Spaces:
Sleeping
Sleeping
Commit History
-
feat: add cold-start notice banner to dashboard
Yellow dismissable banner below tab bar explains the scale-to-zero delay. Auto-hides after 60s or on clicking X.d62e2b1 -
feat: redirect / to /dashboard
0cc34c9 -
fix: allow hf_blog_post.md through Docker build context
1c3b084 -
refactor(dashboard): streamline Shepherd tour for production UX
3sha- Replace long multi-step tour with four concise workspace steps - Add Skip on every step, neutral copy, and improved footer layout - Tighten tour styling (skip as text control, canClickTarget false) - Bump onboarding storage key to v4 after upstream merge Made-with: Cursoraa65daf -
Please work
b595345 -
docs: complete README and blog post rewrite for hackathon submission
README: - Sharper problem framing with the 3AM scenario - Explicit results table with before/after on all 4 tasks - Added explanation of WHY reward curve plateaus (expected behavior) - Added explanation of WHY loss rises (GRPO healthy training signal) - Added explicit scoring criteria alignment table - Full task descriptions with correct response chains - Architecture diagram with both agents' full action lists - Design principles explaining anti-shortcut measures hf_blog_post.md: - Complete rewrite with full technical depth - GRPO mechanism explained (group ranking, advantage, gradient) - Dense reward table with design rationale per row - Results framing: addresses the per-step vs episode-level distinction - Training setup code included - Theme alignment table - Anti-shortcut design explained9778d24 -
fix: Learning tab curriculum refresh + TECHNICAL_ANALYSIS.md
- grade event now triggers refreshCurriculum() immediately - battle_end event also triggers refreshCurriculum() immediately - Both use server-sent curriculum_level instead of hardcoded 1 - Added TECHNICAL_ANALYSIS.md: honest full pipeline docs + theme alignmentd80f206 -
Added example env
e477e10 -
please work this time no
0384157 -
updated notebook to train the model and linked huggingface space api to app.py
1e9a864 -
fix: Learning tab now records every episode + full DASHBOARD_GUIDE.md
- demo_stream and battle_stream now call _curriculum.record_score() at the end of every episode so Learning tab actually populates - curriculum_level and curriculum_episodes now sent in grade/battle_end events - Added DASHBOARD_GUIDE.md — plain English explanation of every element40cbdc9 -
fix: robust AI response parser + dotenv auto-load
- _parse_ai_action now handles model's actual output format (action/INVESTIGATE, action/QUERY_LOGS, details.hosts, etc.) - Maps 20+ synonym action names to valid env action_type strings - Falls back to keyword scan of raw text if JSON parse fails - Never returns None anymore — always produces a valid SecOpsAction - Added python-dotenv auto-load so server reads .env automatically - Added python-dotenv to requirements.txt2038bfa -
fix: randomise env seed each episode + live AI auth warning + /debug/ai endpoint
- Each episode now gets a unique random seed -> different metric values, alert ordering and log lines every run (no more identical 0.990/3.20 every time) - MultiAgentSecOpsEnv also randomised per episode (red agent too) - Print clear WARNING at startup if HF_TOKEN is not set (was silently falling back) - AI call now logs the actual exception so you can see 401/timeout in console - Added /debug/ai endpoint to test the live endpoint in one browser hit - Added _randomise_env_seed() helper that jitters cpu/mem/latency at starta46ff49 -
feat: wire live Qwen2.5-7B endpoint + dashboard AI indicators
- Hardcoded live HF endpoint as default (override via TRAINED_MODEL_ENDPOINT env var) - Added /ai/status API endpoint for dashboard status check - Dashboard: AI status badge in header (pulsing green dot + model name) - Action cards show raw AI JSON output in teal box under each action - Thinking indicators now say 'Qwen2.5-7B thinking...' when live AI is active - Battle feed title updates to show live AI vs Attacker - Episode-end modal shows emoji icons and agent mode label - Added escHtml() helper to safely display AI outputb87cf9d -
feat: live AI inference engine + winning README + HF blog post
- app.py: async _query_ai_model() calls Hugging Face Inference Endpoint when TRAINED_MODEL_ENDPOINT env var is set; gracefully falls back to heuristic playbooks when not configured - demo_stream and battle_stream now support live AI mode - Added _obs_to_text() and _parse_ai_action() helpers - requirements.txt: added httpx for async HTTP - README.md: complete rewrite for hackathon judges (hook, architecture, training results table, all required links) - hf_blog_post.md: mini blog post draft for HF publication41e5dfa
-
Refine dashboard UI to a professional training-platform style.
3shaThis updates typography, spacing, labels, and color system to a cleaner purple/teal visual language while preserving all existing backend behavior. Made-with: Cursorc6034cd -
tMerge branch 'main' of https://github.com/rihandoshi/incident-ai
Okay idk whats happening0c64da7 -
new agent bugs
e3db7a3 -
Jupiter
86feffc -
Looks like we working on round 2 now
afdcfb9
-
fix: clamp grader scores to (0,1) open interval for validator
0aa986d -
fix: accept empty POST body on /reset for validator compatibility
65abbb7
-
fix: allow README.md through .dockerignore
8aa8bca -
v0.2.0 final: Dockerfile fix, version sync, README rewrite, web UI 4th task
ee5b683 -
Final: openenv validate pass, 4 tasks, checklist compliant inference.py
27410ce -
Final (Hopefully???): checklist-compliant inference.py, session server, 4 tasks
8209638 -
More changes, fixed few more bugs
21aac0b
-
another commit, fixed a few bugs yes
fd70b08 -
First commit - already did lots whoops
ce86ff7