AgentSkillsCN

coder

使用 Coder 的 `coder` CLI 和 API 进行协作，尤其适用于 Coder 任务（Coder 工作空间中的 AI 智能体）：创建/列出/监控任务（`coder task create|list|status|logs|send|delete`）、选择任务模板/预设，并排查任务模板的需求（`coder_ai_task` 和 `coder_task`）。触发器：Coder、Coder 任务、coder task、coder_ai_task、AgentAPI、AI Bridge、模板推送。

SKILL.md
--- frontmatter
name: coder
description: "Work with Coder using the `coder` CLI and API, especially Coder Tasks (AI agents in Coder workspaces): create/list/monitor tasks (`coder task create|list|status|logs|send|delete`), pick task templates/presets, and troubleshoot task-template requirements (`coder_ai_task` and `coder_task`). Triggers: Coder, Coder Tasks, coder task, coder_ai_task, AgentAPI, AI Bridge, templates push."

Coder (Tasks-first)

Goal

Use Coder Tasks as the primary unit of work (a task runs an agent inside an isolated Coder workspace), and manage tasks from the terminal reliably.

Key concept: Task ≈ Workspace + prompt

  • •A Coder Task provisions a backing workspace and runs an agent with the task prompt.
  • •Use Tasks for parallel, one-shot agent jobs; use Workspaces for interactive development.
  • •Cleanup is two layers: deleting a Task may not remove the backing Workspace (stop/delete it explicitly when needed).

Quick start (CLI)

  1. •Ensure coder CLI is installed and authenticated
  • •Login: coder login <url>
  • •Print current session token (useful for API calls): coder login token
  1. •Create a task
  • •Direct input: coder task create "<prompt>"
  • •From stdin: echo "<prompt>" | coder task create
  • •Non-interactive: pass --template, optional --preset, optional --name
  • •Automation: add --quiet to print only the task ID
  • •From Beads (recommended for strict scope): python3 scripts/task_from_beads.py <bead_id> --template <tpl> --preset <preset>
  1. •Monitor + iterate
  • •Watch status: coder task status <task> --watch
  • •Send follow-ups: coder task send <task> "<more instructions>"
  • •Inspect logs: coder task logs <task> -o json
  1. •Find the backing workspace (when you need a shell)
  • •coder task status <task> -o json → read workspace_name
  • •Use normal workspace commands after that: coder ssh <workspace>, coder stop <workspace>, coder delete <workspace>
  1. •Cleanup
  • •Delete (no prompt): coder task delete <task> --yes

Git operations from Tasks (important)

HTTPS pushes from Tasks can fail (external auth wrappers, missing scopes, workflow restrictions). Prefer SSH deploy keys for repo write access.

Steps + troubleshooting: references/git-from-tasks.md.

Task templates (required)

If coder task create fails with something like “no task templates configured”, the deployment has no templates that are Tasks-capable.

A template becomes Tasks-capable when it defines:

  • •resource "coder_ai_task" ... (links a workspace app to the Task UI)
  • •data "coder_task" "me" {} (gives access to the task prompt and metadata)
  • •An agent module (Codex CLI, Claude Code, etc.) that runs in the workspace and consumes data.coder_task.me.prompt

Minimal Terraform snippet + notes: references/task-template-snippet.md.

Subscription auth (Codex + Claude)

  • •Codex CLI (ChatGPT subscription): login is interactive by default; for headless Tasks, pre-seed ~/.codex/auth.json in the workspace (e.g., via a sensitive template variable written at startup). If you see an “unsupported model” error, use a ChatGPT-supported Codex model (typically gpt-5.2-codex).
  • •Claude Code (subscription): generate an OAuth token via claude setup-token and pass it to the workspace (e.g., as a sensitive template variable). Avoid storing plaintext tokens in git. If you use OAuth, do not also set CLAUDE_API_KEY (pick one auth mode).
  • •Set/update template variables non-interactively with coder templates push <template> --variables-file <yaml> --yes.

API (automation)

If you need to create tasks from CI/GitHub automation (or without the coder CLI), use the Tasks API (/api/v2/tasks/...) with a Coder session token.

Quick reference: references/api.md.

Safety defaults

  • •Prefer a dedicated, locked-down Task template (least privilege, no prod secrets).
  • •Treat agents as untrusted: restrict permissions in the agent module when supported.
  • •Keep prompts explicit: inputs, repo, constraints, and “done when” criteria.

Troubleshooting

  • •Need the backing workspace: coder task status <task> -o json (look for workspace_name/IDs), then use normal workspace commands.
  • •Task is error/unhealthy: start with coder task logs <task>.
  • •GitHub Actions is stuck in queued: check whether Actions is enabled and runners/minutes exist; see references/git-from-tasks.md.

References (load as needed)

  • •Task template snippet + design notes: references/task-template-snippet.md
  • •Tasks API quick reference: references/api.md
  • •Git auth + deploy keys from Tasks: references/git-from-tasks.md
  • •Beads → Task prompt generator: scripts/task_from_beads.py