agent-decision-log

WHY-layer decision log for AI agents. Records the reasoning behind each branch in an agent run: what options were considered, which one was chosen, the rationale, and what happened.

Sibling of agentsnap (CALLS) and agenttrace (COST + LATENCY). Together they cover the three audit dimensions of an agent run.

Install

[dependencies]
agent-decision-log = "0.1"

Usage

use agent_decision_log::DecisionLog;
use serde_json::json;

let mut log = DecisionLog::new();
let id = log.add(
    vec!["search_web".into(), "ask_user".into()],
    "search_web",
    "Query is specific enough to search without clarification.",
    json!({"turn": 3}),
);
log.set_outcome(&id, "Found 5 relevant docs.");

let d = log.find_by_id(&id).unwrap();
assert_eq!(d.chosen, "search_web");

JSONL persistence

use agent_decision_log::DecisionLog;

let mut log = DecisionLog::new();
// ... log.add(...) ...
log.to_jsonl("decisions.jsonl").unwrap();

let loaded = DecisionLog::from_jsonl("decisions.jsonl").unwrap();

Each decision serializes as one JSON object per line. Good for jq -c piping or DuckDB's read_json_auto.

API

• Decision { id, timestamp, options, chosen, rationale, outcome, meta }
• DecisionLog::new()
• DecisionLog::add(options, chosen, rationale, meta) -> id
• DecisionLog::set_outcome(id, outcome) -> bool
• DecisionLog::find_by_id(id) -> Option<&Decision>
• DecisionLog::last() -> Option<&Decision>
• DecisionLog::len() / is_empty()
• DecisionLog::to_jsonl(path) / from_jsonl(path)
• Decision::chose_listed_option() -> bool

License

MIT