fix(analytics): read pi's arguments wrapper in tool/file shapes + rollup repair pass (warren-677c)

[jayminwest]

Summary

Pi toolCall blocks wrap their args as {"name":"bash","type":"toolCall","arguments":{"command":…}}, but PI_TOOL_SHAPE (src/core/tool-shape.ts) read only payload.input.command / payload.command, and readPathsWith (src/core/file-shape.ts) read only payload.input. In production every pi bash rollup row (10,583/10,583) landed command=NULL and every pi edit/read/write row (3,944/3,944) landed empty file_paths — /analytics/behavior command mining has been frozen on Jul 28–31 claude-code data while 311/319 runs are pi (warren-677c).

Changes

• src/core/tool-shape.ts — pi readToolUse falls back to payload.arguments.command; precedence is input > arguments > top-level. Tests pin the exact observed prod payload.
• src/core/file-shape.ts — readPathsWith now takes an args-container key list: claude-code reads ["input"] (arguments dialect stays pi-only), pi reads ["input", "arguments"].
• repairToolCallRollup (src/runs/tool-calls-backfill.ts) — the boot backfill only targets runs with no rollup rows, so mis-extracted rows never re-enter its candidate set. The repair pass keyset-walks ToolCallsRepo.listRunsWithRollup, deletes each run's rollup rows, and replays its retained tool events through the same extraction seam the bridge/backfill use. Runs whose events were pruned are skipped, never deleted (the rollup is only derived state while its source survives). Idempotent end to end.
• scripts/repair-tool-calls.ts — one-off prod entrypoint: WARREN_DB_URL required with no default, --dry-run counts candidate runs. events covers 318/319 prod runs, so the corpus is fully repairable.

Testing

• New unit tests pin the pi arguments-wrapper shape for both command and file-path extraction, dialect precedence, and claude-code's non-reading of the pi dialect.
• Repo tests cover listRunsWithRollup keyset paging + deleteForRun (sqlite + postgres suites).
• Repair tests cover: re-extracting damaged rows the boot backfill skips, idempotency, events-pruned skip guard, paging, and result-join replay.
• Local check:all: 11/12 gates pass; check:coverage fails only on the pre-existing warren-4de5 macOS test-order flake (fails identically on a clean tree).

Follow-up (not in this PR)

After merge + deploy, run the one-off repair against prod:
WARREN_DB_URL=… bun run scripts/repair-tool-calls.ts (optionally --dry-run first), then re-check /analytics/behavior.

🤖 Generated with Claude Code