[medium] perf(terminal): skip the ps/stty walk when /proc shows no controlling TTY

[elhoim]

BLUF

• Priority: medium. Removes up to 16 ps spawns per render on Linux when no ancestor has a terminal. The result is unchanged.
• Problem. When the /proc probe (perf: stop re-probing the terminal width on every line of every render #501) finds no tty on any ancestor's stdio, we fall through to the portable walk: 2 ps per ancestor (up to 8 ancestors), then tput cols. With no TTY, tput returns 80 or $COLUMNS. That result is a number, so the session "no TTY" cache never kicks in and every render pays for the whole walk again.
• Fix. The native walk already reads /proc/<pid>/stat. It now also reads tty_nr, the field ps -o tty= prints. If every ancestor the ps walk would visit has tty_nr == 0, that walk can't find anything, so we skip it. tput cols still runs, so the width is the same.
• Measured. In a detached no-TTY run, spawns per render drop from 8 to 1 and PATH probes from about 88 to 11. Stdout is byte-identical. CPU median falls 26% on the default config and 19% on the heavy config (25 interleaved passes).

Details

• terminal-native.ts adds probeTerminalNative(), which returns { width, noControllingTTY }, and parseTtyNrFromStat(). probeWidthNative() stays as a width-only wrapper.
• noControllingTTY is set only when all of these hold:
  • every ancestor's stat was read;
  • every visited ancestor has tty_nr == 0;
  • pid 1 also has tty_nr == 0 (the ps walk checks pid 1 too, e.g. a shell running as a container's init);
  • the ancestor at the depth limit, whose stat the walk would not otherwise read, also has tty_nr == 0.
• The full fallback still runs when:
  • a stat can't be read;
  • some ancestor has a controlling terminal but its stdio isn't a tty (redirected stdio, or a foreign-uid ancestor whose fd links can't be read);
  • the platform isn't Linux.
• getTerminalWidth precedence is unchanged: CCSTATUSLINE_WIDTH (feat(terminal): honor CCSTATUSLINE_WIDTH env var to override probe #380), then the session "no TTY" cache (perf: stop re-probing the terminal width on every line of every render #501), then the probe. The cache is still written only for a null result.
• The ps loop moved into probeAncestorWidth() without changes.
• Relation to feat(terminal): read COLUMNS from Claude Code before probing #612 (not ours; it also edits src/utils/terminal.ts, so expect a small merge conflict): the two complement each other.
  • feat(terminal): read COLUMNS from Claude Code before probing #612 reads COLUMNS before probing. On Claude Code 2.1.153+ that means zero spawns.
  • This PR covers the case where COLUMNS isn't set: older Claude Code, other hosts, and manual runs. Today that path costs 7 or more ps spawns plus tput. With this PR it costs one tput.
• Follow-ups (not in this PR):
  • tput cols without a TTY only reports $COLUMNS or terminfo's default (80), so it could be replaced by reading COLUMNS directly. That changes the no-COLUMNS result from 80 to null, so it belongs in its own PR.
  • macOS/BSD still runs the ps/stty walk on every render, and a numeric width is never persisted. feat(terminal): read COLUMNS from Claude Code before probing #612 fixes that for current Claude Code. A sysctl/libproc-based walk could fix it otherwise, but I can't test macOS here, so I haven't implemented it.

Measurements

Setup: setsid -f reparents the run to init, so no ancestor holds a pty. That matches Claude Code spawning the statusline with no terminal in its ancestry. COLUMNS unset, p1 payload, git cache warm.

Spawn counts (strace -f -e execve), per render:

arm	spawns	execve incl. PATH probes
main	8 (7 × ps, 1 × tput)	88
patched	1 (tput)	11

Stdout from main and patched is byte-identical on the default, git0, heavy and livelike configs, and also with COLUMNS=137. No terminal-width.json cache is written in either arm, which confirms the result is numeric (80).

Interleaved bench: 25 passes, round-robin, CPU is user+sys including reaped children. Host load1 was min 53.9, median 61.3, max 66.8 on 6 cores, so compare ratios rather than absolute numbers.

arm	CPU median (ms)	CPU p90 (ms)	wall median (ms)
control node -e 0	114	145	820
main, default config	1989	2087	15621
patched, default config	1471 (−26%)	1536	11167
main, heavy config	2318	2658	18304
patched, heavy config	1880 (−19%)	2135	15037

Checks

• bun run lint (tsc + eslint): clean.
• bun run build: OK.
• bun test src/utils/__tests__/terminal*: 57 pass. New tests:
  • parseTtyNrFromStat;
  • conclusive and inconclusive cases for probeTerminalNative: tty_nr set on an ancestor, on pid 1, and at the depth limit; unreadable stat; our own tty_nr ignored; non-Linux;
  • in terminal.test.ts, a conclusive result gives exactly ['tput cols'], an inconclusive one still runs the ps/stty walk, and a native width spawns nothing.
• Full bun test on this loaded host: 2297 pass, 74 fail. All failures are timing-related (fetchUsageData error handling, custom-command capture, Ink TUI menus). Re-running only the affected files on unmodified main gave the same result there (174 pass, 69 fail on both). None of the failures are in terminal code.

🤖 Generated with Claude Code