runnerless↗

Runnerless concept · not available yet

Custom AI hooks for repository workflows

AI is most useful in a workflow when its task and authority are clear. Runnerless could expose hooks for issue triage, comment responses, code analysis, review and proposed changes while letting the repository operator choose which tools each hook may use. The current toolkit’s Codex integration is a narrower, implemented starting point.

Design direction for Runnerless. The workflow below is not implemented or available to install.

The proposed workflow

01 /

Give the hook a bounded job

An issue hook might summarize a reproducible bug; a review hook might inspect a particular diff; a code-change hook might propose a patch. These tasks need different context and success criteria. One unrestricted assistant with access to every repository action would make that distinction harder to review.

02 /

Separate analysis from permission to act

Reading files, calling network tools, posting comments and changing code should be separate grants. Repository text, issue bodies and artifacts are inputs to analyze, not a source of authority to expand those grants. A proposed patch could go through a normal PR and existing checks before any merge action becomes eligible.

03 /

Keep the action inspectable

A useful hook should leave an explanation of its input scope, tool calls and proposed workflow action. Current ci-toolkit already distinguishes requesting a Codex review from trusting its verdict and from authorizing a merge. Runnerless could extend those explicit boundaries to more models and event types.

A concrete example

Workflow example

Proposed hook: issue triage
  Input: issue title + body
  Tools: read selected repository files
  Output: suggested labels + explanation
  Action: propose a comment
  Code writes and deployment: not granted

Know the boundary

What is still to build

General AI hooks, arbitrary model selection and configurable tool execution are planned, not released. AI and network-dependent workflows are outside the one-second local-check target. No AI credentials are collected by the early-access form.

Common questions

What AI behavior exists today?

ci-toolkit requests Codex reviews for opted-in PR heads and validates trusted review evidence through its gate. General issue or code-change hooks are not implemented.

Would an AI hook be able to merge its own patch?

The proposed model separates tool grants and workflow actions. Current toolkit merging already requires an authorized command and the existing GitHub checks.

Runnerless / early access

Want this in
your workflow?

Get early access ↗

Register interest and tell us what slows your CI down. The optional follow-up questions come after your email is saved.