job-search

name	job-search
description	Run the full job search harness — seek fresh postings, score them, and prepare tailored resumes for the top matches (cover letters optional, offered at the end).
allowed-tools	Read, Write, Bash, Agent(job-seeker, job-preparer)

Run the complete job search pipeline.

Step 0: Determine Active Sources

Source selection is data-driven and per-user — it lives in the harness DB and the user manages it from the TUI/web Settings → Sources panel. Do not prompt the user about sources, and do not write any config file. The job-seeker agent reads the enabled set directly from the DB (harness-db sources enabled).

The 7 available sources are: linkedin, indeed, adzuna, ziprecruiter, greenhouse, remotive, research

Optional transient override (args only): if the skill was invoked with args, treat them as a one-run override that does not change the stored selection:

• --skip=<csv> → run all enabled sources except these (e.g. --skip=research)
• --only=<csv> → run only these (e.g. --only=linkedin,adzuna)

To apply an override: run harness-db sources enabled to get the stored set, then split on commas / trim / lowercase the arg and subtract (--skip) or intersect (--only), and pass the resulting enabled_sources list to job-seeker in its spawn prompt. With no args, spawn job-seeker with no source override — it reads the DB itself.

Steps 1–5

job-preparer cannot prompt the user — it runs as a subagent and its questions do not surface. You (the main agent running this skill) own every user decision. job-preparer runs in phases and returns control to you between them. Ask the user in plain text (no AskUserQuestion), matching the Step 0 style above.

1. Search. Spawn the job-seeker agent (subagent_type: job-seeker). It reads the enabled sources from the DB (harness-db sources enabled) — or uses the transient enabled_sources override if you computed one in Step 0 — searches those sources in parallel for remote, Canada-eligible senior roles, deduplicates against the SQLite DB, inserts new postings, and saves an audit log to job-data/jobs/search-YYYY-MM-DD.json. Wait for it to complete.

   Present the detailed search report. job-seeker returns a detailed report (also written to job-data/jobs/search-report-YYYY-MM-DD.md) with three sections: (1) every source and the number of positions found in each, (2) any content-fetch problems, and (3) any other execution issues. Show this report to the user verbatim before moving on to scoring — do not summarize it away. If the agent's reply did not include the report, read it from $JOB_DATA_ROOT/jobs/search-report-YYYY-MM-DD.md and present that.

2. Score & rank. Spawn job-preparer with phase: score. It scores every new/stale posting in parallel and returns a ranked top-N table (N is set by the JOB_TOP_N per-user config value, default 5) — scoped to the current batch (the most recent scoring date), not all-time — that includes each job's URL. Wait for it to complete.

3. Ask which jobs to prepare. Present the ranked table to the user (you may drop the URL column for readability). Then ask, in plain text:

   Reply with the rank numbers of the jobs to prepare (e.g. 1 3 or 1, 2, 4), or none to stop.

   Wait for the reply. Parse space/comma-separated integers; map each rank back to its URL using the table from step 2. If the reply is none or contains no valid ranks, stop here — the run is complete.

4. Prepare resumes. Spawn job-preparer with phase: prepare and selected_urls: [<the chosen URLs>]. It marks them selected, prepares a tailored resume + PDF for each under job-data/output/YYYY-MM-DD/, writes final-report.md (cover letters shown as "not generated"), and returns a prepared_jobs handoff (company, url, output_dir, resume_yaml_path per job). Wait for it to complete.

5. Offer cover letters, then optionally generate them. Cover letters are off by default. Ask the user, in plain text:

   Resumes are prepared and the final report is written. Generate cover letters too? Reply yes to generate them (or name specific companies), or no to finish.

   Wait for the reply.

   • If no / empty / no clear affirmative: stop — report the summary from step 4.
   • If affirmative: spawn job-preparer with phase: cover-letters and prepared_jobs: [<the handoff from step 4, filtered to any companies the user named>]. It generates the cover letters, updates final-report.md, and returns a summary. Wait for it to complete.

Finally, report the summary table (rank, company, title, score, status) including whether cover letters were generated, and the path to final-report.md.