Skip to content

complete-prd.md

Latest commit

 

History

History
58 lines (45 loc) · 4.42 KB

complete-prd.md

File metadata and controls

58 lines (45 loc) · 4.42 KB
Raw

Finish a Ralph loop run: commit any remaining changes and open a pull request.

Arguments: $ARGUMENTS — either the spec index (e.g. 45), bot:<featureId> for a bot-feature run (e.g. bot:opinionTargeting), or perf:<target> for a performance-optimization run (e.g. perf:CountryPopulationCollector).

Steps

  1. Resolve the spec folder.

    • Spec-index form: Docs/Specs/<index>_<name>/.
    • bot:<featureId> form: the standing spec Docs/Specs/52_bot-feature-eval-harness/ (bot features have no per-feature spec/plan — see the Constitution's bot-feature carve-out).
    • perf:<target> form: the standing spec Docs/Specs/26_07_18_18_benchmarkdotnet-perf-harness/ (perf-optimization attempts have no per-target spec/plan — see the Constitution's performance-optimization carve-out).

    Read .ralph/prd.md and .ralph/activity.md. Determine run status: how many tasks have "passes": true out of the total, and which are "category": "unity-manual".

  2. Run git status (never -uall) and git diff. If uncommitted changes remain, stage them and create a commit using the /commit skill (it handles branch rules and the version bump). If the working tree is clean, still run the /commit version-bump flow only if no commit exists on this branch yet; otherwise proceed.

  3. Create the pull request using the /pr skill. Create it as a draft when any PRD task is unfinished or the activity journal records a blocker; create it ready for review only when every task passes and no blocker remains. In addition to the /pr skill's standard body format, append a section:

    ## Ralph run
- Spec: Docs/Specs/<index>_<name>/
- Tasks passed: X/N
- Manual Editor verification needed: <list of unity-manual tasks, or "none">
- Unfinished tasks / blockers: <from activity.md, or "none">

    Unchecked test-plan items must include every unity-manual task.

    bot:<featureId> form only — follow the Ralph run section with an ## Eval verdict section sourced from Docs/BotFeatures/<featureId>/eval_history.json (latest entry) and eval_summary.md:

    ## Eval verdict
- Gates: score=<pass/fail>, command-on=<pass/fail>, command-off=<pass/fail>
- Mean delta: <mean> (median <median>) vs epsilon <epsilon>
- Improved (mean > 0): <yes/no>
- Winning parameters: <winner's parameter set, or "no search">
- Attempts: <count>
- Details: Docs/BotFeatures/<featureId>/

    perf:<target> form only — follow the Ralph run section with a ## Benchmark verdict section sourced from Docs/Benchmarks/history.json (latest entry/entries for the target) and Docs/Benchmarks/summary.md:

    ## Benchmark verdict
| Benchmark | Baseline mean | Current mean | % change | Verdict | Allocated bytes |
|---|---|---|---|---|---|
| <name> | <baseline ns> | <current ns> | <%> | <pass/fail> | <bytes> |
- Attempts: <count>
- Details: Docs/Benchmarks/

  4. After the pull request exists, clear the completed run artifacts: replace .ralph/prd.md with a short # Ralph PRD header stating that no run is active, reset .ralph/activity.md to its standard header only (# Ralph Activity Journal, its intro line, and ---), and delete every file under .ralph/logs/ (per-iteration loop stdout/stderr logs — not needed once the run's outcome is captured in the PR). Stage and commit this cleanup as a separate follow-up commit using the /commit skill, then push it to the same pull-request branch. Do not combine this artifact cleanup with the implementation commit or the PR-creation commit.

  5. Report the PR URL and the follow-up cleanup commit.

Rules

  • Do not flip any "passes" flags here — this skill reports state, it does not verify tasks.
  • Do not refuse to create a PR merely because the Ralph run is incomplete. A partial PR preserves verified progress and makes the remaining tasks and blockers reviewable.
  • If the loop left blockers in .ralph/activity.md, quote them in the PR body verbatim rather than paraphrasing them away.
  • For bot:<featureId> runs: never fabricate the eval verdict section — read it from the committed eval_history.json/eval_summary.md; if those files are missing or empty, say so rather than inventing numbers.
  • For perf:<target> runs: never fabricate the benchmark verdict section — read it from the committed Docs/Benchmarks/history.json/summary.md; if those files are missing the target's entries, say so rather than inventing numbers.