State & artifacts

A run produces three things on disk:

Path	What	Source of truth?
`state.db`	SQLite index of the whole run	Queryable index
`results/<run-id>/<stage>/*.jsonl`	Every message exchanged with each agent	Yes — durable raw record
`work/<run-id>/hunt/<task-id>/`	Per-Hunt-task scratch dir (PoC compile/run)	Scratch

All three are gitignored.

SQLite schema (`src/state.ts`, `bun:sqlite`)

Table	Holds
`runs`	One row per run: repo path, timestamps, status (`running`/`completed`/`aborted`/`failed`)
`recon_outputs`	The raw Recon payload, keyed by run
`tasks`	Hunt tasks: attack class, scope hint, target files, priority, `source` (recon/gapfill/feedback), status (`pending`/`running`/`done`/`failed`)
`findings`	Hunt findings + validation verdict + dedupe group + `is_canonical`
`traces`	Reachability verdict per canonical finding
`dedupe_groups`	Root-cause clusters with a canonical member
`costs`	Per-stage cost / token / turn / duration rows (powers `--max-cost-usd` and `status`)
`artifacts`	Pointers to the `results/` JSONL files and scratch dirs

You rarely touch this directly — audit status and audit report read it for you — but it's a plain SQLite file if you want to query it.

Each line is one JSON event, appended as it happens. Useful kinds:

`kind`	Meaning
`meta`	Stage + model + start time
`user`	The input sent to the agent
`assistant`	A model message (text / thinking / tool_use blocks)
`result`	The SDK result (cost, tokens, turns, errors)
`repair_request`	A schema-validation repair turn was issued
`schema_errors`	The validation errors that caused a failure
`final_payload`	The validated output that was persisted

When a stage misbehaves, this is where to look — e.g. schema_errors + final_payload show exactly what the model emitted and why it was rejected.

--resume <run-id> continues an existing run:

The run's status is reopened to running.
Tasks left running (interrupted mid-flight) or failed (transient/quota error) are flipped back to pending so they're re-attempted.
Recon is skipped if its output already exists; already-validated findings and already-traced canonicals are skipped too.

This is why a quota exhaustion leaves the in-flight task pending rather than failed — so resume picks it up cleanly once quota returns.

rm -rf state.db state.db-* results/ work/

Outputs in results/ are not scrubbed of anything the agents read from the target (which can include .env / secrets). Treat them as sensitive.