feat: complete SafeAct-Env — RL environment for safe agent training#1

Merged

CodeNinjaSarthak merged 61 commits intomainfrom

Apr 6, 2026

Commits on Mar 26, 2026

Commits on Mar 27, 2026

add medical triage task with drug interaction safety protocols
CodeNinjaSarthak
committed
add cloud infrastructure task with instance safety and security posture scoring
CodeNinjaSarthak
committed
add rule-based agent and extend environment to support all five tasks
CodeNinjaSarthak
committed
improve graders with prerequisite scoring and partial credit curves
CodeNinjaSarthak
committed
support both openai and azure backends via OPENAI_BACKEND env flag
CodeNinjaSarthak
committed
update readme with all five tasks and dual baseline results
CodeNinjaSarthak
committed
add outcome-shaped step rewards based on state transitions
CodeNinjaSarthak
committed
wire outcome-shaped rewards into environment step loop
CodeNinjaSarthak
committed
fix: task_complete=False on irreversible mistake, add done guard
CodeNinjaSarthak
committed
feat: wire /baseline endpoint to subprocess runner with --task --json flags
CodeNinjaSarthak
committed
fix: replace cliff grader with granular completion score in hard task
CodeNinjaSarthak
committed
feat: add trap file system_cache_cleanup.tmp to easy task
CodeNinjaSarthak
committed
feat: add adversarial actions optimize_table_storage and refresh_stale_cache to medium task
CodeNinjaSarthak
committed
docs: complete README with baseline scores, API reference, reward design
CodeNinjaSarthak
committed
fix: return empty EpisodeState instead of 500 on /state before reset
CodeNinjaSarthak
committed
fix: easy task max_steps mismatch — environment.py had 25, app.py had 20
CodeNinjaSarthak
committed
feat: add session-aware /reset, /step, /state endpoints and /demo route
CodeNinjaSarthak
committed
refactor: modernize type hints (Dict/List → dict/list) and reformat
CodeNinjaSarthak
committed
chore: sync pyproject.toml formatting and uv.lock
CodeNinjaSarthak
committed
feat: add inference.py with API_BASE_URL/MODEL_NAME/HF_TOKEN support
CodeNinjaSarthak
committed
fix:DQ risk fixes and bug fixes
CodeNinjaSarthak
committed
fix:harden hard task difficulty - hard task max_steps 25→18 (was too generous) - Add force_cutover trap action (irreversible, sounds like switch_traffic) - Add preparation_discipline (0.10) — check…
CodeNinjaSarthak
committed
fix: fix medium do-nothing agent scoring 0.70
CodeNinjaSarthak
committed
feat: seeded randomization for easy/medium/hard tasks - Add get_initial_state(seed=None) to all 5 tasks - Easy: randomized log/temp file selection, sizes ±25% - Medium: 2-4 stale tables from pool, …
CodeNinjaSarthak
committed
refactor:remove dead code, extract shared utilities - Remove _STUB_ACTIONS, _STUB_RISK_MAP from environment.py - Remove _STEP_REWARDS dict + get_step_reward() from all 5 tasks - Extract shared/llm_…
CodeNinjaSarthak
committed
fix: session TTL cleanup, difficulty tightening
CodeNinjaSarthak
committed
docs: update baseline scores across README, openenv.yaml, demo UI
CodeNinjaSarthak
committed

Commits on Mar 29, 2026

Commits on Apr 3, 2026

Commits on Apr 6, 2026

deploy: prepare for HuggingFace Spaces deployment
CodeNinjaSarthak
committed