Category: ISSUE · Issue-assigned workflow:
issue-writer(Writer Agent).
Issues arrive in all shapes: one-liners without context, vague feature requests, bug reports that don't say what the expected behavior is. Someone has to ask questions, chase down details, and restructure the issue so it's actually implementable — work that bounces between triager and reporter, stretching a fast fix into a multi-day discussion. Writer Agent takes on that refinement: it reads the issue, asks the original author clarifying questions when needed, and creates a structured, actionable replacement issue.
- Reads the assigned issue and its discussion thread.
- Gathers repository context — project structure, conventions, existing patterns — to ground its questions in reality.
- Evaluates the issue for clarity, completeness, and testability:
- Is the problem statement unambiguous?
- Are acceptance criteria present and measurable?
- Is the scope defined?
- Are edge cases addressed?
- When the issue is ready, creates a new linked issue titled
AI Created Issue: <original title> with:
- Background
- Requirements
- Acceptance criteria
- Implementation notes
- When the issue needs more input, posts clarifying questions as a comment addressed to the original author and waits for a reply.
- Posts visible progress and completion comments on the original issue.
The writer bot is read-only for the repository: it can explore code and documentation for context, but it never creates branches, edits files, or opens pull requests. Its only side effects are comments and the new improved issue.
Use the writer agent when an issue is:
- Ambiguous or contradictory,
- Too broad or missing scope,
- Lacking testable outcomes,
- Missing acceptance criteria.
Well-structured issues that are ready to implement should go to a Coding Agent instead.
The writer bot will not start on an issue that already has an active coding-agent session.
Writer-specific settings share the agent.budget.* namespace with the coding
agent. The writer agent is simpler — it only uses read-only tools and issue
mutations — so its configuration surface is smaller.
See the full table in AGENT.md for the complete configuration reference.
- Automatically when a writer bot is assigned to an issue.
- The bot posts a confirmation comment, then either asks clarifying questions or creates the improved issue.
- When the bot is waiting for the author to answer questions, only the original reporter's reply continues the workflow — comments from other users do not advance it.
- A stuck session can often be resumed by commenting
try again,please retry, orredoon the issue.
The writer agent uses the model's native tool-calling API when the AI Integration has Enable native tool calling turned on. Turn it off only when a provider/model behaves poorly in agentic workflows; this forces the legacy JSON-in-prompt fallback. See TOOL_CALLING.md.
- Create an AI Integration and a Git Integration in the admin UI.
- Create a bot with Issue-Assigned Workflow Configuration: Issue: Writer Agent (select the seeded empty No PR workflows PR configuration to keep the bot silent on PRs, matching the historic writer-bot behavior).
- Configure your repository webhook to send Issues events to the bot URL.
- Grant the bot repository read/clone access for context, and issue write access to post comments and create the improved issue.
The bot begins work the next time someone assigns it to an issue.
- Writer sessions that need missing business context wait for the original issue author; the bot cannot proceed until it has enough information.
- Large repositories or long issue discussions may exceed the model context window; the bot sends bounded context.
- Small or heavily quantized local/ollama models may ask low-quality questions or produce poorly structured output. Production use is recommended with Anthropic Claude or OpenAI GPT-4/5 class models.
- Issue Agents overview — configuration reference and shared settings
- Coding Agent — the implementation sibling
- PR Workflows overview
