verify-auto-work

Files (8)

• SKILL.md
• README.md · features/
• agents.md · features/
• continuity.md · features/
• delivery.md · features/
• review.md · features/
• start.md · features/
• fixture.mjs · scripts/

SKILL.md (first lines)

# Verify Auto-Work

Use this repository-only verifier for an expressly commissioned implementation or verification assignment. Read [the feature index](features/README.md) and only the selected feature recipes. Package checks, content walkthroughs, native runs, and local delivery simulations are different evidence.

From the Workflow repository, build once with `rtk npm run build:targets`. Prepare an owned fixture with `rtk node .agents/skills/verify-auto-work/scripts/fixture.mjs prepare`. Its JSON returns the actual base, workspace, evidence directory and goal. Inspect with the same driver using `inspect BASE`; confirm the baseline and built package exist before starting an agent. The fixture has no persistent service, ports, login, or production dependencies.

Start a fresh native subject in the returned workspace with the absolute built package's `skills/auto-work/SKILL.md` path. The subject coordinates native Plan, Implement, Review and Correct children under its task-local selection. Keep the outer observer out of phase judgments; it may supply commissioned trial-human selections, plan approval and result acceptance in the native task. Where using an installed CLI, inspect its help first and use supported workspace sandbox, workspace selection and structured output options. Supply the skill path directly; do not install or activate a plugin. Record actual tools and versions: CLI availability does not establish authentication, delegation or editor capability.

For native Codex trials, record the actual CLI binary path and version and check compatibility with the selected model. If several installations exist, compare the shell-resolved CLI with any existing app-bundled CLI before changing models or installing an update.

Give the subject only the real assignment, fixture and built instructions, not expected findings or the observer's oracle. Make the workspace and fixed built candidate its project read boundaries, also in reviewer handoffs. For the shipped handoff procedure, explicitly commission an additional owned `BASE/phase-transfer` directory for task-local assignment/raw-output copies, with coordinator/child file ownership in the assignment. It contains no observer oracle/baseline/transcripts and is distinct from excluded `BASE/evidence`. Preserve these artifacts alongside evidence after cleanup. This explicit trial permission is not automatic product permission.

Apply boundaries from the first tool call, including cwd/search roots; searching the fixture parent traverses excluded evidence. The subject may edit workspace and assigned transfer files, never verifier, built package, observer evidence or driver. Verify actual instruction reads against the candidate; reading observer recipes or another plugin copy contaminates acceptance. Record competing installed instructions; a prompt does not enforce isolation. Save subjects/reviewers, actual calls/results and snapshots under evidence. Observe separate reviewer identities and no workspace writes during Review. Without native delegation, report the gap; observer review cannot count as native Auto-Work success.

The driver simulates an external gate under the observer's ownership. Only the observer invokes `deliver BASE REVIEWED_DIGEST authorized-local-simulation`, and only for a commissioned local simulation after examining the actual review evidence. This demonstrates gate behavior and candidate binding, not a production authorization boundary against a malicious process. Without verified real project enforcement, the subject must not claim a real Dark delivery is available.

After each drive, inspect actual files and outcomes against the selected recipe. On failure, inspect fixture health and source/package drift before fixing only commissioned product or verifier defects. On macOS, a `sandbox_apply` failure in a nested CLI trial can come from the outer/inner sandbox setup; investigate that setup before retrying commands or treating the failure as a fixture defect. Preserve effective protection boundaries while resolving it. Use the same observed failure to choose a targeted rerun; do not tune the expected result to accept it. Cancel only owned native agents if needed. Run `cleanup BASE` after saving evidence: it removes the owned workspace and retains evidence. Verify the reported evidence directory still contains the raw transcript and final snapshot. Never kill by process name or remove arbitrary temporary directories.

When the change affects selection, native invocation, approval, acceptance, correction, review delegation or delivery, run Light and Dark with a seeded-defect correction and fresh re-review. For agent/model selection changes, exercise both Codex and Cursor with two different available models and supported settings under [agent selection](features/agents.md). Otherwise run affected recipes. Record exercised and unexercised features/hosts, raw native calls and the limits of execution metadata. Finish with results, evidence paths, cleanup, limitations and open learning collection. Trials are not Workflow Review of the plugin change; that remains separately commissioned unless already covered by commissioned Auto-Work.

From geldmacher/workflow, MIT.

More from geldmacher/workflow