Features:
- Token streaming UI with word-by-word display
- Settings sidebar: temperature, top_p, max_tokens, validation toggles
- HR task quick actions: Resume, Job Description, ATS Keywords, Job-Resume Match, Recruiting Strategy
- Blocked response handling with policy messages
- Validation status display (intent, valid flag, issues)
- Token usage tracking (input/output counts)
- Inline temperature override via temp=X.X in messages

Configuration:
- Environment-based config (RUNPOD_BASE, ENDPOINT_ID, API_KEY, CLIENT_TIMEOUT_S)
- Domain defaults: 180 max_tokens, temp=0.3, validation enabled, low-trust intents blocked
- Railway deployment ready with Procfile, railway.json, requirements.txt

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Use 'python -m chainlit' instead of 'chainlit' command
- Add nixpacks.toml with proper Python 3.12 setup
- Update Procfile and railway.json start commands
- Install packages with pip --user flag

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Changed cl.Action parameter from 'value' to 'payload'
- Updated all action callbacks to use action.payload
- Fixes: 1 validation error for Action - Field required

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Changed payload to dict: payload={'prompt': ...}
- Updated callbacks to extract prompt: action.payload['prompt']
- Fixes Pydantic validation error: Input should be a valid dictionary

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Removed validation display (intent, valid flag, issues)
- Removed usage tracking display (input/output token counts)
- Cleaner UI with just the streamed response

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Implement ChatML formatting with system prompts for 5 HR domains
- Add validators for all domains (resume, JD, match, recruiting, ATS)
- Add micro-retry logic with clarifiers for failed validations
- Update to vLLM endpoint with ChatML support
- Add domain detection from user prompts
- Enhance UI with validation status indicators
- Update settings: remove block_low_trust, add micro_retry toggle
- Add 1-shot examples for resume and JD domains
- Test results: 4/5 domains passing validation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Lower word threshold from 80 to 60 words (more realistic)
- Add rationale validator to ensure score explanation is included
- Update system prompt to explicitly request score reasoning
- Update clarifier to emphasize explanation requirement
- Test result: Now passes validation with clear score rationale

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Update extract_text to support both raw vLLM and handler endpoint formats
- Update extract_usage to support both response shapes
- Handler format: output as dict with choices[0].text
- Raw vLLM format: output as list with choices[0].tokens
- Tested with both formats - all passing

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Implement classify_with_llm() with temp=0, 60 tokens for fast classification
- Add router labels: resume, jd, match, recruiting, ats, small_talk, general_qna
- Add confidence-based routing: high (≥60%) → LLM, medium (≥40%) → hybrid, low → heuristic
- Add LLM Router toggle in settings (default: enabled)
- Skip HR validation for small_talk and general_qna domains
- Show detected domain with confidence in response footer
- Update build_chatml to handle small_talk and general_qna
- Test results: 95% accuracy on all HR domains, correctly identifies small talk

Fixes domain detection issues - no more false positives on casual conversation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Update welcome message with EvalMatch Career Copilot branding
- Emphasize natural language interaction with LLM router
- Add clear capability descriptions with emojis
- Professional, user-friendly messaging

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

- Add explicit routing rules for recruiting_strategy domain
- Include 'strategy to hire/recruit/get/find/attract/source developers' patterns
- Add 'hiring pipeline / talent acquisition / onboarding plan' patterns
- Improve heuristic fallback with 'find developers', 'get developers', 'strategy to'
- Update small_talk and general_qna prompts to be concise (1-3 sentences)
- Remove seeding for non-HR domains
- Improve JSON parsing with better error handling

Test results: 95% confidence on all recruiting queries
- 'Help me create a strategy to get relevant new age AI developers' → recruiting_strategy ✅
- 'How can I develop a pipeline for full stack developers' → recruiting_strategy ✅
- 'Strategy to hire senior DevOps engineers' → recruiting_strategy ✅

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Critical Fixes:
- Fix extract_text to join all tokens instead of returning only first token
  This was truncating all responses to the first token only
- Add missing edge cases: string output, root-level text, handler format
- Fix streaming to preserve whitespace and formatting using re.split(r'(\s+)')
  Previous implementation removed line breaks and bullet formatting

Minor Improvements:
- Fix 'hi' detection to work without trailing space (exact match at boundaries)
- Make ATS keyword detection more specific (avoid generic 'keyword' term)
- Remove ats_keywords from SEEDED (comma-separated output doesn't need '1. ' prefix)

Test Results:
✅ extract_text correctly joins token arrays
✅ extract_text handles all edge cases (string, dict, root-level)
✅ Streaming preserves newlines and formatting
✅ Heuristics more accurate with edge case fixes

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

…mples

Final improvements to variant hint pack system:
- Replace print with proper logging.Logger with basicConfig
- Add max_variants support to template system (enforced in select_variant)
- Add sampling override example (repetition_penalty for ATS keywords)
- Remove redundant DEBUG_VARIANTS boolean guard (rely on logger level)
- Ensure console handler always exists for dev environments

All 33 unit tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Improvements to logging and sampling configuration:
- Replace basicConfig with logger-specific StreamHandler to avoid clobbering embedded host logging
- Extract merge_sampling() helper function for cleaner, testable code
- Add 4 comprehensive unit tests for merge_sampling (min_tokens, direct overrides, multiple overrides, no mutation)

All 37 unit tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

…ssion

Defensive improvements to logging and sampling:
- Set logger.propagate = False to prevent double emission if parent handlers exist
- Add temperature/top_p clamping to [0, 1] range in merge_sampling
- Add stop coercion to list for decoder compatibility
- Add 3 comprehensive tests for defensive behaviors (temperature, top_p, stop)

All 40 unit tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Improvements for reliability and browser compatibility:
- Add CLIENT_RETRY_ATTEMPTS and CLIENT_RETRY_BACKOFF_S environment variables
- Implement exponential backoff retry for timeout exceptions in call_endpoint()
- Add timeout-specific error messaging with guidance for users
- Fix upload accept config: replace "*/*" with specific MIME categories
- Sanitize accept patterns in useUpload to prevent react-dropzone warnings

Timeout retry behavior:
- Only retries on httpx.TimeoutException (not other errors)
- Exponential backoff: 0.75s, 1.5s, 3s, 6s...
- Default: 1 retry attempt (configurable via CLIENT_RETRY_ATTEMPTS)
- Logs debug messages when backing off between retries

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Major improvements to recruiting strategy and response parsing:

Streaming-aware variant enhancements:
- Add comprehensive Boolean sourcing patterns (JMS, ActiveMQ, RabbitMQ, Kinesis, Pulsar, Reactor, Netty, backpressure, CDC/Debezium, etc.)
- Include 4-6 specific channels with cadence (LinkedIn Recruiter, GitHub code search, Stack Overflow tags, Kafka/Reactive meetups, Java user groups, conference speakers)
- Add personalized outreach guidance mapping event-driven/JMS/Reactive signals to Kafka/Pulsar
- Include screening prompts for backpressure/idempotent consumers and batch→near-real-time migration
- Add metrics tracking (response rate %, pass-through %, time-to-hire)
- Update clarifier to match expanded hint structure

Extract_text improvements:
- Add support for chat-style responses (message.content, delta.content)
- Add support for streaming responses (delta content)
- Enhanced fallback chain: text → output_text at multiple levels
- Better null safety with isinstance checks
- Comprehensive docstring documenting all supported formats

These changes improve recruiting strategy quality for streaming/event-driven roles and ensure compatibility with both completions and chat API formats.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>