AI/Tasks/OrchestratorUpgrades/Implementation_Overview.txt
|
IMPLEMENTATION PROMPT SET - ORCHESTRATOR UPGRADES
Purpose: - These prompts convert the design-phase specs into implementation tasks that produce production code and tests. - Run phases in order. Each phase assumes prior phases are complete. Execution model: - For each phase, modify production code and tests in the orchestrator execution boundary. - Keep edits bounded and deterministic. - Preserve safety invariants: never weaken authorization, confirmation, sandboxing, refusal rules, or execution safety. Primary target surfaces: - src/TechToolbox.Agent/Orchestrator/* - src/TechToolbox.Agent/Execution/* - src/TechToolbox.Agent/Tools/* (only direct execution, validation, and routing seams) - src/TechToolbox.Agent/Tests/*Orchestrator* and related focused test files Global contracts: - State model precedence: architecture policy > execution semantics > explicit user intent > heuristic recovery details. - Security pipeline order: resolve -> validate -> authorize -> execute -> normalize. - Shared contracts remain stable across phases: run-state semantics, decision schema, tool authorization ordering, cancellation and deadline semantics, retry and repair budgets, path and file-operation safety, telemetry redaction, and retention behavior. - Logs, telemetry, and diagnostics must be redacted and resource-bounded. - Mandatory cross-phase requirements: - Version every shared contract for state, decisions, tool results, telemetry, and memory records, with migration or fail-closed rejection behavior for unknown versions. - Publish an ownership matrix assigning run-state, decision validation, repair, authorization, path checks, retry classification, execution, finalization, telemetry redaction, and memory precedence to single owners or components. - Define precedence rules and conflict outcomes for user intent vs. memory preference vs. policy blocks vs. retry or deadline exhaustion. - Use explicit numeric resource ceilings for total run duration, iteration limits, tool counts, model calls, repair attempts, loop history, telemetry retention, and output-size limits. - Add end-to-end acceptance scenarios covering safe normal execution, blocked mutation, cancellation, retryable and deterministic failures, no-progress loops, low-confidence location selection, prompt injection, and concurrent orchestrator use. - Add rollback/compatibility guidance for any contract migration or compatibility wrapper introduced during rollout. Per-phase output format: 1) Summary 2) Files changed 3) Code changes 4) Tests added/updated 5) Validation results 6) Known risks Validation baseline per phase: - dotnet build src/TechToolbox.Agent/TechToolbox.Agent.csproj -c Release - dotnet test src/TechToolbox.Agent/Tests/TechToolbox.Agent.Tests.csproj -c Release --filter "Orchestrator|Execution|Tool|Safety|Retry|Loop|Policy|Telemetry|Weather|Authorization" Stop conditions: - Stop if a requested change would violate safety invariants. - Stop after three focused repair cycles for phase-related failures and report exact failing assertions. |