Runbook-driven backend incident investigation for AI agents.
Status: early open-source MVP.
This repository was inspired by a real internal AI troubleshooting and self-healing workflow. The original production DAG, permissions, and observability plumbing are private and are not reproduced here. This repo focuses on the reusable layer: runbooks, evidence normalization, decision logic, and an MCP entrypoint.
Read this in Chinese (Simplified Chinese)
I turn backend noise into the smallest, cleanest evidence — with an honest declaration of its own boundary. I deliver trustworthy input, not conclusions.
This layer does one thing: give the agent clean, trustworthy evidence that honestly marks where its own sight ends. It does not draw conclusions for the agent, and it does not prescribe the investigation path. It reduces input noise and variance, and leaves the judgment of where the answer is to the agent that holds the full context.
So a Runbook here is not a path prescription — it is an evidence recipe + an exit condition:
- It declares which evidence this class of problem needs, and how to collect each cleanly (the recipe).
- It declares the falsifiable preconditions of its own hypothesis: once the evidence refutes those preconditions, it returns "not applicable — hand back to free investigation" instead of dragging the agent down a path that is already wrong (the exit condition).
A good Runbook's most valuable feature is not its path. It is knowing when to shut up.
Many online incidents are not hard because they are unique. They are hard because engineers keep replaying the same investigation sequence by hand:
- Compare actual behavior with the expected result.
- Check whether Redis is already wrong.
- Check whether the database source of truth is wrong.
- Check the trace to see where the workflow stopped.
- Decide whether the issue is stale cache, missing side effects, or abnormal persisted state.
Example:
- A detail page returns the wrong asset state.
- The expected investigation order is stable: inspect cache, inspect DB, inspect trace, inspect external dependencies.
- The useful input for the agent is also stable:
trace_id, expected result, actual result.
agent-debugger exists for that pattern. It distills these recurring troubleshooting habits into executable evidence recipes: the agent gets clean, bounded evidence instead of guessing in raw noise — without being stripped of its own judgment about where the answer is.
- A runbook selector that scores incident patterns and picks the best-matching investigation path.
- An executor that calls adapters in a fixed order defined by the runbook.
- Evidence normalization so tool output becomes compact, structured findings instead of raw payload dumps.
- A decision engine that emits an applicability and boundary verdict (applies / not-applicable then hand back) — it tells you whether this path still holds, instead of settling the root cause for the agent.
- An MCP server entrypoint so the investigation flow can be exposed to AI tools.
The zero-config path is the fastest way to understand the project. It uses replayable fixtures and does not require Langfuse, Postgres, or Redis credentials.
Requirements:
- Node.js
>= 18.17 pnpm
Run:
pnpm install
pnpm demo
pnpm benchmark
pnpm checkWhat you get:
- A runnable incident walkthrough from fixture input to structured report.
- A benchmark over the built-in replay cases.
- A metadata consistency check for runbooks, adapters, and evidence policies.
Important:
pnpm demoandpnpm benchmarkvalidate replayable investigation cases.- They are meant to prove the investigation model and repository structure, not to claim full production integration coverage.
The default demo replays this kind of incident:
- Actual: an order was created, but the downstream task was never generated.
- Expected: a task record should exist after order creation.
- Investigation order: trace -> persistence -> idempotency/cache.
The output shows:
- which runbook was selected
- which evidence items were confirmed
- which conclusion fired
- which next actions were recommended
After the zero-config demo, you can connect the MCP server to your own observability and storage systems.
Build the server:
pnpm buildCreate a config file:
cp agent-debugger.config.example.yaml agent-debugger.config.yamlExample:
adapters:
langfuse:
base_url: https://cloud.langfuse.com
secret_key: ${LANGFUSE_SECRET_KEY}
public_key: ${LANGFUSE_PUBLIC_KEY}
db:
type: postgres
connection_string: ${DATABASE_URL}
allowed_tables: [orders, tasks]
redis:
url: ${REDIS_URL}
key_prefix_allowlist: ["idempotency:", "task:idempotent:", "order:view:", "task:view:"]
runbooks:
- ./runbooks/request_not_effective.yamlAdd the MCP server to your AI client:
{
"mcpServers": {
"agent-debugger": {
"command": "node",
"args": ["/path/to/agent-debugger/dist/mcp/server.js"],
"env": {
"LANGFUSE_SECRET_KEY": "sk-...",
"LANGFUSE_PUBLIC_KEY": "pk-...",
"DATABASE_URL": "postgresql://...",
"REDIS_URL": "redis://..."
}
}
}
}Then provide a concrete incident:
Investigate
order_id=order_123. Actual: order was created but no task was generated. Expected: a task row should exist.
- It is not the original internal production system.
- It is not a generic autonomous bug-fixing platform.
- It does not ship the private DAG orchestration, permission system, or internal repair workflows from the original environment.
- It does not grant unlimited automatic repair authority.
- All adapters in this MVP are read-only.
- SQL queries are guarded against write operations.
- DB access is limited by a table allowlist.
- Redis access is limited by a key-prefix allowlist.
- Langfuse span fields are filtered by allowlist before being turned into evidence.
| Runbook | Scenario |
|---|---|
request_not_effective |
A request succeeded but the expected side effect did not happen |
cache_stale |
Cached state appears inconsistent with persistence |
state_abnormal |
Persisted business state itself looks incorrect |
Current built-in context coverage is intentionally narrow:
request_not_effective:request_id,order_idcache_stale:order_id,task_idstate_abnormal:order_id,task_id
If you want broader locator support such as trace_id or user_id, add a custom runbook through runbooks: in the config file.
Custom runbooks are supported through runbooks: entries in the config file. Each custom runbook should include sibling .selector.json, .execution.json, and .decision.json metadata files.
Incident Input (context_id + symptom + expected)
↓
[Runbook Selector] Matches signal weights via *.selector.json
↓
[Executor] Calls adapters in order defined by the runbook
↓
[Adapter Layer] Langfuse / PostgreSQL / Redis -> Evidence[]
↓
[Decision Engine] Emits an applicability & boundary verdict (applies / hand back)
↓
[Reporter] Structured IncidentReport
- Architecture Design
- Evidence Model
- Runbook Specification
- Adapter Specification
- Evaluation
- Release Checklist
- Release Announcement Draft
- v0.1.0 Release Notes Draft
- Changelog
- Security Policy
See CONTRIBUTING.md
MIT