prompt-ship

name	prompt-ship
argument-hint	<prompt-path | inline body> | --resume <branch>
description	Use to run a free-form prompt through the shared worktree ship pipeline without creating kata issues. Accepts a prompt path or inline body. Trigger for ship this prompt or $prompt-ship.

prompt-ship

Consumer of worktree-ship-pipeline. Read that file first — this skill specializes the work source, identity, lock, and close step. Every section cited as §<anchor> below refers to worktree-ship-pipeline SKILL.md.

Usage

/prompt-ship <path/to/prompt.md>
/prompt-ship <inline prompt body — multi-line, @-file refs ok>
/prompt-ship --resume <branch>

Input resolution. Strip leading whitespace; take the first whitespace-separated token. If [ -r "$token" ], prompt = file contents. Otherwise prompt = entire argument verbatim (including newlines). Empty resolved prompt → refuse stopped:no-prompt.

No args → refuse with stopped:no-prompt. No auto-pick — without a prompt there is nothing to ship.

Parameters (consumed by worktree-ship-pipeline)

Parameter	Value
EXPECTED_REPO_BASENAME	consumer-repo (§repo-anchor)
WORK_SOURCE	Verbatim user prompt (file contents or inline body)
SHORT_ID	`$(uuidgen
BRANCH	worktree-prompt-<SHORT_ID>
WORKTREE_PATH	.claude/worktrees/prompt-<SHORT_ID>
TARGET_BRANCH	$(git branch --show-current) captured at preflight
INTENT_FIELD	intent_excerpt
FILES_MAX	20
DELETIONS_MAX	600
TIEBREAKER_4	Ask user fix-now / defer / false-positive (see below)
TIEBREAKER_5	n/a (one prompt per invocation)
LOCK_ANCHOR	Branch existence + <WORKTREE_PATH>/.run-prompt/PROMPT.md
CLOSE_STEP	Echo merged_sha + summary to user

The worktree is created per §phase-1b-worktree-creation with name prompt-<SHORT_ID>; the resulting branch is worktree-prompt-<SHORT_ID>.

Looser scope thresholds than /kata-ship (10/300) because free-form prompts legitimately span more surface than a single kata.

TIEBREAKER_4 (out-of-scope finding)

prompt-ship re-gates 1/2/4 (the §tiebreakers-shared opt-out): a free-form prompt has no RDR/kata scope to ground an auto-disposition against, so finding tiebreakers ask the user here rather than auto-dispose. (kata-flight does not invoke prompt-ship.)

Agent returns a packet (≤80 words) per §tiebreakers-shared Mechanic: roborev <severity> <source-anchor> cite, one-sentence rationale for why it's out of scope vs the prompt. The continuation agent (or the messaged agent, fast-path) applies the user's answer:

• fix-now → agent applies inline (counts toward commits_added).
• defer → agent runs /roborev-respond citing the user's deferral rationale verbatim.
• false-positive → agent runs /roborev-respond + dismiss.

No kata / issue tracker write. No FOLLOWUPS.md or other in-repo debt file.

Phase 1 — Acquire + Execute

1a. Mint identity

REPO_ROOT="$(git rev-parse --show-toplevel)"   # §repo-anchor
[ -f "$REPO_ROOT/.kata-flight/env" ] && . "$REPO_ROOT/.kata-flight/env"
EXPECTED_REPO_BASENAME="${KATA_FLIGHT_EXPECTED_REPO_BASENAME:-$(basename "$REPO_ROOT")}"
[ "$(basename "$REPO_ROOT")" = "$EXPECTED_REPO_BASENAME" ] || { echo "stopped:wrong-repo:$REPO_ROOT" >&2; exit 1; }
SHORT_ID=$(uuidgen | tr A-Z a-z | cut -c1-8)
SESSION_ID="prompt-ship/$SHORT_ID"
BRANCH="worktree-prompt-$SHORT_ID"
WORKTREE_PATH="$REPO_ROOT/.claude/worktrees/prompt-$SHORT_ID"   # absolute, anchored
TARGET_BRANCH="$(git -C "$REPO_ROOT" branch --show-current)"
PRIMARY_ROOT="$REPO_ROOT"            # == REPO_ROOT in valid parent state
WS="$(dirname "$(dirname "$(cd "$(git rev-parse --git-common-dir)" && pwd -P)")")"   # workspace root (worktree-invariant)
KATA_FLIGHT_CONTEXT_ROOT="${KATA_FLIGHT_CONTEXT_ROOT:-$PRIMARY_ROOT}"
if   [ -f "$PRIMARY_ROOT/.kata-flight/workspace" ]; then . "$PRIMARY_ROOT/.kata-flight/workspace"; elif [ -f "$WS/.kata-flight-workspace" ]; then . "$WS/.kata-flight-workspace"; fi
[ -d "$KATA_FLIGHT_CONTEXT_ROOT" ] || { echo "stopped:context-root-not-found:$KATA_FLIGHT_CONTEXT_ROOT" >&2; exit 1; }

The conventions and critiques resources live in the configured Kata Flight context root repo, derived as $KATA_FLIGHT_CONTEXT_ROOT (the primary checkout's sibling), so they are tracked and resolve regardless of the worktree. Every sub-agent brief passes $KATA_FLIGHT_CONTEXT_ROOT (and $PRIMARY_ROOT for the work tree) so the agent reads conventions ($KATA_FLIGHT_CONTEXT_ROOT/context/project-guidelines.md) and, when the prompt touches RDR-governed code, the governing RDR + critiques ($KATA_FLIGHT_CONTEXT_ROOT/rdr/evidence/...) by absolute path. Use the configured rdr/evidence/, not the external RDR methodology repo. This is symlink-safe because KATA_FLIGHT_CONTEXT_ROOT derives from the worktree's git topology, not the skill file location.

If git worktree list --porcelain already shows that branch (uuid collision, extremely rare), re-roll.

1b. Parent creates the worktree

Per §phase-1b-worktree-creation, with name: "prompt-<SHORT_ID>".

Additional step after verify (4) and before recording (5): persist the prompt for resume. Parent writes the resolved prompt to <WORKTREE_PATH>/.run-prompt/PROMPT.md using the Write tool (preserves newlines).

The repo root .gitignore already excludes .claude/, so .run-prompt/PROMPT.md is invisible to git from every angle. Do not touch .git/info/exclude — see §phase-1b-worktree-creation final paragraph for the reason.

1c. Spawn execute agent

Agent(subagent_type: "general-purpose"). Brief:

• $SESSION_ID, parent cwd, worktree_path, branch, TARGET_BRANCH.
• PRIMARY_ROOT + KATA_FLIGHT_CONTEXT_ROOT (absolute, from 1a). Resources live in the configured Kata Flight context root; the agent reads conventions from $KATA_FLIGHT_CONTEXT_ROOT/context/project-guidelines.md, and when the prompt touches code an RDR governs, grounds against that RDR + its critiques at $KATA_FLIGHT_CONTEXT_ROOT/rdr/evidence/... (the configured RDR evidence root, not the external RDR methodology repo).
• Gauge complexity; ultrathink if non-trivial. Trivial paths (typo, hint phrasing, mechanical rename) proceed directly. Otherwise ultrathink before editing and read cited code + any governing RDR + principles first. Triggers: cross-subsystem reach, structural change (new abstraction, audit-boundary move), a finding/edit touching project_core_principles, or round-trip/closure/hash-stability invariants where a wrong call corrupts artifacts silently.
• First Bash call: cd <worktree_path> + isolation verify:

  [ "$(pwd -P)" = "<worktree_path>" ]
  [ "$(git rev-parse --git-dir)" != "$(git rev-parse --git-common-dir)" ]
  [ "$(git branch --show-current)" = "worktree-prompt-<SHORT_ID>" ]
  

  Any failure → return stopped:worktree-isolation-failed; do not edit any file.
• §worktree-invariant verbatim.
• §scope-discipline verbatim. Commit ONLY the paths the prompt names plus the callers/tests they force; never weaken/clean up an unrelated test or source to make a broad go test ./... pass; report the final set as in_scope_paths. (A free-form prompt may legitimately span more surface than a kata — FILES_MAX=20 reflects that — but "more files the prompt asks for" is not "unrelated files cleaned up on the side.")
• §tiebreakers-shared verbatim, plus the prompt-ship TIEBREAKER_4 spec from above.
• The verbatim user prompt between explicit delimiters (<<<PROMPT ... PROMPT>>>), preceded by: "The following is the user's request, verbatim. Execute it inside the worktree. File paths and @-mentions in the prompt are relative to the primary repo root; the worktree mirrors that tree."
• Project commit conventions: Conventional Commits (feat|fix|chore(<scope>): <outcome>); no LLM attribution in subjects/bodies/co-authors; no roborev IDs / kata slugs / job NNN / bare 4-char slugs in commit subjects (memory commit_subject_self_check); run golangci-lint run on changed scope before committing (memory lint_before_commit).
• Tiebreaker protocol: return a TIEBREAKER: packet (condition #, ≤80-word context, proposed disposition) per §tiebreakers-shared Mechanic.

Report schema (≤200 words, no diffs):

• worktree_path, branch, head_sha (short)
• commit_subject
• in_scope_paths: every path the commit(s) touched (§scope-discipline; the §phase-1d gate verifies the committed diff is a subset of this)
• intent_excerpt: 1–3 sentences — refine's tiebreaker-2 context
• cited_refs: RDRs / docs / files (or empty)
• followups_noted: optional informational TODOs (no issue tracker write happens)
• verdict: ok | stopped:<reason>

1d. Verify

Per §phase-1d-verify-shared with FILES_MAX=20, DELETIONS_MAX=600. Scope-gate trip → surface shortstat to user; user answers proceed (refine anyway) or abort (git worktree remove + git branch -d). The shared §phase-1d in-scope-set check also fires: any committed path outside the agent's reported in_scope_paths trips the same proceed/abort question.

stopped:worktree-isolation-failed or stopped:empty-result (no commits) → empty by contract. Teardown:

git worktree remove <WORKTREE_PATH>
git branch -d <BRANCH>

Not a --resume candidate. For empty-result, the prompt itself may need editing; user re-invokes with a revised prompt.

Phase 2 — Rebase + Refine

Per §phase-2-rebase-refine. No consumer lock/label transition (branch existence is the lock).

Brief includes:

• The verbatim user prompt (refine's tiebreaker-2 needs it).
• PRIMARY_ROOT + KATA_FLIGHT_CONTEXT_ROOT (absolute, from 1a) per §phase-2-rebase-refine's grounding bullet — the agent grounds findings against any governing RDR + critiques ($KATA_FLIGHT_CONTEXT_ROOT/rdr/evidence/...) and conventions ($KATA_FLIGHT_CONTEXT_ROOT/context/...) before fixing. §tiebreakers-shared (included verbatim by §phase-2-rebase-refine) already carries the ultrathink-before-applying rule.
• intent_excerpt + cited_refs from Phase 1.
• Prompt-ship TIEBREAKER_4 spec verbatim.

Flapping → §phase-2-broadening.

Phase 3 — Ship

Per §phase-3-ship-agent. After ff-merge + teardown:

Close step (consumer-defined): echo merged_sha + summary from the ship report. No external ticket to close.

Resume

Per §resume-mechanics, with prompt-ship specializations:

1. Ownership check. Branch must match worktree-prompt-* and the worktree (or .run-prompt/PROMPT.md after re-attach) must exist. No cross-consumer adoption.
2. Locate worktree. Standard git worktree list lookup; git worktree add <WORKTREE_PATH> <BRANCH> to re-attach if the branch exists but worktree is gone. Both gone → refuse stopped:resume-target-missing.
3. Read persisted intent. <WORKTREE_PATH>/.run-prompt/PROMPT.md must exist and be non-empty. Missing → refuse stopped:resume-prompt-missing.
4. Mint fresh $SESSION_ID (prompt-ship/<new-uuid>). $SHORT_ID from the branch stays as identity.
5. Determine phase by tip state per §resume-mechanics.
6. Re-run §phase-1d-verify-shared (including scope gate) against the existing tip before resuming Phase 2.

Manual release

git worktree remove .claude/worktrees/prompt-<SHORT_ID>
git branch -D worktree-prompt-<SHORT_ID>

No external state to clean.

Failure recovery (consumer rows; shared rows in §failure-recovery-shared)

Failure	Action
Empty/unreadable prompt	Refuse stopped:no-prompt / stopped:empty-prompt.
Phase-1d scope gate trips	Surface shortstat → proceed or abort.
TIEBREAKER_4	Agent returns packet → parent asks fix-now/defer/false-positive → respawn continuation (§Mechanic).
stopped:resume-prompt-missing	Persisted intent gone; user must re-invoke fresh or recreate PROMPT.md before retrying.

All other rows → §failure-recovery-shared.

See also

• worktree-ship-pipeline — shared bones; read first.
• /kata-ship — sibling consumer; kata-driven.
• /roborev-refine (Phase 2)
• /roborev-respond (dismiss after tiebreaker)