Deterministic and structured approach to slop cannoning.
plan → /slopship → clarify → human releases → machine runs
Require the binary
command -v slopshipper
slopshipper version
If missing, stop and ask the user to install the CLI through their approved host package or release workflow. Do not download or run installers from this skill. Do not invent a second runtime.
Bootstrap the run
slopshipper status --json --fields state,run_id,next_action,allowed_commands,required_evidence,intake_revision,required_reviewers,completed_reviewers,delivered_units,delivery_mode,blocker,decision_question
If status reports UNINITIALIZED, obey its slopshipper init next action.
If it reports multiple open runs, show their IDs and ask which one to resume.
Do not replace a blocked run. If status shows RUN_DONE and the user requested
new work, create a new run; otherwise report the completed run.
Turn the agreed plan into intake immediately after init. Keep units concrete,
bounded, and dependency-ordered; give each unit verifiable
acceptance_criteria and declare the run's risk_tier; use stdin so the
workflow does not leave evidence scratch files in the repository:
slopshipper intake --input - --dry-run --json <<'JSON'
{
"run": "run-id-from-status",
"delivery_mode": "pr-hold",
"required_reviewers": ["autoreview"],
"risk_tier": "low",
"series_bound": 1,
"units": [
{"id": "u1", "title": "Implement and verify the agreed change", "blockers": [],
"acceptance_criteria": ["the agreed behavior is proven end to end"]}
]
}
JSON
Inspect the dry-run projection. If it matches the agreed intake, repeat the
same command without --dry-run. Do this for every mutation: validate first,
then apply. A dry run never advances canonical state.
verify --cmd --dry-run cannot predict an exit code; when it returns
outcome_undetermined: true, validate the command and current state instead of
expecting a post-verification transition.
Show the human a compact intake summary: units, delivery mode, required
reviewers, and exact intake_revision. Wait for explicit release approval,
then run the exact slopshipper release --revision N command printed by
next_action.
Status leash
- Run the field-masked
slopshipper status --jsoncommand above and obey its next step. See status.md for the full status field contract. - Validate every mutation with
--dry-run --json; apply only when the projection matches intent. Successful--jsonoutput is the resulting status document. - Advance only through named CLI commands with the evidence status requires. Re-read status after errors or whenever output was not JSON.
- Treat placeholders in
next_actionas fields to fill, not literal text. Never materialize intake, review, delivery, or other evidence payloads in the repository; send them through stdin. - Before guessing a payload field or enum, run
slopshipper schema --command <name>. See protocol.md for raw input and error handling.
Machine loop
After the human releases: build → verify → review → deliver, always driven by status.
# after verify succeeds and status asks for review
slopshipper review --input - --dry-run --json <<'JSON'
{"run":"run-id-from-status","reviewer":"autoreview","verdict":"clean","artifact_ref":"autoreview://local"}
JSON
Deliver only after every required reviewer is present in completed_reviewers.
Use stdin evidence and match the intake delivery mode:
slopshipper deliver --input - --dry-run --json <<'JSON'
{"run":"run-id-from-status","delivery_mode":"pr-hold","pr_url":"https://github.com/example/repo/pull/1"}
JSON
After each projection is accepted, repeat without --dry-run.
Talk to the human
Collaborator voice: short prose + optional tables. Lead with what changed or what you need — never a wall of CLI JSON. Plain words over machine dumps.
Three human moments
- Release — confirm table (what/how/review), wait, then
slopshipper release --revisionusingintake_revisionfrom status JSON. - Required reviewers — once at intake: pick from the registered
identities (
slopshipper reviewerslists them;autoreviewandbugbotare built in). Store via intakerequired_reviewers. Do not auto-fire reviewers. - Decide —
slopshipper ask --question …, thenslopshipper decide --answer ….
Error recovery
Failed verification records BLOCKED and exits 6. Show the compact failure,
re-read status, surface blocker verbatim, and ask how to recover. After the
human confirms the recovery, record it with slopshipper retry --reason "…";
never retry silently. For a known external blocker before verification, use
slopshipper block --reason "…" and follow the same recovery rule.
With --json, failures return error.kind, error.message, and
error.exit_code. Malformed input exits 2; illegal transitions or unmet guards
exit 3. Fix the input or re-read status instead of bypassing the gate. Empty
next action means the run is done or needs human inspection.
Post-delivery babysit
Delivery opens a change request; the unit is delivered, not settled, and
later units already build while it waits. Record only what the forge really
shows, via slopshipper observe: merged settles the unit,
checks_failed and review_feedback pull it back through the build loop
with the cause recorded. Pass --unit when several units are delivered.
AWAITING_SIGNALS means every remaining unit waits on external signals;
keep watching the change requests and record signals as they land. Never
invent a signal.
Post-review flow
cleanrecords that reviewer. Delivery requires a distinct clean result from every identity inrequired_reviewers.findingsmoves directly toREWORK; summarize findings, then obey the next build command.ambiguousmoves toNEEDS_DECISION; ask the human and record the answer.
Mindful spend
Strong model for BUILD; cheaper review (Bugbot / lighter autoreview). Never default "most expensive everywhere."
Companion tools
Run whichever installed tool matches a required reviewer: the autoreview
binary, Cursor /review-bugbot, or the registered custom reviewer's own
surface. Never simulate a reviewer.
Done
Stop when RUN_DONE (every unit settled), blocked pending human recovery,
or waiting at release or decision. AWAITING_SIGNALS is not done — report
which change requests still wait. SQLite holds the canonical event log.