imbue-mngr-pi-coding-usage

pi data provider for mngr usage. pi loads a single explicit extension (mngr_pi_coding's lifecycle extension, via pi -e), so the per-message usage writer lives there -- it already holds each assistant message's cost and tokens. This package owns the two usage-specific pieces:

• The reader: an aggregate_usage_source hookimpl claiming the pi-coding source, aggregated session-incrementally (pi reports cost per message, summed per session).
• The writer gate: on_after_provisioning drops a pi_emit_usage marker in each pi agent's state dir. The lifecycle extension only emits usage events when that marker is present -- so the events are written exactly when their reader (this package) is installed.

What gets captured

pi computes per-message cost client-side (usage.cost.total), so cost is REPORTED (no estimation). Each assistant message_end appends one cost_snapshot to events/pi-coding/usage/events.jsonl with the reported cost, tokens (input/output/cacheRead/cacheWrite), the provider-qualified model (<provider>/<model>), and the session id (the pi session file's basename). cost_mode is API_KEY (pi bills a real provider key).

Why the writer is in the harness extension

Unlike Claude (statusline shim) or OpenCode (auto-loaded plugin/*.ts), pi loads exactly one explicit extension owned by mngr_pi_coding. A standalone writer would require a harness change either way, so the writer is folded into the lifecycle extension (which already extracts the message usage), and this package stays the reader + gate.