Skip to content

iCoT Transcript Format

iCoT writes an ignored local transcript to <example>/.icot/transcript.json unless --no-transcript or --print is used. The transcript is local review/debug evidence; it is not required for openudon build, and it should not contain secrets.

{
  "version": "openudon.icot-transcript.v1",
  "time_utc": "2026-05-23T00:00:00Z",
  "turns": [
    {
      "label": "Workflow goal",
      "answer": "Get weather in Toronto and send the report using Google Gmail."
    }
  ],
  "events": [
    {
      "kind": "catalog_plan_call",
      "data": {
        "candidates": ["gmail:google-discovery/gmail-discovery-v1.json"]
      }
    },
    {
      "kind": "draft_flow_review_result",
      "data": {
        "issues": [
          {
            "severity": "warning",
            "code": "llm_flow_review_disconnected_report",
            "message": "Gmail does not consume the report body.",
            "slot": "steps.gmail.with.raw",
            "evidence": "raw is not bound to the render step"
          }
        ]
      }
    }
  ],
  "session": {
    "project": {},
    "intent": {}
  }
}

Fields:

  • version: transcript wire name; current value is openudon.icot-transcript.v1.
  • time_utc: write time in UTC.
  • turns: local prompt labels and answers, including automatically accepted defaults in normal and fast mode.
  • events: structured stage observations such as catalog planning, request-mapping drafts, operation-detail requests, confirmation edits, advisory flow-review results, and bounded --review-repair attempts. Flow-review results may include gap_kind, remediation_action, and clarifying_question; forced ambiguity prompts are recorded as draft_flow_review_question.
  • session: the final iCoT session snapshot used to render project.md and intent.hcl.

Transcript events may include operation IDs, API document paths, symbolic credential binding names, and LLM warning evidence. They must not include credential values or pasted secrets.

When present, session.decision_evidence records compact decision history with stage, slot, value, source, confidence, reason, and optional alternatives. This is public review context, not private model reasoning. Replay reports may also count draft_repair_attempt, draft_repair_rejected, and the final draft_flow_review_result issues for M28 evaluation. M29 flow-remediation evidence uses the same public decision history and does not include hidden chain-of-thought.