What it does

From any spec, the AI assistant can implement the spec's code for you:

1. Inspect — the agent checks which repository is connected to the spec's project.
2. Propose — it drafts complete file contents, a commit message, and a pull request description, then submits them as a pending proposal.
3. You approve — nothing is written to your remote until someone with implementation permission clicks Approve.
4. Execute — on approval SpecBridge creates the branch (sb/<SLUG>-<title>), commits all files as one commit, and opens a pull request whose body embeds SB#<SLUG> so code linking and status tracking work automatically.

Rejecting a proposal leaves your repository untouched. Every propose/approve/reject step is written to the audit log.

Who can use it

The org:agent-implement capability is granted to:

• Organization admins,
• Tenant owners,
• Members with the dev or testing discipline.

Management and stakeholder roles can chat with the AI but cannot propose or approve implementations.

In the web app

Open the AI chat panel (⌘/Ctrl + L) while viewing a spec and ask for an implementation, e.g. "implement this spec" or "propose an implementation following tasks.md". The panel shows each tool step live; when a proposal is ready you get an inline card with the branch name, file count, and Approve & open PR / Reject buttons. After approval, links to the commit and pull request appear in place, and the spec's Development tab picks up the new code link via webhooks.

From the terminal

Run sb ai and use the same flow interactively — see CLI commands.

Requirements

• The spec's project must have an active git connection (GitHub App, GitLab OAuth, or Azure DevOps).
• For GitHub, the SpecBridge App needs Contents: write and Pull requests: write permissions.
• Proposals are capped at 10 files / 100k characters per run.