pi-usage-maintenancelisted

# Pi usage maintenance This is a manual-only runbook for maintaining Pi usage reporting. Do not invoke it from an ordinary coding prompt, and do not silently collect or upload telemetry. ## Invariants - Treat Pi's append-only JSONL/session records as the source of truth. SQLite, caches, charts, and summaries are rebuildable indexes, never the only copy. - Preserve raw records and timestamps; normalize only in derived tables. - Separate observed provider-reported usage from estimates. Label missing, sampled, or inferred values instead of filling them with guesses. - Keep historical pages immutable while allowing a bounded live tail to update. - Aggregate by default. Do not expose prompts, responses, paths, account identifiers, or secrets in reports unless the user explicitly requests a redacted sample. ## Procedure 1. Confirm the requested scope, Pi data root, timezone, retention policy, and whether this is a dry run. Resolve paths explicitly; never scan a home directory or delete by a broad glob. 2. Snapshot or checksum the input files before changing derived state. Record the highest source offset/timestamp processed so a rerun is idempotent. 3. Rebuild or incrementally sync the index from JSONL. Use stable event IDs; tolerate truncated final lines and report them for later repair. 4. Recompute daily and monthly aggregates, including request counts, tokens (input/output/cache when available), latency, errors, and provider/model dimensions. Keep