Prompt file imported from galarzafrancisco/ai-monorepo (.codex/prompts/start-task.md). Copyright stays with the author.

description: Your goal is to pick up a task and work on it, taking it all the way from "not started" to "done". argument-hint: [TASK_ID=] [ROLE=]

Input

• TASK_ID: id of the task to work on
• ROLE: who is picking up the task (dev, reviewer, pm etc). Usually dev.

Your AGENT_NAME will be codex-{$ROLE}. For example, codex-dev.

Workflow

Prep

1. Get a session ID: .codex/scripts/workflow/start_session.sh
2. Fetch the task using MCP tool mcp.tasks.get_task
3. Prepare the local environment .codex/scripts/workflow/start_task.sh {AGENT_NAME} {SESSION_ID} {TASK_ID} {TASK_NAME} where TASK_NAME is a branch-safe-short-name you make up based on the task
4. Assign the task to yourself using the MCP tool mcp.tasks.mark_task_in_progress with the {TASK_ID} {AGENT_NAME} {SESSION_ID} and {BRANCH_NAME}

Work

5. Get instructions for your role from .codex/roles/{ROLE}.md (for example, .codex/roles/dev.md)
6. Implement & make frequent comments to the task

Finish

7. Open PR to agents/main by running .codex/scripts/workflow/open_pr_to_agents_main.sh "PR Title" "PR Body" -- you will get a link to the PR.
8. Mark task as in review by calling the MCP tool mcp.tasks.mark_task_for_review with {TASK_ID} {AGENT_NAME}, {PR_LINK}
9. Watch CI
10. If CI passed:

• Merge pull request
• Update task - done & comment saying CI passed by calling the MCP tool mcp.tasks.mark_task_done with {TASK_ID} {AGENT_NAME} {COMMENTS}.

10. If CI Failed:

• Put the task back to in progress with a comment explaining the failure by using the MCP tool mcp.tasks.mark_task_needs_work with {TASK_ID} {AGENT_NAME} {COMMENTS}
• Start work again

11. When done, checkout agents/main and pull