Getting help

Before you report

1. Run the doctor. If a line says FAIL, its reason is often the whole answer.
2. Check the common failures table for the exact message you see.
3. Try the same task on a stronger model if you can. A weak local model can produce failures that look like bugs but are a capability ceiling; note whether it still happens.

Make a bundle

podman compose exec agent-server python -m disco.agent_server.doctor --bundle /tmp/disco-doctor.json
podman compose cp agent-server:/tmp/disco-doctor.json .

The bundle is one JSON file containing the build identity, the check results, your DISCO_* settings with every secret value replaced by ***REDACTED***, both servers’ health bodies, the sandbox probe, and the shape of your last ten conversations: event kinds, tool names, exit codes and error text. It never contains prompts, files, model output or keys. Read it before attaching it if you want to be sure.

Where to post

• Bug report — the template asks for the surface, deployment mode, OS, model, steps, the doctor table and the bundle. Paste the log window around the failure from podman compose logs agent-server, redacted.
• Discussions — usage questions, deployment help, ideas.
• Security — report privately; never in a public issue. See the security policy.

What a good report has

• The doctor table and the bundle, or the exact reason you could not run it.
• What you expected, what happened, and the smallest sequence of steps that reproduces it.
• The build: the version from /health, which the bundle also carries.
• Logs from the window around the failure, with anything that looks like a key, token or cookie removed.

Deeper diagnostics

With DISCO_INSPECT=1 in .env and the agent-server restarted, /api/debug/trace/<conversation id> returns routing decisions, per-step model timing and tool scopes, and /api/debug/evidence/<conversation id> the full redacted event log of one run. Both are owner-scoped. The bundle is the lighter export; use these when a maintainer asks.