Managed session log file is never written

[csvance]

Bug: Managed session log file is never written

Summary

ManagedSession.log_file is constructed and quoted to the user in error messages, but no code path ever writes to it. When a session subprocess crashes, the agent is told "Check log: " and finds nothing there.

Reproduction

1. Call mcp__kaimon__start_session with a project whose subprocess crashes during boot.
2. Tool returns Error: Session failed to start — Process died at <ts>\nCheck log: <path>.
3. The reported path does not exist (or is empty).

Root cause

In src/session_manager.jl:

• ManagedSession(project_path; ...) (line ~23) sets log_file = joinpath(log_dir, "$(basename(project_path)).log").
• spawn_session! (line ~122) launches the subprocess via Tachikoma.pty_spawn(cmd; rows = 24, cols = 80, env) and stores only the PTY handle in ms.pty. Stdout/stderr go to the PTY master fd; nothing drains them to disk.
• Error messages in src/tool_definitions.jl:726,730 quote ms.log_file unconditionally.

Compare with spawn_extension! in src/extension_manager.jl:236-240, which actually opens and redirects to the log file:

log_io = open(ext.log_file, "a")
println(log_io, "\n--- Extension ... starting at ... ---")
flush(log_io)
proc = run(pipeline(cmd; stdout = log_io, stderr = log_io); wait = false)

It also closes log_io in the exit-watcher @async block.

Impact

When a managed session fails to start (bad project, missing dep, instantiate error, version mismatch), the only surfaced error is the generic "Process died at " recorded by _push_session_error!. The actual stack trace and Julia stderr output are lost. This makes debugging session-startup failures effectively impossible without re-running the same command manually outside Kaimon.

Related files

• src/session_manager.jl — ManagedSession struct (line ~11), spawn_session! (line ~122)
• src/tool_definitions.jl:726,730 — error messages that quote ms.log_file
• src/extension_manager.jl:230-267 — reference implementation that does write its log