Decision log — one immutable record per adoption

The engine keeps the artifact (ledger), the evidence (trajectory, candidates, audit) and the run (runstore). A product also needs the decision as one object: what was proposed, what the guard said, what evidence and cost it saw, and which decision it descends from.

from agentdescent import DecisionLog

log = DecisionLog("~/.agentdescent/decisions.jsonl")
record = log.record(
    kind="adoption", artifact_id="skill", diff_id="d1",
    accepted=False, category="leakage:gold",
    evidence={"leaked": ["321"]}, cost={"calls": 12})
print(log.children(record.record_id))

DecisionRecord fields: record_id, kind, artifact_id, diff_id, accepted, category, detail, evidence, cost, parent_id, created_at. DecisionLog is append-only JSONL with record, append, iteration, get, children and count. Malformed lines (including valid JSON that is not an object) are skipped rather than crashing an audit, non-finite floats are written as null, and a torn final line is closed before the next append so it cannot swallow a new record. record_id is the primary key: append refuses a duplicate. Writers in one process share a per-path lock; separate processes need an external lock.

This is a building block: the engine does not yet write decisions on the adoption path by itself. get, children and count scan the file, which is the right trade for an audit log of this size and not for a high-QPS stream.

parent_id is what makes a run a tree: a session decision points at the previous session, an adoption at the session that created it. That is the link a dashboard, a replay tool or an audit export needs.