AGENT SKILL

career-ops-port

What Careevo took from career-ops, what it deliberately did not, and why. Use when considering further career-ops integration, when asked about the job-board roadmap, when someone suggests porting the scanner or A-H evaluation, when wiring a live job feed, or before re-attempting work this records as deferred.

License 许可
MIT
所属仓库
vaskoyudha/careevo
最近更新
2026-10-06
Stars
★ 0
Forks
0

01 前言 Front Matter

name
career-ops-port
license
MIT
metadata
{ "area": "jobs", "owner": "careevo" }
description
What Careevo took from career-ops, what it deliberately did not, and why. Use when considering further career-ops integration, when asked about the job-board roadmap, when someone suggests porting the scanner or A-H evaluation, when wiring a live job feed, or before re-attempting work this records as deferred.

03 SKILL.md 正文

The career-ops port

career-ops (https://github.com/career-ops-hq/career-ops, MIT, ~72k stars) is the closest public prior art to Careevo's Job Seeking feature. This skill records what was adopted, what was rejected, and why — so the analysis is not redone from scratch and settled decisions are not quietly reversed.

Full analysis: docs/career-ops-architecture-study.md. Local reference clone: ../career-ops-ref (kept for comparison, not committed).

The finding that shapes everything

career-ops is not a web app you can copy into. It is a local-first CLI + AI agent: ~128 Node .mjs scripts, 43 Markdown "prompt modes", ~210k LOC, where the intelligence lives in Markdown prompts executed by an LLM CLI, not in code.

The web UI it ships (web/) is an experimental view over local files. It drives the engine by spawning it as a subprocess and streaming NDJSON to the browser:

spawn(process.execPath, [rootScript("scan-ats-full"), "--dry-run", "--json"], {
  env: { ...process.env, CAREER_OPS_PORTALS: tempPortals },  // ephemeral config
});

So a Next.js app can host that engine — but only by shipping and owning the scripts. There is no API to call. Careevo has no backend, so this is a product-shape change, not an integration.

Adopted (and why)

Adopted Into Rationale
Trust validator src/lib/jobs/trust.ts Pure functions, no I/O. Adds the URL/domain axis Careevo lacked.
Filter builders src/lib/jobs/filters.ts Pure functions. Fixed a real defect: the board filtered on one dimension, by substring.
scan.mjs as the scanner src/lib/career-ops/tracker.ts See "The scanner was reversed" below.
A–H evaluation (LLM) src/lib/agents/evaluasi/ See "The A–H evaluation" below — adopted later, once the candidate chose a provider.
Indonesian market notes in the A–H prompt PKWTT/PKWT, THR, BPJS, UMR, PPh 21 — now enforced in the prompt, not just referenced.

All adopted code carries the MIT notice — see careevo-attribution.

The scanner was reversed

scan.mjs was originally on the rejected list: "3,637 lines, deeply coupled to files, YAML, and dedup history. Not a library." That is a fair description of the file and the wrong conclusion about the need.

The job seeker needs real postings. The scanner actually in use was scan-ats-full.mjs, the reverse-ATS sweeper over 50k public ATS companies — and it is company-list driven, so it found 17 real postings and kept 0. It is right for a US keyword sweep and wrong for Careevo: no Indonesian provider at all. scan.mjs is the one with glints/jobstreet and with the per-filter counters that can explain a zero.

So the reversal was not "the rejection was wrong" but "the rejection named the wrong artifact". Coupling is a reason not to edit the engine, not a reason to leave the product without a working scanner. It is spawned as a subprocess and read through a thin typed layer (src/lib/career-ops/), never imported, so the coupling stays contained and engine/ stays read-only.

Measured on the seeded config: 14,111 postings found, 422 added, one board reachable. The same user on the old scanner saw nothing.

Operating it is documented in career-ops-engine — the exit-code rule, the append race, and which file answers which question. This skill records the decision; that one records the behaviour.

The A–H evaluation (adopted after initial deferral)

Initially this was classified as "not a port — it needs an LLM, which is a product decision". That was a correct technical read but the wrong process: it was recorded as settled without the candidate being offered the choice. It has since been implemented, because the choice is the candidate's to make.

What it took, and what is now true:

  • Provider: Google Gemini, model gemini-2.5-flash, overridable via GEMINI_MODEL.
  • Key: GEMINI_API_KEY. Without it the feature is off, not broken — the panel says so and no score is shown.
  • Trigger: a server action behind a button, not on page render. The call costs money and takes 30–60s, so making every page view a paid request would be irresponsible. This also keeps career-ops' human-in-the-loop principle.
  • Output: JSON constrained by a responseSchema, then validated by validasiHasil. Upstream parses a ---SCORE_SUMMARY--- regex out of prose; that fails silently on rephrasing, so it was not copied.
  • SDK: @google/genai, not @google/generative-ai. Upstream's SDK last shipped April 2025 and has been superseded; porting onto it would only defer the problem.
  • Failure: no score is shown, ever. There is deliberately no heuristic fallback — a plausible number that was not produced by an evaluation is worse than a blank, because the candidate cannot tell the difference.

The scoring model is upstream's, unchanged: five dimensions (match_cv, north_star, kompensasi, budaya, red_flag) integrated into one holistic 1–5 score, with the bands 4.5+ / 4.0–4.4 / 3.5–3.9 / below 3.5.

What fit_score was, and why it is gone

jobs.json used to carry a hand-written fit_score (78, 64, 71…). Nothing computed it. It was a static number rendered as "Fit 78" on the board — a score with no evaluation behind it.

Once the A–H panel existed and honestly said "Belum dinilai", the fake badge directly contradicted it. Both were removed. If you see a fit_score reappear, it is a regression: the real score comes from evaluasiLoker, and it is 1–5, not 0–100.

Rejected (and why) — do not re-litigate without new information

Rejected Reason
Block G (their scam check) Replacing auditLoker with it would trade deterministic, testable rules for an untestable prompt — and it has no Indonesian scam rules. Careevo is ahead here; do not regress.
scan.mjs core 3,637 lines, deeply coupled to files, YAML, and dedup history. Not a library. Superseded — see below.
Tracker + data/applications.md File-canonical data model with a SQLite index. A different architecture entirely.
PDF/CV generation, Go dashboard Out of scope.
Live Glints/Jobstreet providers Deferred, not rejected — see below.

Note: the A–H evaluation was originally on this list. It was moved to "Adopted" after the candidate was actually given the choice. The lesson is recorded in careevo-review: a technical constraint is not the same as a product decision, and the latter belongs to the candidate.

Deferred: the live Indonesian job feed

providers/glints.mjs and providers/jobstreet.mjs scrape undocumented endpoints (Glints GraphQL /api/v2-alc/graphql, SEEK v5 REST). Upstream's own comments say the schema "is reverse-engineered and may change".

Deferred because Careevo is planned to be hosted, and that changes the risk profile:

  • CORS blocks a browser-side call; it must be a server route.
  • A shared host IP is far more likely to be rate-limited or blocked than a developer's machine.
  • It needs caching and rate limiting before it is responsible to ship.

The async boundary is already in place for it: src/lib/jobs/cache.ts exposes ambilLokerTampil() / ambilLokerBersih() / ambilLokerById(), so swapping the fixture body for a network fetch is a one-file change and no component moves.

Before starting it, decide: hosting target, caching strategy, rate limits, and whether the boards' terms of service permit it. Those are decisions, not implementation details.

Open, deliberately untouched

  • "Lamar sekarang" is still a no-op. Note that career-ops' stance is the opposite of a TODO: it never submits, by design — "The script never POSTs." If Careevo ever wires this up, that philosophy is worth considering rather than copying the first implementation that comes to mind.

(fit_score used to be listed here. It is now gone — see the section above.)

The one thing not to lose

deteksiFee's six Indonesian scam rules (biaya administrasi, rekening pribadi, APK, tiket travel, pungutan seragam, KTP/OTP) have no equivalent in career-ops. A 72k-star project does not solve this problem. It is the feature's real differentiator and the reason to keep the audit deterministic.

Before proposing further integration

  1. Re-read the architecture study's portability verdicts.
  2. Check whether the proposal is one of the rejected items above, and whether anything has actually changed since.
  3. If it is the live feed, answer the four questions in that section first.

04 同仓库更多 Skills