[Ark0N]

When Claude sessions are created at the same moment (the codeman skill's spawn_workers, or a multi-tab Run), some of them can silently end up outside tmux: they run on Codeman's direct-PTY fallback, so they don't survive a server restart, while the create call still reports success. This PR fixes the cause, a temp-file name that wasn't unique, with a one-line change plus a regression test. No API, UI or behavior changes otherwise.

Cause

Every claude launch goes through ensureStatusLineExporterScript() (src/hooks-config.ts), which keeps ~/.codeman/statusline-exporter.sh up to date for the plan-usage chip. When the script's content needs changing (a fresh data dir, or a release that changes the script), it writes a temp file and renames it into place, so live sessions never execute a half-written file.

The temp name was statusline-exporter.sh.<pid>.<Date.now()>.tmp. Within one server process, two creates in the same millisecond produce the same temp path. The first writer renames it away, and the others then fail with ENOENT on chmod or rename. TmuxManager.createSession() treats that as a mux failure, and Session falls back to a direct PTY:

[Session] Failed to create mux session, falling back to direct PTY: Error: ENOENT: no such file or directory, rename '/home/…/.codeman/statusline-exporter.sh.2505760.1791133404959.tmp' -> '/home/…/.codeman/statusline-exporter.sh'

Fix

The temp suffix now comes from randomBytes(6), the same pattern already used by the skill-file writer in the same file and by user-store.ts. Concurrent writers each rename their own complete, executable copy into place; the content is identical, so whichever lands last is correct.

Testing

• New test in test/hooks-config.test.ts: freezes Date.now() (which is what two creates in one millisecond amount to), runs eight refreshes at once, and checks they all succeed, the script is fresh and executable, and no temp files are left. On master it fails with the same ENOENT ... rename error as the log line above; with the fix it passes. hooks-config and statusline-cli-flag suites pass (104 tests), and tsc, eslint and prettier are clean.

• End to end on a fresh isolated instance (CODEMAN_INSTANCE, its own tmux socket, data dir and cases dir), firing four concurrent claude POST /api/quick-start calls with real Claude Code 2.1.289:

  	round 1	round 2
  master	2 of 4 in tmux, 2 direct-PTY fallbacks	2 of 4 in tmux, 2 fallbacks
  this PR	4 of 4 in tmux, 0 fallbacks	4 of 4 in tmux, 0 fallbacks

  On master, quick-start returned success for all four in both rounds.

Notes

• How often this happens in practice: only while the script needs rewriting, so on a long-running install it's the first concurrent creates after an upgrade that changes the exporter script, or on a fresh install. When it hits, the affected sessions lose tmux persistence without any visible error.
• src/custom-model-hosts.ts has a similar fixed temp name (<file>.<pid>.tmp) on an async write path, but its writers are user-driven route handlers and it also has a separate read-modify-write question, so it's left out of this PR.
• No changeset here; it goes in with the release.