Agent Session Record

​

What a record looks like

{
  "type": "session",
  "seq": 0,
  "id": "h",
  "formatVersion": "0.1",
  "agent": {"name": "claude-code", "version": "2.1.305"},
  "sessionId": "0f3c2a7e-91b4-4d0e-8f43-2b6c1d7a9e10",
  "producer": {
    "name": "example-converter",
    "version": "1.0.0",
    "method": "converter",
    "instance": "conv-7f1",
    "lossy": ["thinking.signature"]
  },
  "authority": "local",
  "startedAt": "2026-10-26T14:02:11.204Z",
  "lineage": {"root": true},
  "scope": "primary",
  "workspace": {
    "roots": [
      {
        "id": "r1",
        "path": "/work/app",
        "vcs": {"type": "git", "head": "5d1c0e2", "branch": "main"}
      }
    ]
  },
  "profiles": ["core", "coding"],
  "contentLevel": "full"
}

{
  "type": "message",
  "seq": 3,
  "id": "r3",
  "messageId": "msg_01",
  "role": "user",
  "author": "human",
  "origin": "prompt",
  "audience": "both",
  "content": [{"type": "text", "text": "Fix the failing parser test."}]
}

{
  "type": "tool_call",
  "seq": 5,
  "id": "r5",
  "callId": "toolu_01A",
  "callIdSource": "native",
  "inferenceId": "inf_01B",
  "name": "Bash",
  "kind": "execute",
  "title": "Run tests",
  "input": {"command": "cargo test -p parser"}
}

{
  "type": "tool_result",
  "seq": 6,
  "id": "r6",
  "callId": "toolu_01A",
  "linkage": "exact",
  "status": "failed",
  "content": [{"type": "text", "text": "test parse_empty ... FAILED"}],
  "exec": {
    "command": "cargo test -p parser",
    "cwd": {"root": "r1", "path": "."},
    "exitCode": 101
  }
}

{
  "type": "vcs",
  "seq": 14,
  "id": "r14",
  "action": "commit",
  "root": "r1",
  "commit": "9b2e41c7d0a35f1e8c4b6a2d9f0e1b3c5a7d8e9f",
  "branch": "main",
  "evidence": "observed"
}

​

Why it exists

• the native logs of existing harnesses can be converted now, with losses stated and without vendor help;
• a harness can adopt it with little implementation work;
• an app that runs other agents and the agents it runs can each keep their own stream, linked so that work is counted once;
• a live ACP v2 stream can be recorded into it, with declared and tested losses;
• ATIF trajectories, OpenTelemetry spans and Agent Trace or Git AI attribution can be generated from it, instead of being captured separately.

​

How it works

• One stream per session. A stream is UTF-8 JSON Lines, one record per line. Line 1 is the session header. Records are immutable: a later fact or a correction is a new record.
• Native IDs are kept. A session is identified by its agent, an optional namespace and the harness’s own session ID, kept verbatim.
• A small required core, plus profiles. The core, coding, lineage, usage and messaging profiles are promises about what a stream records. A producer that claims a profile emits the corresponding record, or an explicit capture state, whenever an event covered by that profile happens. Coding is one of these profiles, and nothing in the core depends on it.
• Fitted to real tools. A field or record type is standard when at least two of Claude Code, T3 Code, Codex and OpenClaw record it. Other facts these tools record get a documented place in the agent registry, so nothing is dropped. There is no experimental tier.
• Gaps are stated. A missing field means unknown, and nothing is made up. Capture states, gap records and declared losses say what a record does not contain.
• Links need evidence. Lineage, copies and links between sessions are recorded only from evidence in the source, never inferred from similar text or timing.
• Room for vendor data. Vendor data goes under _meta with reverse-DNS keys. Vendor record types and enum values are namespaced, and every enum is open.

​

What it is for

1. Search and memory across harnesses and assistants.
2. Portable handoff: starting a new session, in any harness, with information from a prior record.
3. Review and audit: who asked for what, what started each turn, who or what approved each action, what was sent and what was written to memory.
4. Attribution and blame evidence: finding the session that produced a change.
5. Evaluation and training export.
6. Cost and usage analytics.
7. Replay viewers.

​

What it does not do

• A record does not restore processes, credentials, sandboxes or provider-private continuation state. An adapter may support running a session again in its original harness; the adapter’s capabilities are stated for each harness.
• Replaying a record for display never re-executes tools.
• ASR is not a live protocol and does not replace ACP. It does not replace telemetry or evaluation formats either; it can feed them.
• Some assistant features are not in 0.1, such as payments, on-behalf-of identity and live voice. The roadmap lists them.
• Conformance does not make a record safe to share. Records can contain source code, secrets and personal data. See Privacy and security.

​

Where to start