Jobseek Label Daily

Overview

This routine samples public job postings from the last 24 hours, labels them with Codex project custom agents, validates the outputs, merges accepted records, and uploads accepted gold data to viktoroo/jobseek-postings-labelled.

The Python labeller code is deterministic orchestration only: database reads, Jinja task rendering, JSON Schema/custom validation, merge, QA, and HuggingFace upload. Do not add provider SDK calls or call OpenAI or Anthropic endpoints from apps/crawler/src/labeller. LLM judgment happens through the already-running, subscription-backed Codex session and project custom agents configured in the active harness; the repo mirrors durable subagent contracts under .agents/labeller/.

Agents

Use these project custom agents:

• jobseek-labeller-normalizer: GPT-5.6 Luna at low reasoning; raw HTML task input -> normalized.html.
• jobseek-labeller-splitter: GPT-5.6 Luna at medium reasoning; split_sections task input -> split-out.json.
• jobseek-labeller-extractor: GPT-5.6 Terra at high reasoning; extract_all task input -> extract-all-out.json.

The committed .codex/agents/*.toml files pin these settings. Do not let the labeller agents inherit the Sol/high orchestrator default: normalization and splitting are clear, repeatable tasks, while combined extraction needs more semantic judgment without requiring the flagship model.

Each agent invocation message must be exactly:

INPUT: <rendered-input-path>
OUTPUT: <output-path>

The Jinja task prompts in apps/crawler/src/labeller/prompts/tasks/ own task-specific rules and schema phrasing. Do not duplicate or rewrite those prompts in the orchestrator.

Workflow

Run commands from apps/crawler with uv run labeller .... Default RUN_DATE is today; default accepted-record target is 10 unless the user provides overrides. Start with SAMPLE_SIZE=10; if QA rejections leave fewer than 10 accepted local records for the date, sample additional candidates and continue until the accepted target is met or the sampling pool is exhausted.

1. Sample postings:

uv run labeller sample --date "$RUN_DATE" --count "$SAMPLE_SIZE" \
  --out "data/postings-labelled/_runs/$RUN_DATE/sample.json"

2. For each sampled posting ID, set RUN_DIR=data/postings-labelled/_runs/$RUN_DATE/$POSTING_ID.

3. Prepare raw input:

uv run labeller prepare-pre-llm "$POSTING_ID" --date "$RUN_DATE"

Skip the posting if the command exits non-zero.

4. Render and run normalization:

uv run labeller render-task --task normalize_html \
  --input "$RUN_DIR/raw_input.json" \
  --out "$RUN_DIR/normalize-in.md" \
  --output-path "$RUN_DIR/normalized.html"

Invoke jobseek-labeller-normalizer with the two-line INPUT/OUTPUT message. Then check that normalized.html exists, is non-empty, and starts with <.

5. Finalize the deterministic input:

uv run labeller prepare-post-llm "$POSTING_ID" --date "$RUN_DATE"

Skip the posting if coverage validation fails.

6. Render, run, and validate section splitting:

uv run labeller render-task --task split_sections \
  --input "$RUN_DIR/input.json" \
  --out "$RUN_DIR/split-in.md" \
  --output-path "$RUN_DIR/split-out.json"

Invoke jobseek-labeller-splitter, then validate:

uv run labeller validate --kind sections \
  --file "$RUN_DIR/split-out.json" \
  --context "$RUN_DIR/input.json"

On validation failure, re-render with --previous-error "<validator output>" and retry the same agent up to two times.

7. Render, run, and validate combined extraction:

uv run labeller render-task --task extract_all \
  --input "$RUN_DIR/input.json" \
  --sections "$RUN_DIR/split-out.json" \
  --out "$RUN_DIR/extract-all-in.md" \
  --output-path "$RUN_DIR/extract-all-out.json"

Invoke jobseek-labeller-extractor, then validate:

uv run labeller validate --kind extract_all \
  --file "$RUN_DIR/extract-all-out.json" \
  --context "$RUN_DIR/input.json"

Use the same two-retry protocol on validation failure.

8. Merge and validate the posting:

uv run labeller merge --posting "$POSTING_ID" --date "$RUN_DATE" \
  --verdict accepted \
  --out "data/postings-labelled/postings/$RUN_DATE/$POSTING_ID.json"

uv run labeller validate --kind posting \
  --file "data/postings-labelled/postings/$RUN_DATE/$POSTING_ID.json"

uv run labeller validate --kind qa \
  --file "data/postings-labelled/postings/$RUN_DATE/$POSTING_ID.json" \
  --report "$RUN_DIR/qa.json"

If merge or final validation cannot be repaired, use labeller merge --verdict rejected --rationale "<reason>" rather than hand-crafting a record. If QA fails after a valid merge, update only labelling_meta.qa_verdict and labelling_meta.qa_rationale from the QA report.

9. Upload accepted postings only when the user asked to run the routine, not during surface verification:

uv run labeller upload --date "$RUN_DATE"

Use --dry-run for upload verification.

10. End the final response with exactly one single-line runner outcome:

JOBSEEK_ROUTINE_RESULT={"status":"completed","phase":"upload-verification","primary_error":null}

For a fail-closed run, use status: "failed", name the first failing phase, and put a concise redacted description of the first causal error in primary_error. A missing sample artifact or remote date partition is a downstream consequence when an earlier database/model/validation command already failed, so it must not replace that causal error. Emit the marker even when the runner was invoked manually; the Hetzner runner consumes it from the Codex JSONL trace.

Legacy Fallback

The older per-section extractors (extract_team, extract_role, extract_requirements, extract_preferred, extract_benefits, extract_globals) remain available for compatibility. Prefer extract_all unless the user explicitly asks for granular retries or debugging of a specific section kind.

Focused Verification

For changes to this orchestration surface, run:

git status --short
uv run --with 'pyyaml>=6' python /Users/Viktor/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/jobseek-label-daily
python3 - <<'PY'
import pathlib
expected = {
    ".agents/labeller/normalizer.md",
    ".agents/labeller/splitter.md",
    ".agents/labeller/extractor.md",
}
missing = [path for path in expected if not pathlib.Path(path).exists()]
assert not missing, missing
print("labeller agent contracts present")
PY
legacy_terms="$(printf '%s' 'son' 'net|op' 'us|Agent[(]|[.]Codex|Claude Code ' 'session|Anthropic ' 'API')"
rg -n "$legacy_terms" .codex .agents .claude || true
cd apps/crawler && uv run labeller --help

Do not run expensive uploads as verification.