The root cause is the liveness oracle itself. `ensure_chrome.sh` in the `scent-media` repo decides whether Chrome is alive by attempting a connection to CDP port 9223, and treats no response as a dead process to be force-killed [9]. What that probe actually measures is whether the running Chrome exposes a debug port. The `th-scent` job in the separate `social-autolike` repo drives Chrome through Playwright and never opens 9223 [11], so a perfectly healthy browser answers the probe the same way a hung one does [12].
Then look at the blast radius, which is inherited from the pattern. The command is `pkill -f "user-data-dir=$PROFILE"`, with `$PROFILE` set to `.profiles/chrome-ig` [10]. The identity being matched is a directory, and that directory is exactly what the other repo was configured to reuse, via `"reuseProfile": ".profiles/chrome-ig"` in `config/accounts.json` [11]. Neither codebase has to reference the other for the kill to land. This is what happens when a handful of jobs becomes a fleet on one machine: process identity leaks out of the repo and into the filesystem, and that filesystem is shared across jobs.
The data loss follows from the signal pair. `pkill -f` sends SIGTERM then SIGKILL, so Chrome cannot finish its shutdown routine and loses its chance to write the profile's cookie SQLite back consistently [14]. On the next launch it recreates an empty cookie DB, and the measured post-mortem is three rows across all hosts, zero cookies for `instagram.com`, and no IG `sessionid` [15]. On 2026-08-24 the 19:00, 21:00 and 23:00 slots all posted nothing [16].
The timing was not luck. `th-scent` unfollow grabbed the profile at 22:56 and `daily-post` called `ensure_chrome.sh` at 23:00 [13], a four-minute window [20] that launchd hit on schedule every night. The scheduler was the only component doing precisely what it was told.
Detection is the other half, and it failed earlier and for a different reason. The author reports that on 2026-08-08 the `ig-1` like lane exited 0 with zero likes on 8 of 12 daily runs [5], roughly two thirds [19], because the watchdog counted occurrences of a completion marker and therefore scored circuit-break and need-login as healthy [6]. An earlier gap was worse: X, Instagram, TikTok and Threads posting was watched by nobody, and IG posting stopped from 7/29 for two weeks before anyone noticed [4]. Adding `count_dead_runs()` and `sum_likes_today()` fixed the reporting [7], not the structure, since the breaking side and the reporting side live in different repositories [8].
Reaching this failure requires a specific combination: one host, headful Chrome, a profile directory reused by two independently scheduled jobs, and a kill by command-line pattern rather than by a PID the caller started. If each job gets its own `user-data-dir`, the whole class evaporates. According to the dev.to write-up, the account being logged out nightly sits underneath ¥1.2M/month in revenue [17], which is a good argument for the cheap fix being isolation rather than more monitoring.
