session_start and task_classified telemetry fire once per user turn, not once per session

Start in packages/opencode/src/session/prompt.ts at loop() around lines 364 and 498, then inspect the telemetry block around line 939 and how session-scoped state or prior assistant messages are available. Reproduce the issue with three messages and inspect telemetry. Done means each session emits one session_start and one task_classified from the opening turn, while the intentionally per-loop trace at line 1088 remains unchanged.

Summary

session_start and task_classified are emitted once per user turn, not once per session. Session counts in telemetry are therefore inflated by roughly the average number of turns per session, and task_classified re-classifies intent on every turn instead of once from the opening message.

Detail

Both events are emitted from a if (step === 1) block in packages/opencode/src/session/prompt.ts:939:

• session_start — packages/opencode/src/session/prompt.ts:953
• task_classified — packages/opencode/src/session/prompt.ts:1004

step is not session-scoped. It is declared inside loop():

• let step = 0 — packages/opencode/src/session/prompt.ts:364
• step++ — packages/opencode/src/session/prompt.ts:498

and loop() is invoked once per prompt, i.e. once per user turn. So step === 1 means "first model step of this turn", and the block runs again for every subsequent message in the same session.

Impact

• session_start overcounts sessions. A session with N user turns emits N session_start events, all carrying the same session_id.
• Any funnel or retention metric built on session_start counts is proportionally wrong. De-duplicating by session_id at query time recovers the true count, so historical data is salvageable, but no current dashboard is doing that unless it was written with this in mind.
• task_classified reports the intent of whatever the user most recently typed, not the task the session was opened for. Its confidence distribution is also skewed by short follow-up messages ("thanks", "yes", "try again") being classified as tasks.
• Cost: an extra pair of tracked events per turn, plus the classifyTaskIntent regex sweep over up to 2000 characters per turn.

Reproduction

Start a session, send three messages, and inspect the emitted telemetry (or set a breakpoint on the step === 1 block). Three session_start events are emitted with identical session_id.

Suggested fix

Gate both events on something session-scoped rather than on step. Options, roughly in order of preference:

1. Check whether the session already has prior assistant messages before emitting — a session that has already produced output is not starting.
2. Keep a per-session "already emitted" claim in the telemetry module, similar to how other once-per-session state is tracked.
3. Emit at session creation instead of in the prompt loop, which is where the event name says it belongs.

Note that packages/opencode/src/session/prompt.ts:1088 also uses if (step === 1), but that one is intentionally per-loop() (it traces the system prompt once per loop invocation, and the comment there says so). Only the telemetry block at line 939 is affected.