(7/7) feat(session): launch the supervisor topology; delete the single-process chassis

[guapisolo]

Summary

The switchover: start_session_server now always builds and starts a SessionServerSupervisor — one topology for every worker count, so the if workers > 1 launch branch never exists — and the single-process chassis is deleted in the same move.

Stacked on #1569 → #1510 → #1563 → #1518 → #1565 → #1566. Base = refactor/session-mp-test-harness. This PR was split for reviewability: the process-lifecycle layer is #1565, the test-harness migration (the equivalence evidence for the deletion below) is #1566. Merge order: #1565 → #1566 → this.

What's here

• router_manager.start_session_server — always supervisor; returns it. rollout_manager holds it and calls check() at the start of both generate() and eval(), so a dead worker fails the rollout loudly instead of black-holing its hash shard.
• arguments.py — --session-server-workers (int, default 8: idle workers cost ~zero CPU, and under concurrent multi-turn load GIL-bound tokenization/deserialization scales with cores; memory scales with N — each worker loads its own tokenizer).
• sessions.py + server.py deleted. Two chassis duplicate the route surface (the same six routes defined twice) and the proxy stack (SessionServer.do_proxy vs the worker's ProxyBackend, held identical only by equivalence tests), and every future behavior change pays that tax twice. The old workers=1 server was already a spawned child; the delta is one extra tokenizer-free router child and a per-request IPC hop measured in µs against LLM-turn latencies.
• Dev doc + header wording updated to the one-chassis reality (docs/developer/multi-process-session-server.md, core.py, worker.py, processing_utils.py).

Testing

• pytest tests/fast/router — 96 passed at this commit (assertions carried unchanged through (6/7) test(session): InloopSessionServer replaces the in-process SessionServer test harness #1566); generate_hub/test_multi_turn + inference_rollout/integration suites pass on the reworked fixtures.

Not here

Round-robin / no-HOL IPC stays cut (benchmark-gated, see the dev doc). The R3 client-strip landed below this stack in #1563. The worker-direct records path (decided, see the dev doc) and the benchmark port are follow-up PRs.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces an opt-in multi-process session server supervisor (SessionServerSupervisor) that spawns and monitors multiple worker processes and a router process to shard sessions, controlled by the new --session-server-workers argument. The review feedback highlights several critical reliability and resource management improvements: ensuring the supervisor is explicitly shut down in RolloutManager.dispose() to prevent port-binding conflicts, handling process spawning failures gracefully to avoid orphan processes, unregistering from atexit to prevent memory leaks, and making the supervisor's shutdown method thread-safe using a lock.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[guapisolo]

A successful CI run https://github.com/radixark/miles/actions/runs/28771003169

[Shi-Dong]

LGTM! Great job.
But let's postpone merging the chain now, as it's intrusive and might break what is currently working on main.