LLM-05: LLM tool-calling implementation tracker

[Chris0Jeky]

Overview

Implementation tracker for LLM tool-calling / function-calling architecture in Taskdeck's chat system.

Spike: #618 (completed) — architecture document at docs/spikes/SPIKE_618_COMPLETED.md

Key Decisions (from spike)

• Custom implementation over Semantic Kernel (~800 LOC, zero new dependencies)
• Extend ILlmProvider with CompleteWithToolsAsync() — incremental, no breaking changes
• 11 tools: 5 read (execute directly) + 6 write (always produce proposals per GP-06)
• New ToolCallingChatOrchestrator wraps ChatService — max 5 rounds, SignalR intermediate states
• Mock provider simulates tool calls deterministically via pattern-matching dispatch
• Cost: ~$0.00088 per 3-round conversation on GPT-4o-mini

Implementation Phases

• Phase 1: Read tools + orchestrator (LLM-06: Phase 1 — Read tools, orchestrator, and provider tool-calling extension #649)
• Phase 2: Write tools + proposal integration (LLM-07: Phase 2 — Write tools and proposal integration #650)
• Phase 3: Refinements — loop detection, cost tracking, feature flag (LLM-08: Phase 3 — Tool-calling refinements (loop detection, cost tracking, feature flag) #651)

Dependencies

• Builds on existing ITaskdeckToolRegistry and AgentPolicyEvaluator (delivered in AGT-02: Add tool registry, policy evaluator, and bounded inbox-triage template #337)
• Unblocks conversational refinement (Conversational refinement loop for ambiguous chat proposals #576)
• Unblocks MCP server tool inventory (MCP-02: Phase 2 — Full MCP resource and tool inventory #653)

References

• docs/spikes/SPIKE_618_COMPLETED.md — full architecture document
• docs/decisions/ADR-0017-agent-tool-registry-review-first.md — tool registry design
• docs/GOLDEN_PRINCIPLES.md GP-06 — review-first automation safety