/cs-harness

Slash Command Source

Parse $ARGUMENTS: the first token is the domain (one of the 18 manifest names under engineering/agent-harness/skills/agent-harness/assets/harnesses/); the rest is the goal. If the domain token doesn't match a manifest file, list the available manifests and ask.

Sequence (gates are blocking — never skip forward)

1. Compile — python3 engineering/agent-harness/skills/agent-harness/scripts/goal_compiler.py --goal "<goal>" --manifest engineering/agent-harness/skills/agent-harness/assets/harnesses/<domain>.json --out .agent-harness/plan.json
2. Exit 3: relay the forcing questions to the user one at a time (recommended answer first), then recompile with the enriched goal. Do not proceed on a vague goal.
3. Exit 4: show nearest_candidates, ask whether to switch domain or refine the goal.
4. Review the plan with the user — show tasks, verifications, and caps. Confirm before initializing: this is the only approval gate in the loop.
5. Init — python3 .../scripts/loop_controller.py init --plan .agent-harness/plan.json --state .agent-harness/state.json
6. Drive — repeat: next → execute the task per its skill's SKILL.md → record → verify. For long goals, spawn the harness-runner agent per task instead of executing inline, one at a time (writes stay serialized).
7. On exit 2 or 5 — stop, show status and the failing evidence; the user decides: fix and continue, waive with a reason, or abandon.
8. Close — close --state .agent-harness/state.json; paste the handoff block (tasks, statuses, evidence, waivers) as the deliverable summary.

Rules

• Never edit checks, manifests, or the plan mid-loop to make verification pass.
• Never report an exhausted budget as success.
• .agent-harness/ is git-ignorable working state; the handoff block is the record.