jayfarei/opentraces/pr-intent-summary-v1

pr-intent-summary-v1

Emits one structured row per commit on a branch, joining each commit to the agent traces that authored it. The row carries:

• the commit identity (sha, subject, body, merge flag)
• a lineage array of contributing traces, each with:
• the user's intent (Burst.intent.most_substantive_spec)
• a deterministic English summary of what changed

(trace_summary.summarize_contribution)

• a stable ot:// resource pointer + bucket path
• a no_lineage_reason for commits without trace coverage

This row shape is the endpoint. Any consumer can read it: the PR body renderer in core.branch_context, a future Slack notifier, a dashboard, a CI check.

Scope contract

Reads scope from run_packet.json. Expects scope.kind == "branch_context" with:

{
  "kind": "branch_context",
  "project_dir": "/abs/path/to/repo",
  "project_id": "...",
  "base": "main",
  "head": "HEAD",
  "merge_base": "abc1234...",
  "commits": [
    {"sha": "...", "short_sha": "abc1234", "subject": "...", "body": "...", "is_merge": false},
    ...
  ],
  "scope_digest": "..."
}

Row schema

See schemas/row.schema.json. One JSONL row per commit in scope.commits order.

Builder

python ~/.opentraces/workflows/pr-intent-summary-v1/scripts/build_rows.py

The script reads OT_RUN_PACKET for the run packet and writes rows to OT_DATASET_OUTPUT. Core lineage fields are deterministic. Optional LLM polish is on by default for the rendered summary; set OT_LLM_POLISH=0 when a byte-identical deterministic row stream is required.

Consumers

• opentraces trail blame pr render | create | update — renders the row stream

as a GitHub PR body and (for create/update) pushes via gh.

Future consumers can read the same JSONL and project to other destinations without changing this workflow.