How bhagyesh-074/oasis works
OASIS is a framework-agnostic supervisory layer for agent teams. Its documented workflow sizes a team from task complexity, reduces it to fit a run budget, executes it through a framework adapter, enforces budgets on model calls, and supports agent replacement; runtime model-call records and automated decisions provide observability. The README also identifies retrieval/context management, runtime verification, and evaluation as product capabilities. Several of these modules were not sampled, so their presence is shown without inferred call paths; edges reflect only shown code references or README statements.
Main entry points: base.py Adapter contract · litellm_hooks.py LLM gateway · admission.py Call admission · runner.py Evaluation runner
Components 24
Team planning
- Task complexity
src/oasis/tce/heuristic.py - Team planner
src/oasis/tce/llm_planner.py - Team reducer
src/oasis/rbe/reducer.py
Execution and control
- Budget model
src/oasis/rbe/budget.py - Call admission
src/oasis/rbe/admission.py - LLM gateway
src/oasis/gateway/litellm_hooks.py - Adapter contract
src/oasis/adapters/base.py - LangGraph adapter
src/oasis/adapters/langgraph_adapter.py - AutoGen adapter
src/oasis/adapters/autogen_adapter.py - CrewAI adapter stub
src/oasis/adapters/crewai_stub.py - Agent replacement
src/oasis/replacement/engine.py - Scripted LLM
src/oasis/gateway/fake_llm.py - Prompt hashing
src/oasis/gateway/_hash.py - Request log
src/oasis/gateway/jsonl_sink.py
Context and verification
- Context retrieval
src/oasis/cm/retrieval.py - Runtime verification
src/oasis/rtpm/l1_embed.py - Decision logging
src/oasis/rbe/decision_log_client.py - Context store
src/oasis/cm/store.py
Evaluation
- Evaluation runner
src/oasis/eval/runner.py - Oracle sweep
src/oasis/eval/oracle_sweep.py - Evaluation statistics
src/oasis/eval/stats.py
API and records
- OASIS API
src/oasis/api/app.py - Application database
src/oasis/db/models.py
Other components
- Run initiator
Connections 20
- Run initiator → OASIS API: submits run
README.md - OASIS API → Evaluation runner: serves runs
- Task complexity → Team planner: guides sizing
README.md - Team planner → Team reducer: proposes team
README.md - Budget model → Team reducer: constrains team
src/oasis/rbe/reducer.py - Team reducer → Decision logging: logs decisions
src/oasis/rbe/reducer.py - LLM gateway → Call admission: checks calls
src/oasis/gateway/litellm_hooks.py - Call admission → Budget model: checks limits
src/oasis/rbe/admission.py - Call admission → Decision logging: logs decisions
src/oasis/rbe/admission.py - Adapter contract → LangGraph adapter: implemented by
src/oasis/adapters/langgraph_adapter.py - Adapter contract → AutoGen adapter: implemented by
src/oasis/adapters/autogen_adapter.py - Adapter contract → CrewAI adapter stub: stub adapter
src/oasis/adapters/crewai_stub.py - Adapter contract → Agent replacement: supports replacement
README.md - LLM gateway → Request log: writes records
src/oasis/gateway/litellm_hooks.py - LLM gateway → Scripted LLM: supports replay
README.md - Scripted LLM → Prompt hashing: hashes requests
src/oasis/gateway/fake_llm.py - Context retrieval → Context store: retrieves context
- Evaluation runner → Oracle sweep: sweeps outcomes
- Evaluation runner → Evaluation statistics: summarizes results
- OASIS API → Application database: uses records