slopstop

One adversarial round against a target artifact — attack it for gaps against its stated goals, verify every claim in it against the real repo, and return numbered findings with severity plus a PASS / FAIL / GOAL DEFECT verdict the caller can branch on.

Push every file in a ticket's tracking directory to the ticket — task plan, findings, the run.jsonl timing log, adversary rounds — so the local record survives where the ticket lives. Bytes move from disk to the backend without passing through the worker. Reports what it pushed; moves nothing and deletes nothing.

Run the cyclomatic-complexity gate over a branch diff with lizard and return every function at or over the configured warn/reject thresholds — file, line, measured CC, its CC at the base commit, the threshold it broke, and whether the did-not-get-worse exemption applies — plus one overall verdict and a ranked list of what was exempted. Mechanical measurement only; never fixes anything.

Publish drafted tickets to whichever ticket system the project uses — create each issue, assign its key, link parents to children, and resolve cross-reference placeholders. The only place backend-specific creation lives; callers pass a draft and get back a letter-to-key map.

Stage 1 of the slopstop process — grill the user to shared understanding, then write the PRD and feature charter into the run dir and stop at gate G-design. Huge-tier only. Invoke as /slopstop:design <topic>.

Mirror the design/ directory to the project's ticket-system documentation store (GitHub wiki, Linear Docs). One-way push; design/ files unchanged; orphan pages pruned. Reads .project-conf.toml for the backend. Use /slopstop:doc-sync.

Run the duplication-check gate over a branch diff with ast-grep and return every clone group at or over the configured min-lines threshold — file, lines, SLOC, hash, scope (intra/cross-file), whether the clone group was pre-existing at the base commit, and a suggested helper signature showing which identifiers vary across instances — plus one overall verdict. Mechanical measurement only; never fixes anything.

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use /slopstop:grill to stress-test a plan — typically before breaking it into tickets.

Implement a ticket's plan until its failing phase-0 tests pass — writes source code, may add tests but never weakens, retargets or removes one, and returns the changes made, before/after test results, and any findings it is reporting rather than fixing.

Map the codebase for one ticket and return structured investigation findings — relevant modules, entry points, dependencies, conventions to honor, risks, and a predicted file map — as the worker's result, writing nothing to disk.

Verify that a set of freshly-written tests is pinned to the behavior it names — that a failing test is red for the RIGHT reason, or, under --backfill, that a passing test goes red when the behavior it claims to cover is broken. Returns a per-test verdict with evidence plus one overall PASS / FAIL / PINNED / NOT PINNED.

Write the phase-0 failing tests that define a ticket's contract before any implementation exists, run them, and return the test files, node-ids, test command, and the observed failure output proving they are red.

One round of clean-context code review of a diff — find, verify each finding against the real code, apply what survives, report a verdict. Runs in its own forked context so the session that wrote the code never reviews it.

The single lifecycle entry point — take one or more tickets and drive each through its whole lifecycle (investigate, red tests, adversary, implement, gates, review, PR, merge, archive), interleaving them, launching workers for judgment work and doing every mechanical step inline. Invoke as /slopstop:run <TICKET> [TICKET...].

Inspect a branch diff for AI slop — tests rewritten to pass, assertions inverted, vacuous or tautological tests, swallowed errors — and return each finding with its location, the signal that identified it, a severity, and one overall verdict. Reports only; never fixes.

Stage 2 of the slopstop process — read the PRD + charter from the run dir, cut the umbrella/leaf ticket tree per the five-section standard, drive the adversary loop over it, write the approved tree to the ticket system, and stop at gate G-tickets. Invoke as /slopstop:tickets <run-id>.

Mechanically prove whether each named test would already have passed against the pre-branch code by re-running it, by node-id, in a scratch worktree checked out at the base commit. Returns a per-node-id verdict (vacuous / meaningful / could-not-determine) with the exit status as evidence, plus one overall verdict.