chuzom/loophole

CI for AI agents — a swarm works until a falsifiable verifier PROVES the goal is done. Bring your own agent; loophole is the trusted gate.

Trust Signals

name	description	required	default
contract	Path to the Goal Contract (from `loophole init`). Ignored when `goal` is set.	no	loophole.json
goal	Ad-hoc goal text — an alternative to --contract for a one-off check. Requires `verify`.	no	""
verify	Hard verifier command for an ad-hoc `goal` (exit 0 = done).	no	""
workspace	Directory the swarm works in.	no	.
executor	A registered framework adapter (see `loophole executor list`).	no	""
executor-command	Bring-your-own executor CLI, e.g. 'claude -p {task}'. Sandboxed either way; the verifier boundary is unchanged.	no	""
planner-model	provider:model for planning.	no	chuzom:moderate
executor-model	provider:model for execution.	no	chuzom:complex
max-rounds	Round circuit-breaker.	no	25
version	loophole-agents version to install (pip version spec). Empty = latest from PyPI.	no	""
fail-on	'not-done' (default) fails this step when the run isn't verified done; 'never' always exits 0 so the workflow can branch on the outputs instead (report-only mode).	no	not-done
comment	Post/update a sticky PR comment with the Residual-Risk Report, and create a Check Run annotated with any write-allowlist violations. Only meaningful on pull_request-triggered workflows; requires `permissions: { pull-requests: write, checks: write }` in the caller's workflow (the default GITHUB_TOKEN otherwise lacks these scopes).	no	false
github-token	Token used for --comment. Defaults to the job token.	no	${{ github.token }}

name	description
status	'done' \| 'paused' \| 'failed'
verified-done	'true' \| 'false'
exit-code	loophole's own exit code: 0 = done, 1 = not done, 2 = config error
result-json	Absolute path to the full machine-readable run result.