Prose Test

Prose Test

Run prose-test cases (design: design/prose-tests.md; runner: tests/prose/run.cjs). Each case is dispatched to a prose-orchestrator agent that owns the whole run — world, walk, delta, assertion, cleanup — and returns only a verdict. Transcripts never enter this session.

Walks cost tokens: this skill runs on command only, never as part of a routine gate.

Step 1: Select

Parse the arguments:

• Case ids (comma or space separated) → node tests/prose/run.cjs select --cases <a,b>
• --all → node tests/prose/run.cjs select --all
• --diff <ref> → node tests/prose/run.cjs select --diff <ref>
• No arguments → node tests/prose/run.cjs select --diff main

If the selection is empty: report that no cases intersect and stop. Otherwise show the selected case ids and proceed.

Step 2: Dispatch one orchestrator per case

Cases are independent — dispatch up to 4 concurrently, each a prose-orchestrator agent whose prompt is the case id and the instruction to run it end to end.

Never run the walker or asserter yourself, and never read a case's assert.md — the orchestrator owns that boundary, and anything you learn about an expected result can leak into a later dispatch.

Step 3: Collate

Report a verdict table from what the orchestrators returned: case id, model, verdict, deterministic checks, path steps passed, world, scope, markers. Quote the evidence line for every failure, and the archive path for every FAIL, FLAKY, or INVALID. Carry the scope line even when it is none undeclared — it feeds the case's file ratchet — and carry markers with their type prefixes intact (WALK:/WANDER:/INERT:/ NOTE:); INERT: entries may be collapsed to a count in the table.

Carry the checks through as reported. They were decided in code before any agent saw the case, so they are the one part of a verdict that rests on nothing's judgement — summarising them, or taking a green overall verdict as proof they passed, puts an assumption back where a fact was.

The model column comes from the harness record, not from any agent's say-so. An edited agent definition does not reach a running session until its plugins are reloaded, so a run can silently use the previous model — if the column disagrees with what the agent definitions declare, say so and treat every verdict in the run as unproven.

A FLAKY verdict means a failure did not reproduce on a second run from a fresh world — surface both runs, resolve neither.

UNSCRIPTED QUESTION, AMBIGUOUS and DEVIATION markers are findings in their own right; carry them into the report even when every case passed.

Rules

• Never run cases from a worktree-isolated session. The harness's worktree-isolation guard refuses the engine's complete-family subcommands (task complete, topic complete, workunit complete) inside the walkers' temp-dir worlds, failing every case on environment rather than prose — and pushing walkers toward guard-bypass workarounds the record then contradicts. Run from the primary checkout; exit any worktree first.
• Never run cases against prose that is still under review. While a PR's diff is moving — the author iterating, the reviewer commenting — a walk certifies bytes that may not survive the review, and each further change invalidates the verdict and forces a paid re-run. Runs come after the review settles: verify once, on the shape that will land.
• Never fix anything during or after a run. Not the prose, not a case, not a snapshot, not the harness — however small, however obvious, however certain you are. Every failure, every flake, and every pass carrying notes is surfaced to the user, and what to do about it is decided together. A test that repairs what it finds destroys the record of what it found, and the first fix reached for is regularly aimed at the wrong layer: correct prose has been called defective, and a case has been blamed for a fault in the judging.
• A pass with notes is not a clean pass. Markers are findings in their own right and always warrant a look. A walker that wandered and corrected itself did not fail the prose — but the same wander recurring on the same step is prose that reads misleadingly however correct it is, and that is only ever visible if every one is surfaced.
• A failing case is a finding either way: broken prose, or a stale case. Say which is suspected; the user decides.
• Never regenerate a snapshot as part of a run.