Pi/OMP-lane historian runs never record token usage (subagent_invocations always 0/0)

[VibePDF]

Symptom

Every Pi/OMP historian run shows input: 0 / output: 0 in the dashboard — including healthy 30–90s completed runs — while OpenCode-harness runs show real counts. From the shared DB:

• opencode historian: 1506 nonzero rows, 47.9M input / 1.6M output tokens recorded
• pi + omp historian: 345 rows, all 0/0 — every row, ever

Work demonstrably happened (durations, compartments produced); only the metering is blind.

Cause chain (verified live against 0.41.4 + omp 18.1.11)

1. omp emits usage. Live probe omp --print --mode json --no-session: message_end / turn_end / agent_end all carry message.usage = {input: 20153, output: 17, cacheRead: 241, cacheWrite: 0, totalTokens, reasoningTokens, cost} — structurally checked, not substring-matched.
2. The plugin never captures it. spawnPiHistorianRun builds {agentResponses, toolCalls, assistantTexts, ...} from the child stream with no usage plumbing, and the 0.41.4 bundle contains zero writes of usage into subagent_invocations on the Pi path (61 invocation hits, all unrelated graph terms).
3. No backfill path either. Session-entry-based computePiWorkMetrics reads message.usage from persisted entries — but historian children run with --no-session (ephemeral), so there are no entries to sum.

Net: the data exists in the child event stream, the expected shape ({input, output, cacheRead}) even matches the existing extractPiUsage-style extractor — it is simply never transferred into the invocation row.

Suggested fix

Capture message.usage from the child turn_end/agent_end events in spawnPiHistorianRun (or the finalize step) and persist input/output/cacheRead/cacheWrite on the invocation row, same as the OpenCode lane does from its runner callback.

[magic-alfonso]

Confirmed at source — thanks for the exact cause chain. Pi/OMP children run with --no-session, and spawnPiHistorianRun built its result from the child event stream without reading message.usage, so the invocation row was written 0/0 on every run.

Fixed on master: usage is now summed from the child's message_end events (matching the OpenCode lane's per-message accounting) with agent_end totals as the fallback, and persisted on the same subagent_invocations columns; a stream without usage still completes and records 0/0 rather than failing. The dashboard already renders those columns.

Ships in the next patch.