Docent pr

docent-pr — the PR-review film

You are running the entire docent cascade in PR-review mode against a pull request the user named: survey → treatment → spec → tts → clips → render → open. The output is one MP4 that explains the change the way a principal engineer would: why it exists, whether the design is sound, the core before → after, what ripples, what could break, and a verdict.

Arguments

/docent-pr <repo> <pr-number> [--id X] [--scale S] [--no-open]

• <repo> — a repo path, a GitHub URL (https://github.com/owner/name), or the bare owner/name form.
• <pr-number> — the pull-request number to review. Can be 123 or #123. If the user pasted a full PR URL as <repo>, extract both.
• --id X — override the auto-derived film id (default: <repo-name>-pr).
• --scale S — render scale. Default 1. Pass 0.5 for fast turnarounds.
• --no-open — render without opening the result in the system player.

What to do

1. Pre-flight. Confirm bun, ffmpeg, and the agent CLI (claude or codex) are on PATH. If anything obvious is missing, suggest /docent-doctor and stop.

2. Survey — PR mode:

   docent survey <repo> --mode pr --pr <n> [--id X]
   

   The survey lands at analysis/<id>.md. Surface the path and the load-bearing finding (section 1 of the survey — the triage) before moving on. A PR survey's job is triage: of the diff, what is the load-bearing 5% — that you review with depth — and what is the rest you explicitly deprioritized.

3. Commit to a style. Before the treatment, decide the visual register the film renders in:

   bun run docent style recommend <id>
   

   For PR films the recommender will almost always return engineering — code-heavy, the dark console look. Take the recommendation unless the survey surfaces something exceptional (an exec-level retrospective of a release-blocking PR, for instance, may want executive). Surface the choice to the user in one line — "rendering in engineering — the load-bearing 5% is in pkg/scheduler/...; rationale: file-level code review." — and move on. The spec compilation reads this off the survey's "Style commitment" section and pins it as the spec's style: {preset, intent, rationale} field. The depth-review judge fails the style-committed dimension if the spec ships without it.

3b. Commit to a scene set. Same shape as the style commitment, one layer down — the cognitive moves the film will make.

bun run docent scene-fit recommend <id>

The recommender reads the survey and prints the top scene types with rationales tying each to a specific survey finding. PR films almost always include diff (the change) and closeup (the load-bearing function annotated); tension is also load-bearing for the trade-off. If the recommender returns warningOnDefault: true, the survey collapsed to the default rut (frame/structure/compare/tension/recap) — re-read the triage and ask whether the load-bearing 5% actually wants mechanism, walkthrough, causal-loop, timeline, or something more specific. Pin the chosen scene set in the survey's "Scene-set commitment" section.

4. Treatment.

   docent treatment <id>
   

   Writes treatments/<id>.md. Print the Angle line so the user sees the through-line you committed to before the spec is compiled.

5. Spec — and interrogate it.

   docent treatment <id> --to-spec
   docent review <id> --max-rounds 2
   

   The first compiles the treatment into films/<id>.json. The second is mandatory — review runs the adversarial judge → revise → re-judge loop bounded to two rounds. On the corpus this reliably lifts a first-draft spec by ~7 points / 30 — the difference between a PR-review film that passes the depth contract and one that does not. Surface the verdict score and the weakest dimension before rendering. If review exhausts its round budget, stop and ask — do not ship a film the judge rejected.

6. Render.

   docent build <id> --scale 1
   

7. Open the result (unless --no-open). On macOS: open out/<id>.mp4.

8. Hand back. Three things to the user:

   • the film id (so they can re-render via /docent-build <id>),
   • the verdict score (e.g. 26/30),
   • one sentence of the verdict — what changes, what the residual risk is.

Knowing when to stop and ask

Pause and ask the user instead when:

• The survey surfaces two plausible triages — the diff is too large to cover and you have to pick what to deprioritize.
• The PR is a refactor with no clear functional verdict; ask if the user wants the film to focus on the design trade-off rather than a yes/no ship decision.
• review exhausts its round budget without passing the depth contract. Surface the failing dimensions; do not silently ship a film the judge rejected.

Failure modes

• No GitHub auth — survey --mode pr fetches diff / PR body / review threads via gh. If gh auth login is needed, surface the message and stop.
• Agent CLI missing — survey and treatment shell out to claude or codex. If neither is on PATH, suggest /docent-doctor.
• The judge keeps failing the triage dimension — that is a survey problem, not a render problem. The survey did not pick a load-bearing 5%; ask the user to point at the part of the diff that matters.