feat(coding-agent): let root agents create sibling sessions (RES-1256)

[sethkarten]

Summary

• add await rlm.create_session(prompt, name=..., cwd=..., model=..., thinking=...) for daemon-backed depth-0 agents
• create a separate resident root session through the existing supervisor create and prompt operations
• keep normal rlm(...) calls recursive and reject root creation from nested or inline sessions
• return a typed Python handle with the new session identifiers, path, name, and model
• require the updated runtime during kernel bootstrap and document the operation in the root-agent prompt

Test plan

• npm run check
• env -u RLM_DEPTH -u RLM_MAX_DEPTH -u RLM_SESSION_DIR npx tsx ../../node_modules/vitest/dist/cli.js --run test/agent-session-recursion.test.ts test/daemon-mode.test.ts test/system-prompt.test.ts test/ipython-bootstrap.test.ts test/kernel-bootstrap.test.ts
• uv run python -m unittest test/test_create_session.py test/test_subagent_registry.py

---

Note

Medium Risk
Spawns new daemon sessions with model/auth and supervisor RPC; misconfiguration or partial failure is mitigated by validation and kill-on-error, but session lifecycle and naming still affect multi-agent orchestration.

Overview
Adds rlm.create_session so depth-0, daemon-backed root agents can spin up a separate resident top-level session (a sibling root), not a recursive child via rlm(...).

The coding-agent wires a new kernel host request rlm.create_session to AgentSession.createRlmSession, which validates prompt, optional name / model / thinking / cwd, enforces depth-0 and a host with createRlmRootSession, then delegates to the daemon. The daemon worker talks to the supervisor with create (lifecycle: resident) and prompt, returns session identifiers, and kills the new session if prompting fails.

The Python prime-agent-runtime exposes await rlm.create_session(...) (and on rlm.rlm) returning RLMCreateSessionHandle. Kernel bootstrap now requires callable create_session. Depth-0 RLM system prompts document the API. Shared RLM kwargs normalizers accept an operation label so error messages apply to both rlm.run and rlm.create_session.

^{Reviewed by Cursor Bugbot for commit 0c9c883. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add rlm.create_session API for daemon-backed root agents to create sibling sessions

• Depth-0 daemon-backed agent sessions can now create a separate resident root session with an optional name, model, thinking level, and working directory, and receive its session IDs and file path.
• The new rlm.create_session Python API dispatches a host request through the kernel and returns an RLMCreateSessionHandle typed result.
• Runtime readiness probes now require callable rlm.create_session on both the top-level RLM module and its callable namespace.
• AgentDaemon.createRlmRootSession connects to the supervisor, creates and prompts the resident session, and cleans up partially-created sessions on failure.
• Risk: runtimes missing the required rlm.create_session entry points now fail RUNTIME_READY_CHECK in bootstrap.ts; nested, disposed, or inline sessions reject the operation via AgentSession.createRlmSession in agent-session.ts.

^{Macroscope summarized 0c9c883.}

[snimu]

rlm.create_session sidesteps the spawn policy that rlm.run enforces. _startRlmChildRun refuses to spawn once _rlmDepth >= _rlmMaxDepth, but createRlmSession (agent-session.ts ~L10796) has no equivalent gate, and the siblings it creates are fresh roots: default rlmMaxDepth 2 plus their own create_session, so an agent with spawning disabled can still grow the session tree without bound. The prompt paragraph advertising create_session (prompts/rlm.ts ~L149) is gated on depth === 0 && hasIpython only, so a spawn-disabled agent is actively told about it. This is a resource guard, not a security boundary — bash can reach the supervisor socket anyway — but the inconsistency is worth closing.

Minimal fix:

• reject in createRlmSession when _rlmMaxDepth === 0
• gate the prompt paragraph on allowRecursion too

[written by prime-agent, checked by snimu]

[snimu]

createRlmRootSession (daemon-mode.ts ~L2501) sends a hand-built config subset — cwd, agentDir, sessionDir, provider, model, thinking, telemetryDisabled — with no launchEnv and no apiKey. The normal root-create path (main.ts ~L1007) sends the full config plus launchEnv: collectDaemonLaunchEnv(). The model preflight runs in the parent worker against the parent's credentials, but the sibling launches from the supervisor's ambient env, which can be older: after a new PRIME_API_KEY, --api-key, or custom provider env, create_session succeeds and the sibling's first prompt fails auth.

Suggested fix: forward the current worker's filtered launch env and spread runtimeConfig.apiKey into the config, same pattern as the existing sessionDir/telemetryDisabled spreads.

[written by prime-agent, checked by snimu]