v0.1.2 - the gate-valid instrument

Adapter fix: harness-injected records (task notifications, system reminders) no longer split retry scopes - autonomous loops spanning a background wake-up were structurally invisible on unattended sessions. Coverage contract now explicitly names variation loops and dispatch loops as NOT-checked. Plus v0.1.1 fixes (serve fixed port 8765, report positional run id). 600 tests. Install: uv tool install agent-run-ledger

v0.1.1 - field fixes from release night

Two fixes found by real usage within hours of v0.1.0 going public:

• arl serve now defaults to port 8765 (was: an ephemeral random port that left healthy dashboards running where no browser could find them). The port is documented in --help; if it's busy, serve fails closed with direction instead of silently picking another. --port 0 still gives an ephemeral port.
• arl report <run-id> works positionally alongside --run (conflicting ids and a missing id both fail closed with guidance).

599 tests passing. Install: uv tool install agent-run-ledger.

v0.1.0 — graded repair receipts for agent runs

First public release. Local, zero-egress CLI that reads Claude Code / Codex CLI / OpenAI Agents SDK session logs and emits graded repair receipts — honest-abstain by design: 'clean' never means 'verified correct'.

Start here: arl selftest (watch a receipt fire on a bundled known-bad run, ~30s), then arl sweep ~/.claude/projects (grade months of sessions you already have).

Shipped detector classes: retry_loop (graded L0–L2), artifact_failure (R1/R2). Every verdict states what was checked and what was NOT. Exit codes: 0 clean / 3 receipts fired / 1 unreadable (fails closed).

Install: uv tool install agent-run-ledger (PyPI) or uv tool install git+https://github.com/huyn7539/agent-run-ledger.

License: FSL-1.1-ALv2 — free for everyone including commercial internal use; converts to Apache-2.0 automatically two years after release.