Skip to main content
Evidence is the ordered record of what happened during a run: Tool calls, state mutations, events, faults, callbacks, time changes, assertions, target output, and optional attachments.

​
Read a result

Start with the verdict, then follow the causal order:
  1. task and starting scenario;
  2. agent interactions and Tool calls;
  3. state, event, fault, and time changes;
  4. expected and actual assertion values;
  5. retained logs, screenshots, recordings, or files;
  6. build hash, seed, and reproduction command.
Firedrill writes a central report index at .firedrill/reports/index.html and a portable bundle for every retained attempt. JSON supports automation, JUnit integrates with CI, and HTML is the primary human-readable view.

​
Reset semantics

A full reset restores the world’s baseline state, clock, randomness, pending work, faults, and journal. A Tool-scoped reset restores selected Tool packages while preserving global time, earlier evidence, and unselected Tool state.
Every successful reset advances the world’s generation. Restart state and evidence cursors when the generation changes.

​
Reproduce a run

Reports retain the exact content-addressed build hash and seed. The printed reproduction command runs that build again without silently compiling different source.
Keep the corresponding project-local build directory. If it has been removed, restore the source revision and per-test setup that produced the hash, then compile again.
Local report verification checks file hashes, schemas, identities, and evidence ordering. Local bundles are not cryptographically signed. Hosted attestation is a separate Cloud capability.

Read reports and the Inspector

Navigate the report index, compare runs, and verify a bundle.