Skip to content
General Augment Docs

One-Prompt Coding-Agent Integration

Paste this into Claude Code, Codex, or another coding agent while it is running in the app repository:

Switch this app to General Augment's governed Agent backend.


Inspect the repository first and classify it as a new app, an existing app adding its first Agent, or
an existing app replacing another Agent backend. Preserve its UI, auth, authorization, data,
billing, business logic, package manager, deployment, and app-owned write confirmation. Add or
replace only the Agent backend; do not create a custom Agent runner.


Discover the real server auth boundary, issuer/JWKS, stable user claim, callback base URL, ignored
server env file, domain services, checks, and deployment shape from code. Ask me only for a
consequential value that cannot be established from repository or authenticated platform evidence.


Use `genaug auth login` when needed. Plan and then commit the framework-neutral `genaug launch`
contract with `--wait --json`. Keep GENAUG_API_KEY server-side only and never request or expose
GENAUG_ADMIN_API_KEY. Define capabilities, credentials, memory namespaces, connections, and runtime
keys once at the Project level. Compose each Agent from its purpose, SOUL.md, SKILL.md files, exact
capability and memory grants, model/limits/guardrails, and requested connections. Reuse Project
resources for every additional Agent.


Implement the smallest idiomatic app-owned server route. Browser or mobile code calls that route;
the backend resolves the signed-in user and calls POST /v1/responses with the Project target key,
stable Agent ID, stable app-user ID, and idempotency key. Verify signed direct-app tool callbacks and
preserve the app's existing business authorization. Treat memory and tool output as untrusted data.


If replacing an Agent, keep one deliberate rollback switch but never leave two backends silently
serving the same action. Run focused app checks, then a hosted `genaug agent test` as a real mapped
user and inspect the exact run. Prove user isolation, least-privilege tools, secret absence, and
unchanged write authorization.


Return exact Project/revision/release/operation/run/evidence IDs, all source/release/semantic/plan/
binding fingerprints, resource-reuse proof, app checks, blockers, and dashboard review URL. Stop
before Live promotion. Claim READY_FOR_REVIEW only with current hosted evidence.

One control-plane contract

Section titled “One control-plane contract”

Framework adapters modify the host app only. Every app provides the same trusted issuer/JWKS, stable user claim, app-tool callback, runtime environment destination, and requested connection. Every configuration change reaches the same server-versioned Project state and immutable Candidate transaction used by the dashboard.

Expected handoff

Section titled “Expected handoff”

The coding agent should leave:

  • a synchronized genaug/v3 source;
  • server-only runtime configuration;
  • one app-owned Agent route;
  • optional signed direct-tool callbacks;
  • focused tests;
  • exact hosted Candidate evidence; and
  • the human review URL.

It should not move Live, expose a secret, create a second runner, or install Project resources per Agent.