[RockChinQ]

Requested on LANG-327, based on the customer's lead-scoring rubric (A/B/C keyword tiers with points, urgency/decision-maker bonuses, deductions, compliance-risk words, thresholds ≥60 / 30–59 / <30, and the lead fields to capture). Decision from Chin Rock: medium intent is asked once too.

Changes (LeadCapture 0.3.0)

• Scoring: the judge returns {score 0–100, signals[], risk, risk_reason, reason}. Levels come from the new config items high_score (default 60) and ask_score (default 30). Medium and high intent get the contact request once; low does not. A 0.1/0.2-style strong_intent answer still maps to a score. An answer without a score is recorded as an error and never breaks the reply.
• Compliance risk (risk_prompt, default lists scalping, fake orders, cracking, scraping WeChat/banks, VPNs…): the lead becomes risk, with no contact request and no sales email. It is kept for review on the page.
• Lead fields (lead_fields, default: the 12 fields from the rubric): extracted once from the whole conversation when contact details arrive. They appear at the top of the email (missing ones show as 未提及) and on the page. The subject now reads 【购买意向·高意向 75分】….
• Page: score/level column, risk total and filter, a lead-fields table in the detail view, and per-judgment score/signals/risk.
• The intent_prompt default now describes a scoring rubric; READMEs (en/zh) are updated.

Verification

• plugins/misc/LeadCapture: 31 passed, 8 of them new:

  • tiers high/medium/low decide the single ask;
  • configurable thresholds;
  • risk never asks or emails;
  • email shows score, signals and extracted fields (unknown keys dropped, missing ones shown as 未提及);
  • an unparseable judgment is recorded as an error;
  • score clamping.

• ruff check, check_catalog.py, test_repository.py OK.

• Real model check: brightdata's claude-haiku-5-5 via LangBot Models, with the exact prompts the plugin builds and the customer's rubric.

  Sample	Score	Result
  high	100	7 signals
  trial-only	20	low, matches the rubric's +10 per B signal
  low (free/homework)	0
  ticket scalping	45	risk=true
  Traditional Chinese	100

  Field extraction filled 10 of 12 fields correctly and left the 2 unmentioned ones empty.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!