Logging

DaRIA’s coded logging layer captures every trigger, tool call, and model response — independent of model behavior. The logs are the primary input for the Phase 2 fine-tuning pipeline.

Format

Append-only JSONL (one JSON object per line). One file per UTC day, rotated at midnight UTC.

File naming: daria-YYYY-MM-DD.jsonl

Location: <agent_directory>/logs/ (from AgentConfig.directory)

Example: ~/git/daria/logs/daria-2026-03-27.jsonl

Schema

Three entry kinds. Each entry has a top-level kind, a data object, and a timestamp.

trigger

What prompted a turn — a mention, supervisor nudge, or patrol event.

{
  "kind": "trigger",
  "data": {
    "source": "mention",
    "sender": "ori",
    "channel": "#code-review",
    "text": "@daria can you look at PR #14?"
  },
  "timestamp": "2026-03-27T14:32:01+00:00"
}

source values: "mention", "supervisor", "patrol"

tool_call

Every skill invocation — arguments and result.

{
  "kind": "tool_call",
  "data": {
    "skill": "irc_read",
    "args": {
      "channel": "#general",
      "limit": "50"
    },
    "result": {"ok": true}
  },
  "timestamp": "2026-03-27T14:32:04+00:00"
}

response

Every model output — the text and where it was sent.

{
  "kind": "response",
  "data": {
    "text": "PR #14 looks good — all checks pass and the retry logic is bounded. Approving.",
    "destinations": ["#code-review"]
  },
  "timestamp": "2026-03-27T14:32:09+00:00"
}

Hook Points

The three entry kinds map directly to hook points in daria/daemon/daemon.py:

Entry kind	Hook	Trigger
trigger	_on_mention()	Incoming mention or supervisor nudge
tool_call	_handle_ipc()	Every skill invocation
response	_on_agent_message()	Every model output

Schema Contract

This schema is the contract between Phase 1 (logging) and Phase 2 (pipeline). The Digest stage reads these files verbatim. Do not change field names or structure without updating the pipeline accordingly.

Any schema evolution should be versioned — add a "schema_version" field to entries if the format changes.