Agent Autopsy Graph

Forensic memory for AI coding agents. Every failure becomes a guardrail.

A blackbox recorder that wraps an opencode runtime, records every tool call and rejection, builds a failure graph in Postgres + pgvector, and warns future runs before they make the same mistake.

autopsy.surf · install.autopsy.surf · architecture · demo script

From your project's root:

curl -fsSL https://install.autopsy.surf/install.sh | bash

This brings up the full local stack (postgres, FastAPI service on :4000, Next.js dashboard on :3000), installs the opencode plugin into the current project, writes AAG_URL to the project's .env, and prompts once for an optional Gemini API key (press Enter to skip; Autopsy works fully without it). ~/.autopsy/stop.sh brings everything back down.

Flags: --plugin-only (skip the stack, point at a remote service), --no-start (set up but don't launch), --no-prompt (skip the Gemini key prompt), --stop (tear down the running stack and exit). See install.sh --help.

opencode runtime
   │  events (tool calls, edits, chat, rejections, postflight checks)
   ▼
recorder plugin   ─POST─▶  AAG service  ─▶  Postgres + pgvector
                              │
                              │  classify failure → graph node + edges
                              │  embed task text → vector index
                              ▼
                          /v1/preflight
                              │
                              │  ANN + 3-hop traversal over the failure graph
                              ▼
opencode runtime  ◀────  system addendum + (sometimes) hard-blocks tool calls

Every failed run becomes a guardrail for the next similar task. The dashboard at localhost:3000 shows the live timeline, the failure graph, and which preflight hits Autopsy injected into the agent's system prompt.

For the long version see docs/architecture.md.

Prereqs: uv, node 20+, bun, docker (with compose).

cp .env.example .env
make dev                  # postgres + service + dashboard, Ctrl+C to stop
make plugin-link          # symlink plugin/src/index.ts into .opencode/plugins/
make seed                 # populate the graph with synthetic failures
make trace                # end-to-end preflight smoke test

Service + dashboard logs stream to the same terminal with [svc] / [dash] prefixes. Postgres keeps running between sessions; make compose-down stops it.

CI (lint + typecheck across all three packages, plus shellcheck on install.sh) runs on every PR via .github/workflows/ci.yml.

contracts/openapi.yaml, contracts/db-schema.sql, contracts/events.md, and contracts/fixtures/*.json are the source of truth for routes, schema, event mapping, and demo runs. Update them in the same commit when changing endpoints or tables. Ownership map in docs/ownership.md.

docs/demo-script.md has the exact commands. Summary:

1. Run a "schema field addition" task in opencode → user rejects with "missed migration".
2. Service stores the run, analyzer classifies, graph writer records the failure.
3. Start a similar new task → preflight injects a system addendum and (for high-confidence matches) blocks the offending tool call with a cited rationale; the dashboard shows the warning + blocked events on the timeline.
4. Agent does the right thing the first time.