JobRadar — stage 1: search and hard elimination

Invariants (identical in both JobRadar skills; they override everything below)

1. A human submits every application, even when the user says "go ahead" or "我同意". Stop before any submit, send, upload, recruiter message, "next step" or declaration that could file one; login, captcha and declarations belong to the user.
2. The instance's Profile/ (hard conditions, career profile, self-description) holds only what the user chose to share with the agent: read it directly. Résumés, identity documents and other personal files are read only through the user's own privacy tooling or with their explicit permission for that file. Profile text never goes to Feishu, and career-profile aspirations are never résumé facts.
3. Permanent IDs (JOB-000001, or LEAD-YYYYMMDD-NNN for direction leads) are the only record identity. Never renumber, delete rows to eliminate, or overwrite 用户决定 / 本人备注 / 父记录.
4. Stay inside the project, ~/.config/jobradar/ and ~/.local/state/jobradar/. Never write to another host's home (~/.codex, ~/.config/opencode, ~/.claude, ~/.deepseek) or read its session logs.
5. No global installs (npm -g, brew, pip outside a project venv), global proxy/VPN, hooks or schedules without the user's explicit approval in this session.
6. Do only what the user asked in this turn. Name your mode before acting; anything outside it needs the user's go-ahead.
7. Application or interview status comes only from the user's own report or a submission receipt, never from table rows or finished materials.
8. Unknown stays unknown. Never invent a URL, deadline, quota or eligibility, and never claim a tool or capability you have not observed in this session.
9. Temporary files go in the run's _support/ or system/validation/<run>/.
10. Finish with the stage's closeout command and quote its receipt. Without a passing receipt, report "not complete" and list what failed.

Mode

Say which mode you are in before the first action:

Ranking, shortlists, Feishu delivery, link/deadline verification for chosen candidates → potential-position. Say so and stop; do not stretch stage 1.

Host tools

Read hosts.md, find your host's row, and list the search and page-read tools you actually have before searching. A tool that errors is a gap to report.

Procedure

ROOT = the user's JobRadar instance directory (holds Profile/, system/config/, system/state/, outputs/; layout in templates/instance/ of the code directory), S = this skill's scripts/ directory (resolve the skill link).

1. python3 "$S/radar.py" --root "$ROOT" doctor. Its watch_due (also in the init receipt) lists recurring programmes whose re-check date has come; tell the user first.
2. Pick a new run name YYYY-MM-DD_主题 (never reuse an example or an existing batch unless the user named it). Write the plan to outputs/<run>/_support/plan.input.json (never to plan.json, which init owns) with scope, instances (mainland / international), max_records and, when the user gave one, priority (ordered list, e.g. ["银行", "运营商", "其余集团"]; collect in that order), then radar.py --root "$ROOT" --run <run> init --plan <that file>. If the user changes scope or priority mid-run, write the new plan to another _support/ file and run amend --plan <new.json> --reason "<user's words>" (search runs only; diagnostic and maintenance plans are fixed at init); it keeps the lease, logs _support/amendments.json and refuses a budget below the records already held. Raise max_records this way rather than splitting a batch; only one stage-1 run holds the lease.
3. Collect in routing order (below). Public ATS: collect-ats --config source.json. Feishu campus collection: collect-feishu --config "$ROOT/system/config/feishu-aggregate.json" --keyword <kw> (optional: a missing or unfilled config means this source is skipped; say so in coverage). Agent-extracted records: write JSON per record-schema.md, then ingest --input observations.json. Every ingest appends likely duplicates to _support/possible-duplicates.json; name each pair in 批次说明 (same role under two IDs is linked in stage 2, never deleted). Application-count limit: when you first record a mainland employer, check its (or its group's) limit on how many roles one applicant may apply for this cohort and write it in quota_note (record-schema.md); closeout refuses a mainland employer without one. Recurring programmes: a fellowship or annual programme whose current window has closed is still recorded (availability: closed) with recurring_note quoting the official evidence that it recurs; ingest lists it as watch_missing until you add it with radar.py watch --add <entry.json> (record-schema.md), and closeout refuses before that. Ingest compares only against batch records.json files; records created only in the workbench are caught later by stage-2 validate/allocate.
4. screen — applies the user's hard_rules from $ROOT/system/config/preferences.json and refuses if the hard-condition document (hard_constraint_source) changed since they were reviewed. In an authorized maintenance task, compare the document (read directly from Profile/, invariant 2) with hard_rules (schema in initial_filter.py), update the rules, then set hard_constraint_sha256. Do not simply replace the hash to bypass the check.
5. export — optional local workbook (岗位初筛工作表.xlsx), only when the user asks for one; Feishu (stage 2) is the main workbench. Refuses on missing/duplicate IDs rather than dropping manual notes. Once a batch has a workbook, closeout requires it to be current.
6. closeout — writes _support/closeout-stage1.json; quote it in your report.

init acquires the same persistent project lease used by stage 2. If another run holds it, init queues (--wait, default 90 s) and reports the holder; to wait longer, follow "Waiting for the shared lock" in hosts.md rather than asking the user to release it. Ingest, screen, collection and export are serialized for their full operation. closeout releases this run's lease even on failure; use init with the same plan to resume. Do not remove another run's lock. Search closeout requires current rules and a current screening receipt, plus a current workbook if one was exported; an empty run cannot masquerade as a completed search.

For requested diagnostics/repairs, use --area validation --run <dated-name> on every command. Plan: mode = diagnostic or maintenance, scope, max_records = 0, required_checks = named checks; maintenance may declare exact allowed_changes paths. Store _support/checks.json keyed by check name, each with passed, evidence (relative file under this run), sha256. closeout verifies these files and rule readiness. A diagnostic pass proves only those checks, not search coverage or all-host readiness.

Search routing

1. Official sources first: employers, universities, institutes, government notices and ATS boards actually linked by the employer.
2. Structured collections second: the configured Feishu campus collection, before general platforms; read feishu-source.md. Use it as an index for gaps in official pages; official pages resolve gaps in the collection.
3. Other platforms and social leads: complementary discovery only. Login history biases results; see collection.md for optional adapters.

Postdocs: search by mentor, not by posting. Go institution → department → faculty → recruitment evidence → funding routes; tiers, IDs and route records are in record-schema.md.

Rendering pages. Most mainland SOE campus portals are JavaScript apps (job.10086.cn, zhaopin.com subdomains, chng / chnenergy / china-cdt / spic portals). A plain page fetch returns an empty shell. Use a rendering browser (hosts.md) for these before falling back to tier 3, and record which portals still failed. Check portals.md for known filters, maintenance windows and workarounds, and add what you learn to it in maintenance mode.

Browser hygiene (the browser is the user's own, shared by every worker):

• One browser worker at a time; each opens its own tab and closes it when done.
• Never click links that open a new tab; read the href and navigate the same tab.
• Keep page scripts short (well under 45 s). A search that reloads the page will hang a script; set the form fields, submit, then read the result in a separate call.
• End every browser task with a tab inventory and close what you opened. Tabs outside the automation group cannot be closed by the agent; tell the user which ones remain.

Time zones. Mainland portals and deadlines use Asia/Shanghai; the host machine may not. Record observation times with +08:00, and check portal maintenance windows in CST.

Subagent workers get the invariants, the mode, the record types from record-schema.md (conflicts is an array, employer-linked ATS goes in official_url, quota_note per employer), the browser rules above, and the company/title list of records already in this run and in earlier batches of the same employers, so they skip known roles.

Priority is not proof. Match employer, actual role/program, location, graduation cohort and dates before comparing. Keep both observations with URLs and timestamps when they disagree; unresolved conflicts stay unknown and cannot eliminate a candidate. Never merge different roles because the company matches, and never silently replace an official requirement with collection metadata. A program or organization is not a confirmed vacancy. Query collections read-only with projected fields and the built-in cache, not repeated full-table reads.

Boundaries

• Mainland and international are coequal instances; report coverage separately.
• Historical social batches (origin=user_social_excerpt) are read-only and never mixed in.
• IDs come only from system/state/skill-ids.json; never assign or edit one by hand.
• Jev experiments only on the user's explicit task: read jev.md; results never enter formal candidate tables.

Output layout

Each batch is outputs/YYYY-MM-DD_主题/ with 批次说明.md, _support/ and, only if exported, 岗位初筛工作表.xlsx. Machine records, plans, evidence, receipts and backups go in _support/. Test runs go in system/validation/, never outputs/. The workbook freezes the header and columns through company and role; the two amber manual columns (用户决定, 我的备注) survive refresh by JOB ID.

Report template

• Mode and run name; tools used for search.
• Records by region and granularity (role / program / organization); retained vs eliminated.
• Unknowns, failures, truncation (has_more), blocked sources, coverage gaps.
• closeout receipt path and passed. Do not claim all sources, providers or institutions were covered.