Repository navigation
Harness lifecycle and continuity
An agent runtime or harness maintains a temporary tool loop and possibly a local session cache; ACP maintains the long-lived project/conversation and task ledger. Model inference can be served through vLLM/LiteLLM or a similarly configured route, but the model must not control identity or resource grants. Slurm/Kubernetes remain the resource schedulers.
A user should be able to create a private project in Quantum Guide, attach from a permitted second client, restart a worker, change a compatible model/harness route, and still see the same canonical messages and IDs. A harness restoring its private session without ACP history is a failure of the architecture.
The worker reads its input from a trusted API/queue with a scoped AgentPrincipal and task identifier; it reconstructs context from PostgreSQL rather than assuming the browser or local runtime survived. Enforce bounded run duration, cancellation, retries and a single committed final result per idempotent task. Persist tool provenance and observed errors. A model's own statement “I restarted successfully” does not demonstrate restart success.
Use an isolated authorized test user and a distinctive nonsecret marker. Record project/conversation IDs, send a message, end browser/Jupyter/terminal sessions, replace the worker and clear ephemeral runtime state (not ACP database), then retrieve the same canonical transcript. Verify wrong-user/tenant access is denied. Do not execute destructive production tests without explicit operator approval and recovery plan.
An alternate provider/harness should read canonical context through the same protected contract and write back the same conversation; simply starting a second agent with a copied prompt is insufficient. Test misrouting, slow inference, cancellation, token budgets, forbidden tool calls, missing model server and persistence after runtime crash.
The current source provides a thin acp client that can accept a trusted short-lived assertion on standard input. Portal-mediated terminal pairing is a proposed contract (ADR 0012), not a deployed login broker. Never tell users that SSH/POSIX credentials alone grant ACP access.
References: M4 persistent agent, personal worker source, runtimes and sandboxes, terminal ADR.
ACP
- Home
- Architecture
- Researcher guide
- Clients and integrations
- Models and runtimes
- Operator runbook
- Developer guide
- API and data contracts
- Security and tenancy
- Troubleshooting
- Contributing
- Guide help and support Engineering reference
- Database RLS and migrations
- Memory lifecycle and provenance
- Harness lifecycle and continuity
- ACP release and conformance
- Agent security checklists