self-extend
Use when evolving the agent's own capabilities with a new skill, custom subagent, plugin, hook, MCP server, or project rule.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- self-extend
- triggers
- [ "user", "model" ]
- description
- Use when evolving the agent's own capabilities with a new skill, custom subagent, plugin, hook, MCP server, or project rule.
03 SKILL.md 正文
Self-Extension
Overview
Extend Devin CLI's behavior by adding files to the project (./.devin/) or to your global Devin config (~/.config/devin/ on Linux/macOS, %APPDATA%\devin\ on Windows). Most changes are picked up at the next session or skill reload.
What to create for each need
| I want to... | Create | Location | Hot-reload |
|---|---|---|---|
| Add a reusable workflow the agent can invoke | Skill | .devin/skills/<name>/SKILL.md (project) or ~/.config/devin/skills/<name>/SKILL.md (global) |
next skill list / session |
| Define a specialized worker for a task type | Custom subagent profile | .devin/agents/<name>.md or .devin/agents/<name>/AGENT.md (project/global ~/.config/devin/agents/) |
next session |
| Add always-on context for a project or globally | Rules | .devin/global_rules.md or .devin/rules/*.md (project); ~/.config/devin/AGENTS.md (global) |
session start |
| Run custom logic at lifecycle events | Hooks | .devin/hooks.v1.json or .devin/hooks.json |
next session |
| Give the agent new API/database tools | MCP server | .devin/mcp_config.json or ~/.config/devin/mcp_config.json |
next session |
| Bundle skills/rules/hooks/MCP for distribution | Plugin | .devin-plugin/plugin.json manifest + skills/rules/hooks/mcp |
install/reload |
Creating skills
Before writing any skill or rule, invoke writing-for-agents to ensure the document follows agent-readable conventions (frontmatter, context pointers, completion criteria, pruning).
Project skill (committed to repo):
---
name: my-skill
description: Use when [specific triggering conditions]
---
# My Skill
## Overview
...
Save as .devin/skills/my-skill/SKILL.md. For personal skills, use ~/.config/devin/skills/my-skill/SKILL.md (or %APPDATA%\devin\skills\my-skill\SKILL.md on Windows).
Skill frontmatter can also set:
allowed-tools: restrict which tools the skill can use (e.g.[read, grep, glob, exec])subagent: true: run the skill as asubagent_generalsubagentagent: <profile>: run the skill as a specific custom subagentmodel: override the model for this skill (e.g.swe-2-max[free],swe-2-flash-high[free],swe-2-flash-medium[free permanent]) - never use paid modelspermissions: add permission grants/restrictions (e.g.allow: [Read(src/**)],deny: [exec])triggers: who can invoke it (default[user, model])
Discover with skill list --path . and skill search --path . --keywords "...". Invoke with /my-skill.
Creating custom subagents
Custom subagents are Devin CLI profiles that specialize in a kind of work (e.g. reviewer, researcher).
Profile definition in .devin/agents/reviewer.md or .devin/agents/reviewer/AGENT.md:
---
name: reviewer
---
You are a careful code reviewer. Read the diff, then report only concrete issues with file paths and line numbers.
Reference the profile in a skill frontmatter:
---
name: review-pr
description: Use when the user asks for a focused PR review.
agent: reviewer
---
Or dispatch directly from a skill:
Run a `subagent_general` or `researcher` subagent (NOT `subagent_explore` when parent is FREE - runs on the CLI default router, possibly PAID) with the reviewer profile, passing the diff and the review checklist.
Adding rules
.devin/global_rules.md and .devin/rules/*.md inside the project are loaded by default. Keep them small; put detailed guidance in skills.
.devin/rules/*.md files can have trigger frontmatter (always_on, manual, model_decision, agent, glob) to load only when relevant.
.devin/global_rules.md is an always-on file inside .devin/.
Adding hooks
Hooks run custom logic at session lifecycle events. Place them in .devin/hooks.v1.json or .devin/hooks.json.
Common events:
| Event | Use |
|---|---|
PreToolUse |
Inspect or block a tool call |
PostToolUse |
Log or modify tool output |
PermissionRequest |
Auto-allow/deny specific permission patterns |
UserPromptSubmit |
Inject context when the user sends a message |
Stop |
Run cleanup before the agent stops |
SessionStart / SessionEnd |
Setup/teardown |
Example .devin/hooks.v1.json:
{
"hooks": [
{
"event": "PreToolUse",
"matcher": "exec",
"type": "command",
"command": "echo 'exec called: {{tool_name}}' >> .devin/audit.log"
}
]
}
Adding MCP servers
MCP servers give the agent new tools. Configure them in .devin/mcp_config.json (project) or ~/.config/devin/mcp_config.json (global):
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { }
}
}
}
For secrets, use .devin/mcp_config.local.json (gitignored).
Packaging as a plugin
A plugin is a directory with a .devin-plugin/plugin.json manifest. It can ship:
skills/<name>/SKILL.md.devin/global_rules.md.devin/rules/*.md.devin/agents/<name>.mdmcp_config.jsonhooks.json/hooks.v1.json
Plugins can be installed from a GitHub repo, git URL, or local folder when Devin CLI plugin support is enabled. See the Devin CLI docs for plugin.json format.
Helper scripts
If a skill needs executable helpers, keep them in the skill's scripts/ directory. Use Python for cross-platform work, Bash for one-off shell integration, or JavaScript if the helper is tool-specific. Document dependencies and exit codes.
Constraints
- Do not ask users for secrets; put credentials in
.devin/config.local.jsonor env vars. - Do not modify repository security policies or CI compliance settings.
- Keep
.devin/global_rules.mdsmall; prefer skills for detailed guidance. - Prefer standard Devin CLI paths; do not reference non-Devin runtimes.
Dynamic Subagent Construction (AOrchestra 4-tuple)
Any subagent can be described as a 4-tuple: ⟨Instruction, Context, Tools, Model⟩. The 7 fixed profiles (architect, debugger, implementer, researcher, reviewer, subagent_explore [PAID - não usar quando parent FREE], subagent_general) are pre-configured 4-tuples. For task-specific specialization, override individual tuple elements instead of creating a new profile:
| Element | What it controls | When to override |
|---|---|---|
| Instruction | What the subagent must accomplish | Task needs a specific goal not covered by any profile |
| Context | Task-relevant working memory | Task needs scoped context (filter out irrelevant files/history) |
| Tools | Actions the subagent can take | Task needs restricted tool access (e.g., read-only exploration) |
| Model | Which LLM executes | Cost-performance trade-off (cheaper model for simple subtasks) |
How to apply: When dispatching via run_subagent, specify task (instruction),
profile (defaults for context/tools/model), and override with task-specific
constraints in the task text. The 7 profiles remain as backward-compatible defaults.
Source: AOrchestra (ICML 2026, arXiv:2602.03786). "Sub-agents should be treated as recipes created at runtime, not fixed roles." Training-free mode: 16.28% relative improvement over baselines. SFT mode: +11.51% pass@1 on GAIA (requires 2K trajectories).
Cross-skills
- Use
ai-coding-dictionaryto align terms like subagent, harness, and context engineering before writing new skills or rules. - Use
continuous-improvementto validate that the new skill or rule actually improves the bundle, not just adds it.
04 同仓库更多 Skills
montador-de-ideias
Use when assembling, expanding, or elevating a raw project idea or initial prompt into an ultra-detailed, first-principles dossier and compiled master prompt with journalistic OSINT market research, Tech Lead and Staff Architect justifications, atomic and modular hardening/optimization, and high-impact contact surfaces.
git-github-adaptive-engineering
Use when any task touches git history, GitHub issues/PRs/releases/CI, repo governance, or branch and merge decisions. Runs the Fase Zero situational-awareness battery, picks one of five delivery flows via the adaptive decision engine, and enforces commit/PR information-density standards. Orchestrates the gh, git-workflows, github-actions, pr, and debug/finish skills.
adaptive-agenda-calendar-routine-engine
Use when planning or executing work across hours, days, weeks, or sprints. Turns the calendar into a strict time-blocked agenda (HH:MM-HH:MM blocks carrying a routine profile, expected inputs, and a handoff output artifact), runs the 5-stage cognitive flow (Radar and Hunt -> Triage plus Steve Jobs conception -> workbench manifest -> >=32-cycle loop -> Close plus Open-Minded Scrum), and unfolds macro calendar milestones into micro agenda blocks through extensions/gcal/wrapper.py.
demo
Use when demonstrating a skill in a fixture project.
duplicate-skill
Use when testing duplicate skill name detection.
duplicate-skill
Use when testing duplicate skill name detection.