distill-pr-intent-orchestrator

Distill PR Intent — Orchestrator

Goal

Run distill-pr-intent across a selected PR set, and persist:

primary memo text (.txt)
sidecar metadata (.meta.json)

All outputs go to shared memory on EFS, and are automatically published to the public S3 bucket by the host timer.

Inputs

repo: fixed openclaw/openclaw
selection: one of:
- last N (e.g. last 500)
- range <start..end>
- since <YYYY-MM-DD>
skill_version: string (default v1)

Output paths (persisted)

/memory/pr-intent/<skill_version>/openclaw-openclaw/<PR>.txt
/memory/pr-intent/<skill_version>/openclaw-openclaw/<PR>.meta.json

Determinism rule

NEVER put timing/telemetry/debug into the primary .txt memo.
All telemetry goes into .meta.json.

Workflow

Determine PR list.

Examples:

# last N merged PR numbers
gh pr list -R openclaw/openclaw --state merged --limit 200 --json number --jq '.[].number'

# range
seq <start> <end>

For each PR (sequential by default):

Run distill-pr-intent <PR>.
Write memo to /memory/pr-intent/.../<PR>.txt.
Write metadata to /memory/pr-intent/.../<PR>.meta.json.

Writing files (memory locking)

Use the locking helpers (required on EFS):

# memo
printf '%s\n' "$MEMO" | memory-write "/memory/pr-intent/<skill_version>/openclaw-openclaw/<PR>.txt"

# meta
printf '%s\n' "$META_JSON" | memory-write "/memory/pr-intent/<skill_version>/openclaw-openclaw/<PR>.meta.json"

Suggested metadata schema

{
  "pr": 17392,
  "repo": "openclaw/openclaw",
  "state": "merged",
  "skill_version": "v1",
  "generated_at": "2026-02-15T19:00:00Z",
  "patch_mode": "PATCH_OK",
  "patch_bytes": 12345,
  "model": "openai/gpt-5.2-codex",
  "error": null
}

Failure handling

On failure, still write .meta.json with error populated.
Memo file optional; if written, keep it short and point at .meta.json.