land-changes

Land requested changes, inspect available landing workflows, or manage repository landing preferences. Use for "land it", "ship this branch", or "push to main".

SKILL.md

GitHub

Land changes

Use the destination's landing procedure and the user's choices to publish the requested change and complete its authorized follow-ups. Inspecting, discussing, or editing this skill authorizes no landing.

Resolve scope and state

From the conversation, identify the change, source checkout, destination, target, and requested stopping point. Before publication, verify the publishing identity and effective remote destinations; source and destination may differ.

Inspect the full diff, uncommitted work, dependencies, PR/stack metadata, and any work already completed. Preserve unrelated changes and isolate preparation when it would disturb them. Prefer one integration commit where the selected method allows.

Read choices and preferences before acting. Resolve invocation arguments and applicable saved choices there. Show and reset requests end after preference handling. --inspect makes no checkout, check-run, preference, or publication changes.

An action completed earlier does not request a new change. Record the intended scope and the outcomes still authorized; that defines when this run is complete.

Discover the destination's workflow

Treat a method as the publication route, an action as work to perform, and a gate as evidence required before proceeding. The user may name outcomes without using these categories.

Follow the destination's guidance and inspect the evidence that applies:

  • Contribution or agent instructions and the repository landing procedure.
  • Task definitions, selected validation, and effective Git/tool hooks.
  • Live host policy, available caller capabilities, and existing PR, queue, or stack state.
  • Automation caused by publication, including follow-on workflows, release/deployment jobs, environments, and available integration metadata.

For GitHub, follow host evidence. For an unfamiliar mechanism, read its native help and official documentation. When the repository's convention remains unclear, read review history. History can inform a recommendation; it cannot grant an exception or follow-up.

Use names from the destination, qualified when necessary. For every relevant choice, establish its mechanism, dependencies, controls, affected scope, and completion evidence. Mark each evidence source inspected, inapplicable, or unavailable, and keep important unknowns or conflicts visible. Capability does not grant user authorization; authorization does not make a capability available.

For a landing request, inspect the chosen route and its dependencies rather than cataloguing unrelated routes. For --inspect, show configured choices, alternatives that need setup, and unknowns with their sources and controls; recommend a route and stop. Discovery is complete when every relevant source is accounted for and every requested outcome has a procedure or a named blocker.

Resolve choices and gates

Apply explicit choices, applicable standing decisions, and valid saved preferences before defaults. Use the selectors or equivalent natural language in choices and preferences. Resolve conflicts by scope and recency.

Without an overriding choice:

  • Use the established route that meets policy without a bypass. Resolve material route ambiguity before publishing.
  • Run relevant repository checks and hooks. A failure blocks by default; retain checks explicitly marked informational. Wait for required evidence, not unrelated or informational CI.
  • Run manual follow-ups only when explicitly requested or covered by a saved preference scoped to this change, host/environment, and effect.

Record each selected action and gate, whether it will run or be skipped, whether it will be enforced or bypassed, its waiting behavior, and the source of that decision. For every bypass or follow-up, cite the instruction or saved key that covers its scope. Include effects triggered automatically by publication. If an effect lacks authorization or conflicts with a user constraint, find an authorized route that avoids it or resolve the decision before publishing. A no-deployment constraint requires evidence before publication that the route avoids deployment; publishing and then watching for deployment is not proof.

Skipping a check preserves its known result; it does not convert failure to pass. Enforce each applicable gate unless a scoped exception covers it and the actor has a usable bypass. Report informational failures separately. Shared protection changes, new privileges, and history rewrites each need explicit scope.

Reuse decisions that still cover the work and resume unfinished authorized actions without asking again. A completed apply for one change does not authorize an apply for another. The resolved plan, including follow-up scope and source, governs execution. If new evidence changes a choice, revise the plan explicitly.

Resolve all choices or mark them as blockers. Continue independent preparation while a blocker is being resolved; publication waits until its blockers are clear.

Prepare and publish

For a direct update, follow direct Git landing. For a PR, queue, stack, or other method, follow the repository's or host's procedure. Preserve attribution and PR/stack relationships. Run selected actions on the final candidate and record their results. Finish independent preparation before asking for an unresolved publication decision.

Immediately before publication, verify the scope, destination, live policy, and candidate. Every applicable gate needs valid evidence or a usable authorized exception, and every resulting effect must fit the resolved plan. If the target or candidate changed, refresh affected evidence and choices while retaining decisions that remain in scope. After an uncertain response, inspect remote state before retrying.

Publication is complete only when the live destination contains the intended change. For a direct push, verify the prepared commit or its ancestry. For a server-created commit, verify the source revision, integration commit, and intended diff from the method's result metadata. PR or queue submission is an intermediate result unless the user requested it as the stopping point. Report a policy rejection as a rejection, then resolve an authorized route.

Complete authorized follow-ups

Safely synchronize local checkouts. For each selected follow-up, read its repository runbook and confirm that its authorization, effect scope, and published revision or artifact still match the plan. Perform only the remaining authorized work. Verify actual activation when command success alone is insufficient. After an uncertain outcome, inspect state before repeating a side effect.

Report publication, local synchronization, validation, CI, and follow-ups separately as passed, failed, skipped, pending, or unknown. Name bypassed gates and unrequested follow-ups. Keep branches and worktrees unless cleanup was requested. Finish when every requested outcome is verified or its exact blocker is reported.