openspec/changes/fix-traced-bugs/tasks.md

1. Layer 1 — Privacy: Complaint ID PII Protection

• 1.1 Add _complaint_id_deny_recognizer() function in src/privacy/redactor.py that returns a PatternRecognizer with entity type COMPLAINT_REF_ID, regex matching REF-, TRN-, TXN-, OD-, BK-, CLM-, CASE-, CMP-, LN-, POL-, CR- prefixed IDs (score 0.99), and a comment explaining intentional exclusion from _ENTITY_TYPES
• 1.2 Register the _complaint_id_deny_recognizer() in PIIRedactor.__init__ alongside the existing Indian recognizers
• 1.3 Verify manually: PIIRedactor().redact("REF-20260501-001") returns text with REF-20260501-001 unchanged; PIIRedactor().redact("account 9876543210") still returns <US_BANK_NUMBER>

2. Layer 1 — Infrastructure: Tesseract Pre-flight

• 2.1 Add a _check_tesseract() function in start.py using shutil.which("tesseract") that prints a formatted error with macOS/Ubuntu/Windows install commands and calls sys.exit(1) if the binary is absent
• 2.2 Call _check_tesseract() in start.py before the uvicorn server launch block (after argument parsing, before training or server startup)
• 2.3 Verify manually: rename/hide the tesseract binary temporarily and confirm python start.py exits with code 1 and prints install instructions

3. Layer 3 — Prompt: Rule 4/5 Tightening

• 3.1 In src/agent/prompts.py Rule 4, append after the tab-navigation instruction: "Do not append a yes/no follow-up question ('Does this look correct?', 'Shall I proceed?', etc.). The numbered summary above is the confirmation request. Stop completely and wait for [USER CONFIRMED]:."
• 3.2 In Rule 5, replace "The most recent user message begins with" with "You have received a user message that begins with" to eliminate turn-ordering ambiguity
• 3.3 Verify by running Trace 07 scenario: agent should present numbered summary and stop without asking an extra question

4. Layer 3 — Prompt: HITL Structured JSON Block

• 4.1 In src/agent/prompts.py Rule 4, add an instruction immediately after the numbered list template: the agent SHALL append <!--ENTITIES:{"provider":..., "incident_date":..., "amount":..., "reference_id":..., "prior_contact":..., "desired_resolution":...}--> with the same values as the numbered list; omit keys for unknown fields; use string values for all fields including prior_contact ("Yes"/"No")

5. Layer 4 — UI: Structured Entity Parsing

• 5.1 Add _parse_hitl_entities_json(reply: str) -> dict function in ui/app.py that extracts the <!--ENTITIES:{...}--> comment, calls json.loads, and returns the dict (returns {} on any parse error or absent block)
• 5.2 Replace the call to _parse_hitl_entities(reply) in _handle_chat with _parse_hitl_entities_json(reply)
• 5.3 Remove the _parse_hitl_entities function and _HITL_PATTERNS dict (now unused)

6. Layer 4 — UI: prior_contact Guard

• 6.1 In _handle_confirm in ui/app.py, change the prior_contact entry from "prior_contact": prior == "Yes" to "prior_contact": True only when prior == "Yes", and omit the key entirely when prior is None, empty string, or "No" — align with the filter condition on line 596

7. Layer 4 — UI: _extract_draft Relaxation

• 7.1 In _extract_draft in ui/app.py, add a third extraction path between the --- primary and the Subject: fallback: re.search(r'(Subject:.+)', reply, re.DOTALL | re.IGNORECASE) that returns from Subject: to end of string when no closing --- is present
• 7.2 Fix the existing "Subject:" in reply fallback to return reply[reply.index("Subject:"):] instead of the full reply string

8. Layer 2 — ML: NER ORG List Extension

• 8.1 In src/ner/train.py _ENTITY_VALUES["ORG"] list, add: "Indian Bank", "Niva Bupa", "Care Health Insurance", "Bajaj Allianz", "Agoda", "OYO Rooms", "Booking.com India", "Razorpay", "BharatPe", "CRED"
• 8.2 Retrain EvidenceNER checkpoint: python -m src.ner.train --output_dir models/evidence_ner
• 8.3 Verify: extract_entities("Indian Bank charged 1000 rupee") returns an entity with label="ORG" and text matching "Indian Bank"

9. Layer 2 — ML: Domain Classifier Rebalancing

• 9.1 In src/classifier/train.py _SUPPLEMENT_TEMPLATES["telecom"], add 10 new templates covering SIM activation, prepaid recharge, porting, deactivation scenarios; add 5 Agoda/OYO/Booking.com templates to ecommerce
• 9.2 Change the supplement_per_class default in _build_supplement() signature from 2_000 to 5_000
• 9.3 Change the --supplement_per_class CLI argument default in train.py from 2_000 to 5_000
• 9.4 Retrain DomainClassifier checkpoint on Kaggle: zip src/classifier/train.py + CFPB CSV, run python -m src.classifier.train --cfpb_csv data/raw/complaints.csv --output_dir models/domain_classifier, download and replace models/domain_classifier/
• 9.5 Verify: classify("I bought a new sim card from BSNL for rs 450 but sim card is not yet activated") returns domain="telecom" with low_confidence=False