Cut PostHog event volume with generation aggregates and subagent rollups

[TheGreatAxios]

Summary

• Folds per-tool $ai_spans into $ai_generation aggregates by default (tool_call_count, tool_error_count, subagent_call_count); opt-in spans via CORBITS_TELEMETRY_AI_SPANS
• Enriches subagent_end with run rollups (model, turns, tokens, tool counts, stop_reason, spawn-time parent_trace_id) and allowlists first-party director ids
• Stamps $process_person_profile: false on ambient events, dedupes plugin_loaded per plugin per process, and documents a generation sample-rate knob (default 1.0)

Verification

• bun run typecheck, bun run build, and bun run test pass (bun run check exit 0 locally)
• Critic re-review clean after parent_trace_id fix

Fixes CL-6816

[linear-code]

CL-6816

[TheGreatAxios]

Primary · Request changes

This aggregates parent-turn telemetry, rolls worker activity into subagent_end, and makes ambient capture anonymous.

Findings

• src/telemetry/ai-observability.ts:207 — A recoverable inference error always emits a failed generation before the successful retry, so one completed parent turn can emit two generations or look failed-only when success sampling is zero.
• src/telemetry/ai-observability.ts:47 — Only task is classified as a subagent call, so preferred spawn_agent dispatches inflate tool_call_count and leave subagent_call_count at zero.
• src/subagent/run.ts:1147 — Worker failures rethrow before accumulated turn, token, tool, and error totals reach subagent_end; clean completion also omits the required stop reason.
• src/subagent/agent-fleet.ts:575 — Fleet setup failure returns after subagent_start but before the terminal emitter is registered; fused task setup failure reports the default completed status at src/subagent/task-tool.ts:936.
• src/exec/runner.ts:660 — Exec mode does not install the turn-start observer used to capture the active parent trace, so worker terminal events omit parent_trace_id.
• src/telemetry/product-events.ts:73 — Plugin dedupe records identity even when capture uses the no-op telemetry handle, permanently suppressing a later real plugin_loaded event in the same process.
• commit 75df24f6 — The subject includes CL-6816, which the repository commit rules prohibit.
• CL-6816 — The representative before/after event-count measurement remains unchecked, so the target reduction is not demonstrated.

Notes

• bun run check exits 0 across lint, typecheck, build, and tests.
• GitHub prevents a formal self-request-changes action because the PR author and authenticated reviewer are both TheGreatAxios; this comment review records the substantive request-changes verdict.

[TheGreatAxios]

Critic · Request changes

This reshapes AI observability around parent-turn aggregates and per-dispatch worker rollups.

Findings

• src/session/run-sink.ts:178 — inference.error immediately emits a failed generation even when the run sink retries. At sample rate zero, an eventually successful turn is recorded only as failed; at rate one, the same trace emits both failed and successful generations.
• src/telemetry/ai-observability.ts:47 — The subagent tool-name check excludes spawn_agent, so its calls and optional spans are recorded as ordinary tools instead of subagent calls.
• src/subagent/run.ts:1147 — A worker that fails after prior activity throws away its accumulated rollups before callers can enrich subagent_end; src/subagent/run.ts:1071 also returns clean completion without a stop reason.
• src/subagent/task-tool.ts:936 — Fused setup failures inherit a default completed status, while src/subagent/agent-fleet.ts:575 returns before fleet dispatch installs its subagent_end emitter. These paths create false-success or unmatched start/end telemetry.
• src/exec/runner.ts:660 — Exec parent turns never update the current turn trace used by subagent dispatch, so src/telemetry/product-events.ts:44 cannot attach parent_trace_id.
• src/plugins/loader.ts:143 — Loads without a telemetry handle still pass through global dedupe with NOOP_TELEMETRY; the identity is consumed without capture and later enabled telemetry cannot emit the plugin event.

Notes

• Permanent coverage should include recovered retries at sample rates zero and one, both dispatch tool names, worker failure after prior activity, clean stop reason, setup failures for both dispatch primitives, exec trace linkage, and no-op-then-enabled plugin loading.
• Targeted telemetry tests pass with 75 tests and zero failures; the independent full bun run check also exits 0.

[TheGreatAxios]

Primary · Approve

This reduces ambient telemetry volume with parent-turn aggregates, truthful worker rollups, anonymous processing, and conservative fallback attribution.

No findings.

Notes

• bun run check exits 0 at ba4c34ad.
• Required GitHub checks and CLA pass.
• The deterministic representative fixture records 98 → 18 billable events, an 81.6% reduction.
• GitHub prevents formal self-approval because the PR author and authenticated reviewer are both TheGreatAxios; this comment review records the approval verdict.

[TheGreatAxios]

Critic · Approve

This emits one terminal parent generation, enriches one worker terminal event per dispatch, and avoids false provider attribution when runtime source identity is unavailable.

No findings.

Notes

• Retry, sampling, setup failure, cancellation, interruption, plugin dedupe, exec linkage, and privacy paths have permanent regression coverage.
• Pre-usage fallback failures retain the attempted model and use the fixed unknown provider bucket instead of claiming a stale selected provider.
• The final diff contains no Interchange, vendor, dependency, or lockfile changes.