Nuclear clamp: every reward source in the codebase now returns (0.05, 0.95)

Clamp all remaining score leak paths: /state, step_rewards, demo SSE

Bulletproof _safe_score for all bad inputs (None, NaN, strings, bool)

Clamp every grader return value strictly inside (0, 1)

Defensive score clamping at all emission points

Widen score epsilon to 0.01 so :.3f formatting stays in (0, 1)

Fix reward.value to always be task_score in (0,1) exclusive

Fix task scores to be strictly in (0, 1) exclusive per OpenEnv spec

fix

fix

fix

fix

Update project origin description in README

Clean up README.md by removing metadata

fix

Remove origin details from README

Refactor README.md by removing metadata and updating content

fix

fix

fix

fix

fix

fix

fix

fix

fix

fix

Reset SQL on new attempt so attempts don't concatenate in the same box

Fix UnboundLocalError: remove duplicate local import of REPAIR_ACTION_BY_NAME

Surface GEPA optimization: prompt history, live banner, smart retry

Fix IndexError: skip empty-choices chunks from HF Router streaming

Add LLM diagnostics: /api/test-llm endpoint + startup/error logging

Surface real error messages and fix LLM exception formatting

Fix chat SSE events to match frontend protocol (result+done instead of success)

Fix chat query failures and benchmark ID mismatches

fix: GEPA current_generation, task_id mapping, Connect DB button, remove difficulty from header

fix: default LLM to HF Router + Qwen2.5-72B, no custom Space variables needed

fix: remove duplicate YAML frontmatter in README

fix: light mode contrast, collapse queries on demo completion

feat: demo mode with reward chart, github diff, single-difficulty rounds, no loop

feat: add Demo button with scripted autoplay showcase

fix: align rl-state API shape with frontend, add error boundary

Initial submission: SQL Agent OpenEnv for Meta+HF hackathon

initial commit