Traces and replay

A trace is the full record of what a session actually did: every turn, every step, the prompt that went to the model, the completion that came back, and each tool call with its arguments and result. One append-only NDJSON file per session, on your disk, readable in any text editor.

Traces answer the questions logs cannot. Why did the agent pick that tool? What exactly was in the prompt when it went wrong? And — via replay — has anything changed underneath a session since it ran?

Developer Preview

Atomic Agent is a Developer Preview. Event shapes inside the NDJSON still move between releases. Treat traces as diagnostics, not as a stable data format to build on.

Tracing is on by default

The interactive entry points all record. run, tui, and serve boot the runtime with traceDefault: true, so a fresh install traces every session you drive by hand.

Two things do not trace by default:

• Background task drains. The task command boots with traceDefault: false. Turns executed by the task runner are not recorded unless you opt in explicitly.
• The sidecar. Embedded hosts stay silent unless configured.

The config key is tri-state, which is what makes that split work:

{
  "tracing": {
    "trace": {
      "enabled": null,
      "maxBytesPerSession": 10485760
    }
  }
}

tracing.trace.enabled	Meaning
null (default)	Defer to the entry point. run / tui / serve trace; task drains and the sidecar do not.
true	Always trace, whatever the entry point. This is how you record background task runs.
false	Never trace, including interactive sessions.

An explicit true or false always wins over the entry-point default.

maxBytesPerSession (default 10 MB) is a hard cap on one session’s file. Once a trace hits it, writes stop — the file does not rotate and the session keeps running untraced.

Where traces live

<stateDir>/traces/<sessionId>.ndjson

With the default state directory that is ~/.atomic-agent/traces/. Override the root with ATOMIC_AGENT_STATE_DIR.

Traces are confidential

A trace contains prompts and tool IO verbatim, with no redaction. That means your messages, file contents the agent read, command output, and anything else that passed through a turn — including secrets that happened to be in a file or an environment the agent touched.

Treat a trace file the way you would treat the session itself. Do not paste one into a bug report, an issue, or a chat without reading it end to end first and removing what should not travel. The state directory is not encrypted.

Commands

atomic-agent trace list   [--limit N]
atomic-agent trace show   <sessionId> [--step N] [--raw]
atomic-agent trace export <sessionId> [--format ndjson|json]
atomic-agent trace replay <sessionId>

list

Summarise the most recent trace files in the traces directory.

atomic-agent trace list --limit 20

show

Pretty-print one session as a human-readable chronology.

atomic-agent trace show s-1234
atomic-agent trace show s-1234 --step 2
atomic-agent trace show s-1234 --raw

--step N isolates a single step. --raw prints the unformatted lines instead of the formatted chronology. A session id with no trace file on disk exits 1.

export

Dump the raw trace for processing elsewhere.

atomic-agent trace export s-1234                  # ndjson (default)
atomic-agent trace export s-1234 --format json > session.json

--format accepts ndjson (the default) or json. Anything else exits 1. NDJSON streams the file as-is, one event per line — good for jq and grep. JSON wraps the events into a single array, easier to load whole.

replay

Rebuild the stable prompt prefix with the current tools, skills, and capabilities, then compare its hash against what the trace recorded.

atomic-agent trace replay s-1234

Output lists each turn and step with its recorded hash, the current hash, and whether they diverge:

  turn  step  drift  recordedHash  currentHash

Replay and prompt drift

replay does not re-run the agent. It answers a narrower and more useful question: would this session get the same prompt prefix today?

The stable prefix is the byte-identical head of every prompt — persona, tool catalog, skill index, capabilities. Replay reconstructs it from the current runtime and compares hashes with the ones the trace recorded. A mismatch means something underneath the session changed: a tool was added or removed, a skill was installed or disabled, or capabilities shifted.

The exit code is the part to script against:

Exit code	Meaning
0	No drift. The current runtime reproduces the recorded prefix.
2	Drift detected on at least one step.
1	The command failed — no such session, unreadable file, bad arguments.

atomic-agent trace replay s-1234 || echo "prefix changed since this session ran"

Why drift matters

The stable prefix is what llama-server caches per session. When it changes, the KV-cache is invalidated and every step reprocesses the prefix from scratch — sessions get slower and more expensive for no visible reason. Replay turns that into a hash comparison you can check in CI after installing a skill or bumping a release.

Reading a trace directly

The file is NDJSON, so the usual tools work:

# Every tool call in a session
jq -c 'select(.type == "tool_call")' ~/.atomic-agent/traces/s-1234.ndjson

# Follow a live session
tail -f ~/.atomic-agent/traces/s-1234.ndjson

Malformed lines are skipped

Trace writes are best-effort. If a line is truncated — a crash mid-write, a process killed during a turn — the reader silently skips it rather than failing the whole file. That keeps a damaged trace readable, but it means a trace is a diagnostic record, not an authoritative account of what happened. Do not treat the absence of an event as proof it did not occur.

Turning tracing off

If you would rather record nothing:

{
  "tracing": {
    "trace": {
      "enabled": false
    }
  }
}

That covers every entry point, interactive ones included. Existing files stay on disk — delete <stateDir>/traces/ yourself if you want them gone.

Related

Why local-first

Traces are one of the files that make the runtime inspectable, and one of the reasons the state directory is sensitive.

Approval gates

What stopped for approval during a turn, and why it is in the trace.

CLI Reference

Every command and flag, trace included.

Troubleshooting

Using a trace to work out why a turn went the way it did.