minhtran3124/agent-harness/visual-planner

Plan → HTML Renderer

The executable interface is authoritative; never hand-transcribe HTML. Resolve <visual-planner-dir> to the directory containing this loaded SKILL.md:

python3 <visual-planner-dir>/render_plan.py <path-or-slug> [output.html]

The renderer resolves a PLAN path or unambiguous slug, writes a sibling untracked PLAN.html, and performs self-checks. Relay its stdout; on failure report its SELF-CHECK FAILED: output without claiming success. template.html controls layout and render_plan.py owns parsing.

--summarize first injects or refreshes the tracked At-a-glance block in PLAN.md (between the generated AT-A-GLANCE:BEGIN/END sentinel lines; its ### Progress checklist is derived from ## Status Log), writing only on change, then renders as usual. writing-plans runs it after saving a plan, rules/wave-parallelism.md runs it at each wave boundary, and finishing-a-development-branch runs it before the shipped commit.

Use view_plan.py <slug> --file|--no-open|--port N|--render only when visual viewing is requested; its serving mode blocks. Do not auto-open from writing-plans or a headless environment.

Optional review overlay

For graph-backed plan review, first emit plan files, gather graph evidence, then render a sidecar:

python3 <visual-planner-dir>/render_plan.py <slug> --emit-files
python3 <visual-planner-dir>/render_plan.py <slug> --review specs/<slug>/.plan-review.json

The sidecar records provenance plus per-file path, status (new|existing|missing), dependents, tests, risk (high|medium|low), note, and optional flows. New/missing graph nodes remain explicit; do not invent risk or coverage. Review output defaults to untracked PLAN.review.html.

Both XML and canonical markdown task forms are supported. Keep parsing and rendering behavior in the tested code; re-run the renderer's unit tests in the harness repo after changes.

References

• references/review-sidecar.md — sidecar schema and evidence rules