Runtime, Storage, and Recovery

The Current Persistence Core

The current recovery model is still centered on the subagent runtime, not on world state.

The most important files are:

• subagent_runs.jsonl
• subagent_events.jsonl
• threads.jsonl
• agent_messages.jsonl

Together they record:

• subagent execution
• runtime events and retries
• agent thread relationships
• internal message flows

File Location

These files are currently written under:

workspace/agents/runtime/

subagent_runs.jsonl

Stores run-level records such as:

• run id
• agent id
• task
• status
• output
• created / updated timestamps

This is the basis for recovering in-flight work and auditing outcomes.

subagent_events.jsonl

Stores runtime events such as:

• spawned
• running
• completed
• failed
• retry counts

It is the direct trail for understanding why a subagent succeeded or failed.

threads.jsonl

Stores agent thread metadata such as:

• thread_id
• owner
• participants
• status
• topic

That is what makes internal collaboration a thread model instead of plain text forwarding.

agent_messages.jsonl

Stores the actual message flow, such as:

• from_agent
• to_agent
• reply_to
• correlation_id
• requires_reply
• content

This is also one of the reasons the WebUI can reconstruct internal collaboration streams.

Sessions And Memory

Outside workspace/agents/runtime/, two other persistence areas matter:

• ~/.clawgo/sessions/
• workspace/memory/

The first is more about primary session history. The second is more about:

• heartbeat.log
• skill-audit.jsonl
• nodes-dispatch-audit.jsonl
• trigger / process / node audit data

Why This Design Matters

Recovery here does not mean “restore a chat transcript.” It means restoring:

• which subagent was running
• what stage it reached
• where the internal thread was
• what the last output or error looked like

That is one of the key differences between the current Agent Runtime and a regular chat shell.