System Map

The Machine Behind the Answer

All layers. No simplified picture. Two paths, one bottleneck, zero free decisions by the model.

INTELLIGENCE_MODULES/Prompts · Capability Maps · cim_policy.csv · execution_mode_signals.csv · tool_intents.jsoncim_policy.csvexecution_mode_signals.csvtool_intents.jsonprompts/capability_maps/inputuser textClassifierPattern-Matching · 0 LLMcategory · route · safety_levelBOTTLENECKrouting_frameoperation_contract7 axes · built once · read by allintent · domain · evidence_need · confidenceSIMPLE PATHThinkingplans · selects tools · 1 LLMselected_tools from contract1 LLMINPUTrouting_frame · operation_contractselected_tools · memory_contextavailable_tool_details (context, no revival)OUTPUT — ThinkingPlanneeds_loop: boolrepeat_count_hint · operation_family_hintselected_tools: [T1, T2, ...]Rule: selected_tools empty → no tool from available_tool_detailsVerifierdeterministic · 1 LLMGuardDecision · anti_patterns1 LLMrejected → ReplanTask Loop0 LLM · deterministicevidence ⊆ required → doneOutputStreaming · 1 LLM · ClaimCheckevidence → no guessing1 LLMCOMPLEX PATHOrchestratortool filter · context · 0 LLMeligible_tools_for_contract()needs_orchestrator: true1. Hard Gates (T_eligible)domain · operation · scope · evidence · risk2. Args Binding (T_executable_now)required_args bound · ToolDescriptor match3. Ranking (within eligible set only)ohne Contract → eligible_tools = [] · fail-closedselected_tools: [T1, T2, ...]context: memory · workspace · runtimeMemory · RuntimeSTM · MTM · LTM · Workspaceconversation scope · home_contextcontextSTMconversation scope · current sessionMTMlast N sessions · configurable windowLTMpersistent facts · home_contextMEMORY_RETRIEVAL_FILTER_ENABLE: false (default)Per-conversation isolation designed, not yet activeMCP HubDiscovery · Dispatch · HealthHTTP-MCP · latency checktool resultMemory MCPsearch · store · retrieveContainer Commanderlist · inspect · read-onlyTime MCPdatetime · timezone{type, content, evidence_tags}Gatedoneblocked → next iterationrequired_evidence ⊆ observed_evidenceVerifier checksClaimType · GuardDecision · anti_patternsAPPROVED → continue · REJECTED / HARD_BLOCK → ReplanOutput checks every claimruntime · hardware · file → Evidence requiredconceptual analysis → freely alloweddone = required_evidence ⊆ observed_step_evidence · else: blockedTool success alone does not complete a task. No guessing. No fallback.LEGENDLLM step — Thinking (1) · Verifier (1) · Output (1) = 3 maxDeterministic — Classifier · Task Loop · Orchestrator · MCP HubComplex path only — Orchestrator · Tool-Executionlive_claim ≠ meaning ≠ operation · routing_frame ist SSOTdone = required_evidence ⊆ observed_evidence · else: blocked
routing_frame

Built once. Read by all. No downstream module re-classifies the request from raw text. What the frame says is authoritative — for Orchestrator, Thinking, Task Loop, and Output simultaneously.

Operation Contractlive_claim ≠ meaning ≠ operation

What the user says, what it means, and what the system is allowed to do — three separate layers. The contract is set once before any tool is considered. Thinking, Replan, and Observer all read the same immutable contract. No module may reinterpret it.

Task Loopdone = required_evidence ⊆ observed_evidence

Done does not mean "tool returned success." Done means the required evidence types were actually collected. No LLM, no guessing — deterministically measurable.

Tool Dispatchtool_name → Task Loop executor → MCP client

The tool name chosen upstream carries straight through the Task Loop executor to a generic MCP client / catalog dispatch — no per-domain branching. Zero LLM, no model involved. A CRON/SKILL/generic domain-router config exists but has no wired consumer in the executed path yet.