AI/Tasks/OrchestratorUpgrades/Orchestrator_Phase04_ValidationRepairLoopProgress_Implementation.txt
|
PHASE 04 IMPLEMENTATION PROMPT - VALIDATION, REPAIR, AND LOOP PROGRESS HARDENING
Goal: Implement structured decision validation, explicit repair budgets, and bounded progress detection so malformed decisions and no-progress loops are caught early. Inputs: - AI/Tasks/OrchestratorUpgrades/Orchestrator_Phase04_ValidationRepairLoopProgress_Strict.txt - AI/Tasks/OrchestratorUpgrades/Overview.txt Required code work: - Add a versioned decision schema and structured validation results with field-level error reporting. - Introduce a decision-repair budget and enforce it globally. - Add bounded validation for long final answers and tool decisions. - Replace coarse repeated-call detection with bounded recent-history tracking and progress signatures. - Distinguish same-call, same-target, and same-evidence loop detection with explicit actions. - Publish the decision schema contract, define validation ordering, and ensure unknown or malformed field combinations are rejected before routing or execution. - Ensure repair output is treated as untrusted and passes through the same validation and policy checks again before execution. Implementation constraints: - Keep execution policy, authorization, and retry logic separate from validation and loop detection changes. - Preserve current behavior except where malformed decisions and loops are prevented or handled deterministically. - Do not implement domain-specific weather routing or memory privacy hardening in this phase. Minimum tests to add/update: - Validation error tests for malformed and unknown tool decisions. - Repair budget exhaustion tests. - Loop-progress detection and no-progress tests. - Regression tests for repeated read/search/NOAA-like behaviors. Validation commands: - dotnet build src/TechToolbox.Agent/TechToolbox.Agent.csproj -c Release - dotnet test src/TechToolbox.Agent/Tests/TechToolbox.Agent.Tests.csproj -c Release --filter "Orchestrator|Decision|Validation|Repair|Loop|Guard|Strict" Deliverable: Return changed files, the final validation and loop-progress contract, and proof that malformed decisions and no-progress loops are contained. |