/github-actions

Make it your own.

In Claude Code, use the slash command and add your context. In Codex, select github-actions from the just-vibe skill picker, then send the same brief.

Version 0.11.0 also supports /jv github-actions, /just-vibe github-actions and /jv:github-actions in Claude. See shortcut setup and context examples.

What the agent does

1. Map each event/ref, actor trust level, token permission and artifact producer/consumer before editing the workflow. Inspect the pinned action/runtime versions and repository policy; consult current primary documentation for changed platform behavior.
2. Keep untrusted PR text out of executable shell source and separate privileged metadata work from untrusted code execution. Validate artifact origin across workflow boundaries and define minimum job permissions.
3. Design job dependencies, cache identity and cancellation by effect: cancelling an obsolete test run differs from interrupting a release halfway through publication. Prevent overlapping destructive jobs without hiding failures or sharing artifacts across untrusted scopes.
4. Validate syntax plus representative trusted/untrusted event paths. Verify required check names remain reachable for applicable branches and that skipped/conditional jobs do not accidentally report an untested release as ready. Describe changes without agent self-attribution.
5. Use the matching bundled evidence collector when available; read its result and limitations rather than treating exit zero as readiness. Revalidate identity before a dependent action.

Inputs

• workflow goal, event model, runtime constraints, and existing workflows.

Optional context: scope, references, constraints, successCriteria, environment, mode, budget.

Scope

Reads

Workflow files, least-required permissions, caching, concurrency, and job structure; no repository secret/admin changes implicitly.

Writes

Apply: only the requested local changes and relevant isolated verification. Inspect/plan requests remain inspection/planning. External actions require their exact action and target in session authorization.

Mode

Apply; workflow goal, event model, runtime constraints, and existing workflows.

Prerequisites

Exact owner/repository and relevant issue/PR/ref; authenticated read access through an available connector or CLI for remote evidence. External writes require the requested operation, appropriate account permissions, and rechecking target state. Local preparation remains useful without write access.

Expected output

• Workflow changes, event/permission rationale, and validation coverage.
• Trigger/permission matrix, job graph, cache keys and trusted/untrusted scenario checks.

How the work is checked

• Untrusted fork contributions do not gain secret-bearing privileged execution; cache invalidation follows dependency changes.
• Review newly prepared commit/PR/message text, including template or hook additions, for agent self-attribution before submission; verify the resulting artifact when available. Do not silently rewrite existing history or remove human credits.

When to stop or clarify

• Missing remote execution remains unverified. Do not substitute a successful YAML parse for a working Actions run.

Handling missing context

Infer

Resolve owner/repository and PR/issue/ref from links, remotes and supplied artifacts; inspect available account and head identity.

Assume

Prepare local text or analyze supplied evidence if remote access is absent; label its freshness.

Ask

Ask only when repository/account/target ambiguity blocks the requested remote action; missing write access does not block local drafting.

Technical guidance

Evidence

Inspect event, checked-out ref, job permissions, secret access, interpolation, action pins and artifact producers.

Method

Keep untrusted contribution code outside privileged jobs; pass event values as data and scope caches/artifacts by trust and content identity.

Pitfall

pull_request_target or workflow_run plus untrusted checkout/artifacts can cross a privilege boundary even if the workflow file is trusted.

Check

Trace one fork contribution and one trusted release end to end; check that neither untrusted shell text nor poisoned artifacts reach privileged execution.

Situational decisions

When pull_request_target or workflow_run consumes attacker-controlled code/artifacts: Separate privileged metadata work from untrusted execution and validate artifact origin.

The coding agent follows this workflow using its available tools. Installation does not grant service access or guarantee an outcome. Read the compatibility notes.

Keep exploring