Agent-Skills.md

Agent Skills: Ship PR

Ship a PR end-to-end: verify locally, push, create/update the PR, watch CI, handle review feedback, and merge without admin override. Use for 'ship', 'ship it', or 'push and merge'. SKIP if the user only wants a PR opened; use `create-pr`.

UncategorizedID: gopherguides/gopher-ai/ship

Repository

gopherguides/gopher-ai
gopherguidesLicense: MIT
172

Install this agent skill to your local

pnpm dlx add-skill https://github.com/gopherguides/gopher-ai/tree/HEAD/plugins/go-workflow/skills/ship

Skill Files

Browse the full folder contents for ship.

Download Skill

Loading file tree…

plugins/go-workflow/skills/ship/SKILL.md

Skill Metadata

Name
ship
Description
"Ship a PR through local verification, push, CI, review, and merge without admin override. Use for 'ship', 'ship it', or 'push and merge'. SKIP if the user only wants a PR opened; use create-pr."

Ship PR

Plugin Resource Resolution

<PLUGIN_ROOT> is notation. Replace it with a concrete absolute plugin root before every resource read or command:

  • Codex: Start from the directory containing the absolute selected SKILL.md path, then ascend two directories (skills/<name> -> plugin root).
  • Claude Code: Bind it to the injected ${CLAUDE_PLUGIN_ROOT} value.

Before decisions or delegation, read <PLUGIN_ROOT>/lib/driver-interaction.md. Read <PLUGIN_ROOT>/lib/decision-gates.md before resolving any workflow choice. Bind the invocation arguments as SKILL_ARGS for $go-workflow:ship by reading <PLUGIN_ROOT>/lib/skill-arguments.md with this compatibility payload:

<claude-skill-arguments> $ARGUMENTS </claude-skill-arguments>

Run source "<PLUGIN_ROOT>/lib/github-rest.sh". Routine PR metadata, formal reviews, CI, mergeability, and squash merge use its REST helpers. Use GraphQL only for review threads, closing-issue references, or required merge queues.

Setup, Arguments, and Re-entry

Read <PLUGIN_ROOT>/lib/ship/bootstrap.md completely and execute the state bootstrap. Standalone ship owns one canonical state file; embedded ship uses the caller file and child path, never creates another loop, changes root promises, or emits a terminal marker.

Parse SKILL_ARGS into LLM_CHOICE (default codex), MAX_PASSES (default 3), NO_MERGE, SKIP_COVERAGE, COVERAGE_THRESHOLD (default 60), GEMINI_TIER, and LLM_EXPLICIT. Supported LLMs are codex|gemini|ollama|fable; tier is flex|standard|priority. --skip-coverage never waives coverage for changed source files.

Read <PLUGIN_ROOT>/lib/ship/state-fields.md completely and persist every argument and workflow field. Then read <PLUGIN_ROOT>/lib/ship/reentry.md completely and execute its re-entry check. Its phase routing and expired-review recovery are authoritative.

If this skill or any supporting file sets WORKFLOW_RESULT=INCOMPLETE, execute Hard Invariant Failure from reentry.md and stop. Embedded ship returns structured incomplete state without a marker; standalone ship persists and emits only <done>INCOMPLETE</done>.

Phase routing:

  • reviewing -> expired-review recovery, then push
  • review-required -> local review
  • fixing|verifying|coverage-check|e2e-testing -> corresponding local-review step
  • pushing -> push and PR creation
  • ci-watch -> exact-head CI
  • bot-watching -> bot watch
  • addressing -> address feedback
  • merging -> final merge

Context and Prerequisites

Read <PLUGIN_ROOT>/lib/ship/context.md completely and execute PR, branch, and worktree-ownership discovery. Shipping the default branch is a hard invariant. Include dirty changes only when ownership and fresh validation are unambiguous; otherwise preserve unrelated work or stop incomplete.

Read <PLUGIN_ROOT>/lib/ship/prerequisites.md when resolving the review backend. An unavailable unpinned default uses external CLIs, then a recorded skip; no automatic sub-agents. Replacing a pinned backend is a missing-intent gate.

Phase 1: Local Review and Verification

Read <PLUGIN_ROOT>/lib/ship/local-review.md completely and execute Steps 5-8. It owns LLM review/fix passes, build/test/lint, generation drift, the final-pass coverage gate, non-UI E2E skip, scoped staging, commit, and loop decision. UI-visible changes require passing E2E; missing browser tooling or an unavailable server blocks before push or merge.

Phase 2: Push and PR

Set phase pushing. Read <PLUGIN_ROOT>/lib/ship/push-and-pr.md completely. Push the explicit head, ensure a non-draft PR exists, then persist the published head SHA and bot-review baseline.

Phase 3: Exact-Head CI

Set phase ci-watch. Read <PLUGIN_ROOT>/lib/ship/ci-watch.md completely. No checks, prior-head checks, or a concurrent head shift are never passing CI. Every completion claim must use terminal successful checks for the exact published head.

Phase 4: Bot Review

Set phase bot-watching. Read <PLUGIN_ROOT>/lib/ship/bot-watch.md completely. Detect registered bots from REST reviews/comments, GraphQL threads, and exact-head statuses; apply its bounded startup wait and approval rules.

Phase 5: Address Feedback

Set phase addressing. Read <PLUGIN_ROOT>/lib/ship/address-bots.md completely. It owns rebase-or-abort, the embedded address-review Steps 2-11 boundary, baseline-before-push ordering, and the return to exact-head CI.

Phase 6: Merge

Set phase merging. Read <PLUGIN_ROOT>/lib/ship/merge.md completely. Never use admin override or bypass branch protection. Use the configured merge strategy (squash when available), expected head SHA, final CI/thread/human-review checks, and merge queue only when required.

Completion Contract

Return shipped only after fresh evidence proves: local review is clean/maxed or durably recorded as void/skipped for an allowed session or backend-unavailable reason and the exact current head passes CI; applicable coverage and E2E passed; changes are pushed; a non-draft PR exists; detected bots approve or none apply; no unresolved or human-requested changes remain; and the PR merged (or --no-merge was supplied). An unrecorded timeout, error, or early exit never satisfies review completion.

Standalone success persists then emits <done>SHIPPED</done>. Embedded success persists its structured child result and emits no marker. After 15 unsuccessful iterations, persist incomplete evidence and stop; never bypass a completion criterion.