self-healing
GitHub允许用户通过自然语言指令动态修改Agent自身配置，包括切换模型、添加工具或MCP服务器、调整图像生成提供商及安装新技能。执行手术式文件编辑，无需重启即可生效，实现Agent的自我修复与升级。
Trigger Scenarios
Install
npx skills add lthoangg/OpenAgentd --skill self-healing -g -y
SKILL.md
Frontmatter
{
"name": "self-healing",
"description": "Update or upgrade the agent's own configuration on request — swap the model, add tools or MCP servers, change the image-generation provider\/model, or install a new skill. Use when the user says things like \"upgrade yourself\", \"switch your model to X\", \"use Gemini for images\", \"make yourself faster\/smarter\"."
}
Self-Healing Skill
Modifies on-disk agent configuration in response to user requests. All changes are surgical edits under {OPENAGENTD_CONFIG_DIR}/. No code changes, no restarts.
Modes — pick the right team first
| Mode | When active | Agent files |
|---|---|---|
normal |
default; no workspace attached | {AGENTS_DIR}/*.md |
coding |
coding workspace attached | {AGENTS_DIR}/coding/*.md |
Always edit the files for the mode you are currently running as. MCP servers and multimodal.yaml are global (shared by both modes).
The normal lead includes generate_image / generate_video by default; the coding lead does not.
Scope
| Target | File | Example request |
|---|---|---|
| Model / params | Agent .md frontmatter |
"switch to gpt-5" |
| Tools | Agent .md → tools: |
"give yourself shell access" |
| MCP servers | Agent .md → mcp: or tools: |
"let yourself use the filesystem MCP" |
| Image / video generation | {OPENAGENTD_CONFIG_DIR}/multimodal.yaml |
"use Gemini for images" |
| New skills | {SKILLS_DIR}/{name}/SKILL.md |
use skill skill-installer |
Workflow
- Identify the target file. Ask "which agent?" only if ambiguous; default to the lead of the current mode.
- Read the current file.
- Compute the minimal diff — change only what was asked. Never reformat unrelated lines.
- Show a
diffblock with a one-line summary. - Wait for confirmation unless the request was already explicit — then apply in the same turn after showing the diff.
- Apply with
edit(preserves the file verbatim). Usewriteonly for new files. - Report what changed, which file, and when it takes effect.
When changes take effect
| Change | Takes effect |
|---|---|
Agent .md |
Next turn of that agent (drift detection) |
mcp.json |
After restarting the affected MCP server, on the next agent turn |
SKILL.md edited |
Next turn of any agent using the skill |
| New skill installed | Immediately loadable by skill("name") |
multimodal.yaml |
Instantly — read on every generate_image / generate_video call |
Process restart required only for: adding/removing agent files, .env changes, and code changes.
Agent frontmatter — valid fields
| Field | Values |
|---|---|
model |
provider:model — e.g. openai:gpt-4o, anthropic:claude-..., googlegenai:gemini-... |
tools |
Extra tools on top of built-in profile: web_search, web_fetch, date, read, write, edit, patch, ls, grep, glob, rm, shell, bg, generate_image, generate_video, plus <server>_<tool> for MCP. Never add skill or team_message — injected automatically. |
Invariants: exactly one role: lead per agents dir; model must contain :.
Recipes
Find your own agent file
grep -l 'role: lead' {AGENTS_DIR}/*.md {AGENTS_DIR}/coding/*.md
Setting up multimodal.yaml
When generate_image / generate_video returns "not configured", or the user explicitly asks.
- Pick a provider (user-specified, or default to
openai:gpt-image-2/googlegenai:veo-3.1-generate-preview). - Write immediately — don't check env vars first. If a key is missing the tool returns a clear error; relay it then.
- Retry the original generation request after writing.
MCP tools on agents
mcp: list (recommended) — grants all tools the server exposes:
mcp:
- filesystem
tools: list (selective) — pick specific tools:
tools:
- filesystem_read
Verify the server is ready before wiring:
curl -sS http://localhost:4082/api/mcp/servers/<name> | jq '{state, tool_names}'
When removing a server: strip it from agent .md files first, then edit mcp.json — otherwise the agent logs agent_config_refresh_failed on its next turn.
Failure modes
- File not found → ask the user to confirm the path / agent name.
- Malformed frontmatter → show the problem; don't attempt repair unless asked.
- Invariant violation (two leads, unknown tool, unsupported provider) → explain and suggest the nearest valid alternative.
Version History
- b4ec0d3 Current 2026-07-24 16:52
