[joshrotenberg]

Spike, not for merge. It implements the agent operations (agent, agent_result, agent_cancel, agent_stop, agent_sessions) directly over claude_wrapper 0.19 and codex_wrapper 0.9. One GenAgentServer.Session process per session holds the wrapper session struct and its provider resume id. No GenAgent server, Ensemble session or backend adapter is on the turn path.

What it showed

• Run 6 used this branch for three patterns in parallel: a single-use subagent, a counterpart session, and a recursive session that started its own reviewers through agent. It produced five ready PRs. The record is in docs: procedure for working one issue through the server #134, under docs/runs/2026-11-01-agent-patterns-run6/.
• For CLI backends the turn path needs one process per session plus the wrappers. The trace and the evidence are in the interface note (docs: agent session spec and interface note #139).
• Two wrapper gaps found by the spike were fixed upstream: codex_wrapper feat(agent): keep a bounded record of ended sessions #142, released in 0.9.0, and claude_wrapper #338, released in 0.19.0.

Why it does not merge here

This package depends on GenAgent, Ensemble and both adapters for its instance path, so a session layer inside it cannot show that the wrappers are enough on their own. It also needs override: true on both wrappers, because the adapters cap their versions. The plan in #139 ("Where it lives") moves the session layer into a wrappers-only library and adds an MCP and CLI adapter. gen_agent_server then becomes the library's second consumer. This branch is the seed for that library's first slice.

Checks

CI 9/9. The base suite has 141 tests, including six session tests against a fake codex. Every extension suite and the MCP release smoke pass. A live two-turn smoke passed on both providers from the dogfood release.