fix(desktop): detach managed agents from the controlling terminal (setsid)

[matthiasdebernardini]

Problem

When the desktop app is launched from a terminal (just dev), every managed-agent process it spawns — the buzz-acp harness, the claude-agent-acp adapters, the agent engine, and its MCP servers — inherits the terminal's controlling TTY. The harness was spawned with process_group(0), which makes the subtree a background process group of that terminal session. macOS job control then intermittently stops the actively-working agent's process group with SIGTTIN-class stop signals.

Observed live: an agent turn froze mid-Bash tool call with the engine, both its MCP servers, and finally the adapter all in ps state T (stopped). The turn hangs indefinitely; the desktop shows the agent "working" forever (the typing/status indicator loops), and the reply only escapes if someone manually kill -CONTs the process group. Packaged .app launches have no controlling TTY and are unaffected — this only bites development runs, but it makes agents unusable under just dev.

Fix

Replace process_group(0) with a pre_exec setsid() in spawn_agent_child (the sole long-lived harness spawn site). setsid() puts the whole agent subtree in its own session with no controlling terminal, so background-TTY job control can never stop it. If setsid() fails (EPERM when already a group leader), fall back to setpgid(0, 0) best-effort — exactly the pre-patch behavior — and never fail the spawn.

Invariants preserved:

• setsid() makes the child both session and process-group leader, so pid == pgid still holds and every existing kill path (sigterm_then_sigkill, orphan-sweep kill(-pid, …)) is unchanged. The orphan-cleanup comment even anticipated setsid'd children already.
• Orphan cleanup after a desktop crash is unaffected: the BUZZ_MANAGED_AGENT marker + PID-receipt sweep was already the safety net (the harness never sat in the shell's job list even before this change).
• Windows (Job Object) path untouched; #[cfg] gating mirrors the code it replaces.

The unsafe in pre_exec is minimal (only async-signal-safe setsid/setpgid) and follows the existing libc::kill precedent in this file. Stale comments in storage.rs / process_lifecycle.rs that cited process_group(0) as the pid==pgid mechanism are updated.

Tests

• New #[cfg(unix)] test spawned_harness_is_session_leader_in_new_session: asserts a child spawned with the same pre_exec becomes a session leader (sid == pid) in a session distinct from the test process, with pgid == pid so the group-kill invariant holds. (It exercises the setsid pattern on a stand-in child — spawn_agent_child needs an AppHandle — see the test doc comment for the limitation.)
• cargo test --manifest-path desktop/src-tauri/Cargo.toml managed_agents — 682 passed.
• cargo clippy --manifest-path desktop/src-tauri/Cargo.toml --all-targets -- -D warnings — clean; cargo fmt applied.
• Reproduced the freeze locally under just dev (agent turn stuck with adapter/engine/MCP in state T), and verified kill -CONT on the stopped group let the turn complete — confirming background-TTY stops as the mechanism this patch removes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PCWMx2BkoFDZtFRiDny16W