Milton / v0.1.0

Understand what your agents actually did.

Milton reads the traces, receipts, ledgers, and Git history you already have. It joins work across layers without inventing custody, separates reported from computed cost, and turns repeated evidence into findings you can inspect and grade.

pip install milton-ai
milton init
milton scan --since 7d
milton accounting --since 7d
milton cost --per-outcome --since 7d

Conservative accounting

Reported and computed amounts stay visible. Exact shared keys—not adjacency or similar totals—control deduplication.

Outcome attribution

Inspect the path from model call through runner and work registry to shipped outcome, including ambiguity and coverage gaps.

Graded findings

Failure motifs, procedure candidates, and memory-hygiene leads carry bounded evidence, review state, and expiry.

Local by default

The dependency-free core starts no service, sends no telemetry, and hashes transcript content unless full recovery is requested.

Status: v0.1.0 pre-alpha release. The deterministic foundation, twelve read-only adapters, accounting, attribution, append-only finding review, and one evaluated procedure-return loop are implemented. Some finding generators remain deliberately offline or narrowed until their held-out evidence is strong enough.

Python 3.12+ · MIT licensed · dependency-free runtime core