Replay - Agent Sentinel

Replay lets you run agent logic without re-calling external systems by returning recorded outputs for instrumented actions.

Use replay mode

from agent_sentinel import replay_mode, guarded_action

@guarded_action(name="fetch_data", cost_usd=0.01)
def fetch_data(url: str) -> dict:
    return {"url": url, "data": "..."}

# First run: records to ledger
fetch_data("https://example.com")

# Later: replay from ledger
with replay_mode(ledger_path=None, strict=True):
    fetch_data("https://example.com")  # returns recorded output

Divergence detection

In strict mode, replay raises ReplayDivergenceError when inputs don’t match what was recorded.

Per-run replay (`ReplayMode.from_run_id`)

SDK 0.2.0 stamps a run_id onto every ledger entry (sourced from the active ExecutionContext). ReplayMode.from_run_id("…") filters a shared ledger file down to a single run before replaying — useful when one ledger file contains many runs.

from agent_sentinel import ReplayMode

with ReplayMode.from_run_id("run-2026-04-18-abc"):
    fetch_data("https://example.com")  # replays only entries from this run

If you’d rather not share ledgers across runs at all, set a different AGENT_SENTINEL_HOME per run so each run gets its own ledger file.