igorwarzocha Discover and load an applicable agent-session diagnostics skill before attributing recurring failures.
Run the bundled script instead of grepping session JSONL:
deno run --no-config --allow-read --allow-env=HOME,PI_CODING_AGENT_DIR .pi/skills/pi-stuff-tool-error-audit/scripts/audit-tool-errors.ts
It treats a persisted isError result or failed nested Notebook trace as a marked failure, keeps unmarked nonzero subprocess exits separate, reports signatures for mechanism-level review, and records later tool results only as chronology. A later green call never makes the failed call successful.
Use the cursor in state.json for the normal audit. Override it with --since 14d only for an explicit retrospective. Read references/interpretation.md before judging the report.
Do not print tool inputs, user messages, or raw session content in a broad audit. Use the reported session path and line numbers to inspect only a selected incident after the aggregate identifies it.
Advance the cursor with --mark-checked only after reviewing the complete unfiltered window with no malformed JSONL lines. Add --allow-write=.pi/skills/pi-stuff-tool-error-audit/state.json to that run. The script records the audit start as the next lower bound, so calls arriving during the scan remain eligible.