| 87cfb0b93c123fc8… | The tool decorator supports runtime behavior flags including requires_context, is_base_tool, subagent_prohibited, approval_required, and bypass_size_guard so that tools can declare execution constraints and capabilities. | active | contextual | yes | |
| 1e6fb502f01f83c0… | All agent tools are implemented as Python functions decorated with @tool(), discovered at startup through entry points, convention scanning, and a static manifest, and executed in-process through a DecoratedToolServer rather than an external MCP server. | active | contextual | yes | |
| 7cd88a389cd8ce2e… | MessageBuilder manages the full message lifecycle by accumulating conversation history and rendering the complete prompt, including system instructions, history, and reminders, on every turn. | active | contextual | yes | |
| e8a92e96ae6440c2… | AgentContext stores the persisted belief state for an agent run, including plan steps, accumulated knowledge, and artifact references, and it is stored in Redis, read each turn, and updated by tools such as update-plan and update-knowledge. | active | contextual | yes | |
| d24931bd7f999167… | LoopAgentState is the per-run state model that tracks turn control, counters, budget, HITL status, and observations, and it is checkpointed so that agent runs can pause and resume reliably. | active | contextual | yes | |
| 3ab50e5bb246b7e6… | The platform executes action agents through a defined runtime sequence in which a Temporal workflow invokes the execute_action_agent activity, PyranaHarness runs the agent, the LLM creates an initial plan on turn 0, native tool-calling iterates until completion, and the system finalizes by building and persisting an AgentOutputEnvelope. | active | contextual | yes | |
| 8140ca1156b03853… | The Context Engine workspace structure consists of three Python packages so that its codebase is organized into three package-level components. | active | contextual | yes | |
| 23dbe4bf9ab98eee… | The extraction pipeline accesses LLMs through a LiteLLM proxy with an OpenAI-compatible interface so that model routing is handled through a standardized proxy layer. | active | contextual | yes | |
| ba8ffb24b71eaaf2… | In the Context Engine, NATS provides L2 agent coordination using nats-py so that agents can coordinate through a messaging backend. | active | contextual | yes | |
| 04ef771307f33415… | In the Context Engine, Redis serves as the cache using the redis client so that frequently accessed data can be retrieved quickly. | active | contextual | yes | |
| 8c65a80a6207ce59… | In the Context Engine, Qdrant provides vector search with 768-dimensional embeddings using qdrant-client so that semantic retrieval is supported through vector indexing. | active | contextual | yes | |
| c3e85f44af55f96d… | In the Context Engine, Neo4j provides the knowledge graph and TEMPR memory using neo4j-driver so that graph-structured knowledge and memory are managed in Neo4j. | active | contextual | yes | |
| af409eec597ceafa… | In the Context Engine, PostgreSQL serves as the source of truth with JSONB and full-text search support using asyncpg so that canonical structured and searchable data is stored in PostgreSQL. | active | contextual | yes | |
| 3e0f5b81fae1a01b… | The Context Engine is defined as a multi-backend Knowledge-Augmented Graph system with five storage and messaging backends so that knowledge, vectors, cache, and coordination are distributed across specialized services. | active | contextual | yes | |
| 16f821960cfddf46… | Tenant overlays configure platform behavior per tenant and do not contain platform code so that customization is isolated from the core platform implementation. | active | contextual | yes | |
| eea0bfd024491486… | Per-tenant overlays contain agent definitions, workflow configs, data connectors, and SQL templates so that each tenant can configure platform behavior with its own resources. | active | contextual | yes | |
| 404f40feaaefeaf5… | Multi-tenancy is implemented through client overlays in the `clients/` directory, with examples including `clients/jnj/` and `clients/patrick/`, so that tenant-specific assets are separated by client. | active | contextual | yes | |
| fd9898854792af89… | CI/CD uses GitHub Actions and changesets for versioning so that automation and release version management are handled in the delivery pipeline. | active | contextual | yes | |
| 6a2d61f3ad69d6c7… | Authentication uses Azure AD, OAuth, and JWT with PyJWT and MSAL so that identity and token-based access are handled through Azure-compatible auth components. | active | contextual | yes | |
| 9c943726343d23aa… | The frontend tech stack consists of Next.js 16, React 19, Vite, and TypeScript so that the user interface is built with modern web tooling. | active | contextual | yes | |
| 904cf06dc6ed5f3e… | The backend tech stack consists of FastAPI, Temporal, PostgreSQL, Redis, and a LiteLLM proxy so that the platform backend supports APIs, orchestration, storage, caching, and model access. | active | contextual | yes | |
| f5cea4a029918cfa… | The output gateway processes results through the pipeline normalize, derive, validate, and persist so that outputs are transformed and stored in a controlled sequence. | active | contextual | yes | |
| 76b5b6066813670c… | Workflow configurations are stored alongside agent configs in `config/workflows/` so that workflow definitions live with related tenant configuration. | active | contextual | yes | |
| 9356fdfe3cfb6128… | Agents are defined in YAML configuration files under `clients/<client>/backend/config/agents/` so that tenant-specific agent behavior can be declared through configuration. | active | contextual | yes | |
| 4ce892105ebbf53d… | A typical platform request flow proceeds from API to PlannerDelegator to ExecutionPlan to GenericExecutor to `execute_action_agent` to PyranaHarness so that planning and execution move through defined orchestration stages. | active | contextual | yes | |
| 9ba6dab3bb5f0e2f… | BatchOrchestratorWorkflow is defined as the workflow that runs batches of entities through a workflow so that multiple entities can be processed in grouped execution. | active | contextual | yes | |
| 6d87acf0243578a5… | SubagentWorkflow is defined as the workflow that runs spawned child agents through the `dispatch-subagent` tool so that delegated subagents can execute independently. | active | contextual | yes | |
| b0763a42d42003be… | GenericExecutorWorkflow is defined as the workflow that executes DAG-based plans and dispatches `execute_action_agent` for each node so that planned tasks run node by node. | active | contextual | yes | |
| ca1c4165419433de… | PlannerDelegatorWorkflow is defined as the top-level planner that creates execution plans and delegates work to action agents so that planning and delegation are centralized in one workflow. | active | contextual | yes | |
| 8f9acf2893675573… | The platform groups tools into the categories discovery, data fetching, artifacts, context management, completion, delegation, hierarchy, HITL, lifecycle, and context engine so that tool capabilities are organized by function. | active | contextual | yes | |
| 0d28522999af4036… | The tool decorator exposes the flags `requires_context`, `is_base_tool`, `subagent_prohibited`, `approval_required`, and `bypass_size_guard` so that tool behavior can be configured through metadata. | active | contextual | yes | |
| a0994f0569e1b94e… | Each discovered tool is wrapped in a DecoratedToolServer and executed in-process without an external MCP server so that tool execution remains local to the runtime. | active | contextual | yes | |
| cf7a54db7021e9f6… | Tools are discovered at startup through entry points, convention scanning, and a static manifest so that the runtime can assemble the available tool set before execution. | active | contextual | yes | |
| f2cdf844bdd9f2eb… | All agent tools are defined as Python functions decorated with `@tool()` from `inproc_mcp/decorators.py` so that the platform can recognize them as executable tools. | active | contextual | yes | |
| 8724c0713b2667d2… | MessageBuilder is defined as the component that owns the full message lifecycle by accumulating conversation history and rendering the complete prompt, including system text, history, and reminders, on each turn. | active | contextual | yes | |
| c591b954736e9321… | AgentContext is defined as the persisted belief state containing plan steps, accumulated knowledge, and artifact references, and it is stored in Redis, read each turn, and updated by tools such as update-plan and update-knowledge. | active | contextual | yes | |
| d5d6600048e01c26… | LoopAgentState is defined as the per-run state model that tracks turn control, counters, budget, HITL status, and observations, and it is checkpointed for pause and resume behavior. | active | contextual | yes | |
| c5ea921309bcf04f… | The execution flow ends by building and persisting an `AgentOutputEnvelope` so that the final agent output is stored as a durable result. | active | contextual | yes | |
| c5d68726a5f1150a… | The native tool-calling loop repeats by having the LLM emit tool calls, executing those tools, observing results, and iterating so that the agent can progress through tool-mediated reasoning. | active | contextual | yes | |
| 8259d7e7e7bc0477… | On turn 0, the harness creates an initial plan through an LLM so that the agent begins execution with a generated plan. | active | contextual | yes | |
| 0cd46f464c84284a… | After the Temporal activity is invoked, the runtime calls `PyranaHarness.execute(config, workflow_input)` so that the harness can run the configured action agent with workflow input. | active | contextual | yes | |
| e1c3881e055668bf… | The Pyrana harness implements native tool calling through LiteLLM in the orchestration layer so that action agents can execute tool-driven LLM loops within the platform runtime. | active | contextual | yes | |
| 3cad798c15e9fe29… | The workspace structure section states that the context engine workspace is organized as three Python packages so that the codebase is partitioned into three package units. | active | contextual | yes | |
| efb40381143279a9… | The extraction pipeline accesses LLMs through a LiteLLM proxy with an OpenAI-compatible interface so that model routing is handled behind a standardized API. | active | contextual | yes | |
| e6fec5181bf3c1e2… | Within the context engine, NATS provides L2 agent coordination and uses nats-py so that agents can coordinate through a messaging backend. | active | contextual | yes | |
| 37ecba9398d5f0c7… | Within the context engine, Redis serves as the cache backend and uses the redis client so that cached data can be accessed through Redis. | active | contextual | yes | |
| df0bcc260c0f8f7f… | Within the context engine, Qdrant provides vector search with 768-dimensional embeddings and uses qdrant-client so that semantic retrieval is handled by a dedicated vector database. | active | contextual | yes | |
| 8a852424d1242282… | Within the context engine, Neo4j provides the knowledge graph and TEMPR memory and uses the neo4j-driver so that graph-structured knowledge is managed in Neo4j. | active | contextual | yes | |
| 15cf197a526fbbdf… | Within the context engine, PostgreSQL serves as the source of truth with JSONB and full-text search capabilities and uses asyncpg so that authoritative structured and searchable data is stored in PostgreSQL. | active | contextual | yes | |
| 575009c8ee2a9113… | The context engine is a multi-backend Knowledge-Augmented Graph system with five storage and messaging backends so that knowledge, vectors, cache, graph data, and coordination are distributed across specialized services. | active | contextual | yes | |