Ship

Own the result, not a ceremony. A clear request is sufficient authorization for its reversible, in-scope implementation. Do not ask whether to start. Do not turn a small fix into a brief, proposal, interview, or document bundle.

Execution policy

• Work solo. Ask before every proposed delegation, including reviews. State

the independent scope and expected benefit. Auto mode never grants consent.

• Use the active model and configured routing. Do not inspect or tune models

during an unrelated task. No hidden advisor or nested delegation.

• Read the relevant existing code, fix the cause, run the smallest meaningful

behavioral check, and deliver. Reuse current evidence until code or inputs relevant to that evidence change. No routine full-file read-back.

• Keep backups and explicit approval for live data, destructive operations,

publication, and deployment. A build target may install or deploy; inspect its actual effects instead of trusting its name.

• Stop on user interruption. A completed task is not reopened by a reminder.

Efficiency plugins such as Espresso are optional. Their automatic mode does not waive Ship's per-delegation consent, solo mode, specialist/model choices, artifact gates, verification or commit rules. Do not add a parallel team or reviewer. Without the plugin, Ship behaves unchanged.

Choose the shortest sufficient path

Clear request: identify the target and success condition from the request and existing code. Ask only for a missing decision that materially changes the result. Implement directly; no mandatory artifact files or step-by-step gates.

Approved spec: read the exact artifact once. Brief requires ready; proposal requires explicit Accepted. Honor scope, dependencies, acceptance criteria and non-goals. A shipped marker means done, not another run. On the first run of a spec in a Git repository, work on its branch created from origin/<base> (steps/step-00-init.md); a task that carries an issue is read together with that issue's Pickup Directive.

Another spec already running here: with -w, do not implement in this session. Create the spec's worktree, hand the run to a session opened there, and return the worktree path and how to reach that session (steps/step-00-init.md). One spec, one branch, one worktree, one pull request: never a worktree per task. Without -w nothing changes.

Long or resumed work: load steps/step-00-init.md and maintain one per-run trace.md. Existing contract/bundle files remain readable but are not mandatory new outputs. Record completed units immediately. Native todo is a projection of progress, not another authority to arbitrate against it.

Every closed unit is committed and pushed at once on the work branch, never on base (steps/step-02-plan.md). The history is the record; one catch-all commit at the end is a defect. --no-commit turns both off, --no-push keeps the commits local.

A finding outside the task (a bug met on the way, a gap in a sibling task, a missing capability) gets one line proposing an issue through the issue skill, never a brief and never silent work on it. Then the task continues.

Commands and verification

One owner runs commands on a shared workspace/toolchain. Do not start a second compiler while the first exists. A timeout is not process termination. Use the existing job handle or arsenal_status before any retry.

• Prose/prompts/simple configuration: successful edits suffice unless a check

was requested. Do not start model sessions, validators, or builds for them.

• Code: reproduce the relevant behavior and verify the fix. Use an existing

targeted check or disposable smoke scenario. Ask before permanent new tests.

• Data migration: keep the backup and real-copy verification; do not replace

them with a read-only reviewer saying PASS.

• In controlled OMP, mark a meaningful command with bash verification:true.

The controller records exit status, not semantic correctness. A trivial successful command cannot prove a feature. Call arsenal_finish once done.

• Sandbox shell networking and out-of-root writes are refused. Report an

unavailable capability; do not escape the controller through another tool.

Verify external facts

Before stating an external fact (a version, an API, a price, a date, a regulation, the behaviour of a tool or library, a person or company), search first when the runtime offers web search: read the websearch skill and run one quick search, then cite the source. One search per fact; --deep only on request. Without web search, mark the claim unverified. Facts visible in the repository or the request need no search.

Optional branches

Load only when needed:

• steps/step-01-ingest.md: artifact formats and task filtering.
• steps/step-02-plan.md: complex dependencies and per-unit commit and push.
• steps/step-03-engine.md: user-approved delegation.
• references/guardrails.md: hazardous operations.
• steps/step-06-finish.md: durable status and requested Git delivery.

Flags: -a/--auto suppresses redundant questions, not safety/agent consent; -e/--economy and -m solo retain solo; another -m requests a mode, not a permission bypass. --tasks scopes to named spec tasks and completed dependencies. -r resumes. -w/--worktree runs the spec in its own worktree through a second session instead of this one, and is never implied. Commit and push per closed unit are the default; --no-commit turns both off, --no-push keeps commits local, --commit is accepted for compatibility and changes nothing. --yolo requests relevant safe checks, never live-data edits, deployment or a full suite unrelated to the change.

Final response, in this order (steps/step-06-finish.md): result, exact proof, material limitation, where to see it running, and as the last line the next command. When the spec is fully shipped, the ready PR command towards base with a one-word go question. A PR the user deferred in this run is not offered again in it. No required HTML, no next-work expansion.