llm-job-pipeline: Command for Claude Code

/jobs-night — one preparation gate, unattended

The default daily run emits prepare_screening with discovery payloads: one call per vacancy. Unscored roles request a numeric score and quoted facts/profile comparison; existing roles below 40 request only stale or missing facts; roles at or above 40 are untouched. The other gate names remain legacy paths.

Arguments: $ARGUMENTS = <gate> <night_dir> <phase> where <gate> is one of screen_companies | score_companies | score_vacancies | prepare_screening, <night_dir> is this night's private directory (vacancies/nightly/<date>/), and <phase> names the pass inside the gate (screen | escalate for score_vacancies, screen for screen_companies, score for score_companies, prepare for prepare_screening).

This is the NIGHT variant of the /jobs-new gate protocol. The pipeline orchestration already ran in scripts/run_daily.py --unattended; your only job is the judgment for ONE gate. The wrapper that launched you resumes the driver after you stop — you never run the driver yourself. Everything below overrides the interactive habits of /jobs-new.

Overrides — read first, they are absolute

1. Never ask a question. There is no human. If something is ambiguous, take the safe direction (skip the item, record it failed in the log) and keep going.
2. Subagent model comes from the wrapper. Read <night_dir>/session.json once and pass its model to EVERY scoring subagent. Python resolved screen_model(), scoring_model() or company_screen_model() for this gate and phase before starting you. Never use a shell to discover it.
3. Spawn only the night-scorer agent type. One payload file = one night-scorer subagent. No other agent type, ever. At most 5 subagents at a time (rolling waves). 1 item = 1 subagent — batching is untested here.
4. File in, file out. Each subagent reads its own payload <night_dir>/score_in/NNN.json and writes its one result to <night_dir>/score_out/NNN.json (same NNN). Subagents have Read and Write only. Neither you nor a subagent has shell, database or network tools.
5. Write each result immediately. The Python wrapper saves completed JSON files while you work and performs a final sweep when you stop or time out. Never run a save command yourself. A malformed file is skipped independently.
6. Log after every wave. Append one line per wave to <night_dir>/scoring_log.md: time, gate, items in the wave, written count, failures with their NNN. Findings and anomalies go HERE, not to an issue tracker.
7. Never resume the driver. The wrapper owns the driver loop: it runs --resume itself after your session ends. When every score_in item is written or recorded failed in the log, finish the End-of-session step below and STOP.