mythos-skill
Deterministic packet compiler for AI agent runs.
Takes raw subagent output (evidence JSONL, verifier findings, raw artifacts) and compiles a schema-validated, hash-provenanced next_pass_packet.json that the orchestrating model reads instead of raw subagent prose.
Part of the mythos-skill project — see the repo for the full JS runtime (ingest, strict gate, readiness) that drives this crate.
What this crate does
- Reads a run directory containing
manifest.json,worker-results/evidence.jsonl,verifier-results/findings.jsonl, andraw/artifacts. - Validates every
source_ref— hashesfile:references against disk, checks line-range spans, enforces provenance rules for substantive evidence kinds. - Promotes evidence into
trusted_factsusing agent-suppliedconfidencewhen present. - Auto-detects
Contradictionentries when different agents assert divergent summaries on the same direct source span, graduating severity by evidence kind. - Emits
next_pass_packet.json,snapshot.json,decision_log.jsonl— byte-deterministic for byte-identical inputs.
Install
Run
Expected run directory shape:
manifest.jsontask.mdraw/worker-results/evidence.jsonlverifier-results/findings.jsonl
Outputs land in state/ inside the run directory.
Minimal evidence shape
Determinism guarantee
The compile_determinism integration test runs the compiler twice on a byte-identical fixture and asserts byte-identical next_pass_packet.json + snapshot.json. A regression here means a non-deterministic code path slipped in.
License
MIT