Prj Orchestrator

prj-orchestrator

Overview

This skill is the cron heartbeat for the PR Jangler module. Each invocation executes exactly one state transition: load the persistent queue at _bmad-output/pr-workflow/state.json, pick the highest-priority next action, dispatch to the target phase skill, persist state, append a structured run-log entry, exit. Idempotent with atomic state writes. System actions (discovery on cadence, daily report at the configured hour) outrank per-PR actions.

Act as a reliability-minded ops engineer. Treat the queue as authoritative state, the run-log as the audit trail, and "never silently fail" as the prime directive.

Conventions

• Bare paths (e.g. scripts/run.py) resolve from the skill root.
• {project-root}/... resolves from the project working directory.
• Configuration variables live under [modules.prj] in {project-root}/_bmad/config.toml.

On Activation

Load configuration from {project-root}/_bmad/config.toml (the [modules.prj] block). Required key: prj_repo. If absent, exit early with a clear "configure prj_repo first" message and a run-log entry tagged misconfigured. Defaults for optional keys are applied by the scripts themselves.

Execute the heartbeat:

python3 scripts/run.py

That single script orchestrates everything: state initialization on first run, action selection, dispatch (with v1 stub fallback for not-yet-built phase skills), atomic state persistence, and run-log append. Flags:

• --dry-run compute next action, log the intent, do not dispatch
• --verbose emit diagnostics to stderr
• --once no-op flag for cron clarity (always single-shot)
• --project-root <path> override autodetect

See python3 scripts/run.py --help for full detail.

Architecture

The orchestrator is intentionally thin. All deterministic work lives in Python scripts. The LLM activation context exists so the Skill tool can dispatch to phase skills once they are installed; in v1 stub mode, the dispatch path falls back to bmad run subprocess invocations or logs a stub: not built entry when the target skill is missing.

State Locations

Owned by this skill. All other phase skills MUST go through scripts/state_io.py to read or write state to preserve atomicity.

• Queue: {project-root}/_bmad-output/pr-workflow/state.json
• Per-PR cache: {project-root}/_bmad-output/pr-workflow/prs/{pr_number}/
• Run-log (daily, append-only): {project-root}/_bmad-output/pr-workflow/logs/{YYYY-MM-DD}.jsonl
• Reports archive: {project-root}/_bmad-output/pr-workflow/reports/{YYYY-MM-DD}.md

Priority Logic

System actions outrank per-PR actions.

1. Daily report (priority 700): local hour >= prj_report_hour_local AND no report sent today.
2. Discovery on cadence (priority 500): queue is empty, OR heartbeat_count % prj_discover_every_n == 0.

Per-PR actions, in descending priority:

3. PleaseAdvise unacknowledged (priority 1000, outranks everything): rises until the user-acknowledged override label is applied.
4. Per-PR next_action set by the previous phase skill. Base 100, with bumps:
   • new_comments_since_triage > 0: +50
   • needs_retriage == true: +30
   • Age tiebreak: +1 per day idle since phase_entered_at, capped at 30.

Terminal phases (ReadyToMerge, Blocked, Rejected, Archived) carry no next_action and are skipped. If no action qualifies, the heartbeat is a no-op: log idle, exit cleanly.

v1 Stub Mode

The other 10 phase skills (prj-discover, prj-triage, prj-verify-claim, prj-review, prj-detect-overlap, prj-plan-fix, prj-implement-fix, prj-validate-adversarial, prj-decision, prj-report-daily) do not exist yet. When the orchestrator dispatches to a target skill whose SKILL.md is not present at {project-root}/skills/{skill-name}/, it:

• Records status: "stub: not built" in the run-log entry
• Does NOT transition state (the PR stays in its current phase)
• Exits cleanly with code 0

This lets the orchestrator be exercised end-to-end before any other skill lands.

Non-Negotiables

• Idempotent per action. If a heartbeat is interrupted mid-run, the next heartbeat picks up cleanly. Atomic writes via tmp + fsync + rename.
• Never silently fail. Every outcome (dispatch, stub, abort, error) emits a structured run-log entry. Errors also surface on stderr.
• No PR content inspection here. The orchestrator reads queue metadata and dispatches. Code review, classification, and verification belong to the phase skills.

Verification

After install, smoke-test with:

python3 scripts/run.py --dry-run --verbose

Expected on a fresh project (no state, prj_repo configured): exit 0, run-log entry showing action: dispatch, skill: prj-discover, status: dry-run. On a project missing prj_repo: exit 2, run-log entry status: misconfigured.

Run unit tests:

python3 -m unittest discover scripts/tests