Spaces:
Sleeping
Sleeping
Commit History
-
Repositioning: world-class personal project (v1.0.0)
Drop all competition/hackathon framing and tighten the project into a standalone, portfolio-grade repo. Behavior of the environment, graders, and API is unchanged; this is reframing + cleanup + consolidation. De-hackathon: - Remove HF Space frontmatter from README; reword "OpenEnv-style" -> "RL-style" and "validator parity"/"validator-friendly" -> "score/log contract" / "numerically stable" across all docs and code docstrings. - Delete the OpenEnv manifest (openenv.yaml) and the dead openenv-core dep. - Rewrite DEPLOYMENT_GUIDE from an HF-submission checklist into a generic Docker/cloud deployment guide (HF Spaces kept as one optional target). - Rename tests/test_validator_parity.py -> tests/test_score_contract.py. - Remove empty stray dirs (Hackathon/, Hackathonenvlearning/, E:Meta/). UI consolidation: - Remove the duplicate Streamlit console (streamlit_app.py) and the streamlit dependency; the React dashboard is now the single served UI. Dependency & tooling: - Single runt…203345c
-
docs: free LLM provider option + positioning whitepaper
- .env.example Style C: reproduce the LLM column at $0 via any OpenAI-compatible endpoint (Groq free tier / local Ollama). docs/BENCHMARK.md gains a reproduce-free section. - docs/WHITEPAPER.md: positioning + methods + real results + honest findings + roadmap. - README docs index links the whitepaper.ad73d06 -
Publish real Azure gpt-4o benchmark results
Fill the LLM column with measured Azure OpenAI gpt-4o numbers (3x3x3 grid): - easy_classification: baseline 1.00 | multiagent 0.80 | llm 0.17 - medium_prioritization: baseline 1.00 | multiagent 1.00 | llm 1.00 - hard_full_management: baseline 0.67 | multiagent 0.09 | llm 0.62 LLM ~3k tokens / ~$0.009 per episode; full sweep ~$0.23. README + docs/BENCHMARK.md (all 27 per-cell rows) + CHANGELOG updated with honest findings.5084129 -
Wave 1: task-aware coordinator, strong hybrid no-key fallback, wider seeds
- CoordinatorAgent is now task-aware (task_id arg + set_task): classifies on easy/medium instead of escalating; keeps risk-first escalation on hard. Multi-agent easy 0->0.8, medium 0->1.0. Footnoted as a finding in BENCHMARK.md. - HybridPolicy falls back to BaselinePolicy heuristics in no-key mode (Planner skips the LLM when no key). No-key hybrid 0/0/0.03 -> 1.0/1.0/0.60. - DEFAULT_SEEDS 3->8 for meaningful CIs (determinism preserved). - +16 tests (multiagent task-awareness + hybrid fallback). Suite: 239 green. Scenario-variant authoring deferred: loader binds one file per task_id; a safe multi-variant loader change is documented for a later pass.662eb9e -
Add reproducible results harness, scenario schema, benchmark docs + OSS scaffolding
- benchmark/results_report.py + scripts/run_benchmark.py: run the matrix and emit results.json/csv/html aggregated by (task,persona,agent) with mean +/- 95% CI across seeds. Stdlib-only, runs key-free for baseline/multiagent. - env/scenario_schema.py: pydantic v2 validator for data/scenarios/*.yaml. - docs/BENCHMARK.md: code-derived methodology (tasks, personas, scoring, repro). - .github/ issue+PR templates, CODEOWNERS; CHANGELOG.md (Keep a Changelog). - +16 tests (results harness + scenario schema). Suite: 224 green.5657821
-
Phase 7: mark core complete in ROADMAP (/v1 move deferred)
cadd51d -
Phase 7: docs consolidation, architecture doc, API metadata + /version
- Enrich FastAPI OpenAPI metadata (description, license) and add a /version endpoint (semver); document the versioning policy. - Add docs/ARCHITECTURE.md (layers, request flow, key components, invariants, design decisions). - README: add a Documentation index (TECHNICAL_REFERENCE/ARCHITECTURE/RUNBOOK/ ROADMAP/CONTRIBUTING/SECURITY), the new health/version endpoints, and a Security & Configuration section. Kept the OpenEnv unversioned paths stable (a breaking /v1 move is deferred to avoid conflicting with validator parity). 204 tests passing, ruff clean.94970d2 -
Phase 6: mark complete in ROADMAP
9bfa208 -
Phase 6: observability ops (probes, lifespan, provisioning, runbook)
- Split health: /health/live (cheap liveness) and /health/ready (DB reachable -> 200/503); keep /health for back-compat. - Add a FastAPI lifespan that logs startup/shutdown (uvicorn maps SIGTERM to a graceful lifespan shutdown). - Add Prometheus scrape config, Grafana datasource provisioning, and an optional docker-compose observability stack (Prometheus + Grafana) under telemetry/. - Add docs/RUNBOOK.md (probes, metrics, alerts, log correlation, incident table, security toggles). - tests/test_observability.py: liveness/readiness, metrics format, /alerts wiring. (Alert rule evaluation was already wired via /alerts.) 203 tests passing, ruff clean.1be2435 -
Phase 5: mark core complete in ROADMAP (additive items deferred)
522a6d6 -
Phase 4: mark complete in ROADMAP (a11y deferred)
c1ffafc
-
Phase 3: mark complete in ROADMAP (release pipeline deferred)
db9c575 -
Phase 2: mark complete in ROADMAP
bddcfb8 -
Phase 1: mark complete in ROADMAP
93fa6ff -
Phase 0: sync docs with closed gaps
Update README 'Important Constraints' and TECHNICAL_REFERENCE 'Operational Notes & Constraints' to reflect wired persistence/learning/telemetry, the opt-in approval gate, and the persistent response cache. Mark Phase 0 done in ROADMAP and flag uncommitted source files for Phase 1.ad65925 -
Phase 0: opt-in HITL approval, persistent LLM cache, fix failing test
- LLMAgent.require_approval (default off) so the raw agent returns its decided action; API/product path can enable via arg or REQUIRE_APPROVAL env. - Stop clearing the response cache every call; cap size; skip cache when approval is required so it cannot bypass the gate. - Replace bare except in confidence retry with logged handling. - Add tests for both approval modes and cache reuse; isolate cache per test. Adds docs/ROADMAP.md (world-class roadmap).9faa270