is wedged: bounded 3 s wait, then process.exit() fires anyway.

Expected effect on the missing-final rate: ~27% → near-zero across all models.

Verify

• bun typecheck clean (filtered + per-package).
• Manual eval run on glm-5.1 / gemini-3-flash recommended to confirm rate drops. (Eddie can re-run a small slice with the patched binary.)

[cubic-dev-ai]

No issues found across 1 file