Skip to main content

Interactive Task Editor Copilot

Bosun now ships an interactive AI copilot inside the task editor. The panel sits next to the YAML editor so you can iterate on manifests without leaving the page. It reads your current manifest, chats about intent, and proposes concrete edits through the same validation pipeline as manual changes.

When to use the copilot​

  • Drafting manifests: Paste an outline and ask the copilot to expand steps, add constraints, or wire for_each loops.
  • Refining instructions: Describe desired behavior in plain English and let the copilot tighten the instructions while staying within schema.
  • Fixing validation errors: Share the error message, and the copilot will adjust the manifest until it conforms.

The copilot is purpose-built for manifest editing. It does not run arbitrary shell commands or commit code; it only edits YAML.

Opening the assistant panel​

  1. Navigate to Tasks → Edit manifest for any task.
  2. Click the speech bubble icon in the toolbar to open the Assistant drawer.
  3. Pick a repository context (required on first open). The dropdown is sorted by recent activity.
  4. Start chatting—each message sends your prompt, repository choice, and current manifest to the copilot.

The copilot stays synchronized with the active manifest. If you switch tasks or revert changes, close and reopen the panel to refresh its context.

Chat loop and manifest proposals​

Each answer follows the same flow:

  1. The copilot analyses the manifest and your instructions.
  2. When it recommends edits, it sends a structured update_task_manifest tool call with a full YAML proposal.
  3. Bosun renders the proposal as a diff. You can preview inline or open a full-screen diff popout for review.
  4. Click Apply changes to copy the manifest back into the editor. Undo works immediately because the editor manages the edit history.
  5. If the suggestion misses the mark, click Reject, optionally supply feedback, and continue the chat.

The panel highlights additions/deletions counts so you understand the scope before applying. Approved changes inherit the editor's validation rules, so schema errors surface instantly.

Safety rails and permissions​

  • The copilot only receives read access to the selected repository and cannot run commands.
  • It mirrors the same AGENTS.md guidance that runtime agents use, so repository guardrails stay intact.

Tips for effective prompts​

  • Mention specific step IDs or edges you want to tweak so the copilot edits the right section.
  • Provide acceptance criteria (e.g., "ensure the run step exits non-zero on failure").
  • Share validation errors verbatim—the copilot understands the JSON schema and can adjust fields precisely.
  • Use follow-up questions to iteratively refine instructions rather than requesting sweeping rewrites.
  • Agent step reference: describes how agents interpret instructions and constraints. Useful when asking the copilot to add new agent steps.
  • Pull request workflows: pair the copilot with the PR agent to build end-to-end automation.
  • Task graphs: consult when directing the copilot to add branching logic.