Inside the 05:15 job-hunt run - six legs, one LLMCapability

Your 05:15 job-hunt run is six legs in a row, and only ONE of them is an LLM. Reading nightly_wrapper.sh top to bottom is the clearest picture of how your system splits judgment from verification.

Six legs fire in order. (0) Network settle — up to 5 minutes waiting for DNS/WiFi after sleep, because roughly half the pre-dawn failures used to be ConnectionRefused before the wait existed. (1) Content — `claude -p /jobhunt-nightly` under a 30-minute watchdog: it sweeps listings, selects 3–5 leads, and writes leads/briefings/YYYY-MM-DD.md. This is the ONLY leg that needs judgment. (2) Render/send — brief_render.py --send turns that markdown into the paper doc and pushes it to the reMarkable 'Job Applications' folder. (3) Verify — verify_nightly.py, which ALWAYS runs. (4) Watch — lead lifecycle alarms. (5) Pursuit — pursuit.py --max 2 drains the GO-inked queue into staged application packages, each under its own 40-minute watchdog.

The load-bearing detail is which exit code the wrapper returns. The content leg's result is captured as RC, but the script ends with `exit $VRC` — the VERIFY leg's code. That's deliberate: a morning where the LLM ran fine but the briefing never reached the device is still a failed morning, and only the verify leg, checking the artifact and the cloud manifest, can tell. Trust the leg that inspects the side-effect, not the leg that did the work.

One more rule worth internalizing: ink_apply.py is the ONLY writer of your briefing decisions. When you mark a lead GO or NO on paper and tag it done, ink_apply is the single door that updates the pursuit queue, the refinements file, the lead record, and the seen-state. One writer per piece of state is why the three JobHunt ledgers don't drift into contradiction the way the reMarkable manifests did.

Worked example

The wrapper documents itself — read the leg map and last night's exit codes.

sed -n '1,20p' /Users/tom/Claude/JobHunt/scripts/nightly_wrapper.sh
grep -nE 'RC=|VRC=|exit \$VRC' /Users/tom/Claude/JobHunt/scripts/nightly_wrapper.sh
tail -20 /Users/tom/Claude/JobHunt/logs/nightly.log

• The header comment IS the spec: six numbered legs with their runtimes and watchdogs.
• The grep shows RC captured from the content leg and VRC from verify — and that `exit $VRC` is what the LaunchAgent (and your alert ledger) actually reads.
• The log tail shows each leg's real exit code from last night — 124 on the content leg means the 30-min watchdog killed it.

Gotchas

• The render leg is guarded by `if [ -f .../briefings/$TODAY.md ]` — no briefing source, no send, and the log says so; that's the upstream-aware fallback (rule 4), not a silent skip.
• The content leg runs with --dangerously-skip-permissions because launchd can't answer a prompt — that headless fence is scoped to this one leg chewing untrusted web listings, not a blanket setting.
• Pursuit stages packages but NEVER submits — the system's hard line; anything that looks like an auto-send would be a bug, not a feature.