Fix PersonaToolExecutor: 4 bugs breaking tool calls (#368)

[joelteply]

Summary

• String(v) destroying object/array params — String({foo:1}) → "[object Object]" (unrecoverable). Fixed with JSON.stringify for non-primitives in both PersonaToolExecutor and AgentToolExecutor
• Force-text at 3 iterations too aggressive — Frontier models (Anthropic/OpenAI) now get 10 iterations before forced text (was 3). Enables read→edit→test→fix chains
• Double correction eliminated — Added executeCorrectedToolCall() so PersonaToolExecutor doesn't re-correct through executeToolCall()
• Loop detection side-effect — Result mapping was re-calling isLoopDetected() (which has side effects). Now uses pre-partitioned set
• Tool group example bias — code-read group's only example was code/search, teaching LLMs to search instead of read known files. Now shows both with guidance

Also filed #399 (persona response latency) and added to gap analysis Phase 2.

Test plan

• TypeScript compilation clean
• Deployed via npm start (133s)
• AI QA: asked personas to use code/read on src/package.json
• Fireworks AI: correct (1.0.8900)
• DeepSeek Assistant: correct (1.0.8900)
• Groq Lightning: tool executed (385 lines read)
• All personas used code/read not code/search

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes multiple defects in the persona/agent tool execution pipeline that were causing incorrect tool parameter passing, premature tool-loop termination, double-correction, and loop-detection side effects; it also updates tool-group prompting guidance to reduce model bias toward searching instead of reading.

Changes:

• Preserve structured tool parameters by JSON-stringifying object/array inputs when converting native tool calls to internal ToolCall format.
• Add executeCorrectedToolCall() to avoid double-correcting already-corrected tool calls, and remove loop-detection re-invocation with side effects.
• Increase provider-tiered “force text” threshold to allow longer tool chains for frontier models; update tool-group examples/guidance and planning docs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/system/user/server/modules/PersonaToolExecutor.ts	Avoids re-correction by calling executeCorrectedToolCall() and updates native tool-call param stringification.
src/system/user/server/modules/PersonaAgentLoop.ts	Makes “force text” threshold provider-tiered, allowing longer tool chains for capable providers.
src/system/tools/server/AgentToolExecutor.ts	Adds executeCorrectedToolCall(), improves native param serialization, and fixes loop-detection side-effect behavior during result mapping.
src/system/rag/sources/ToolGroupRegistry.ts	Adjusts code-read group description/example to encourage code/read for known files.
docs/planning/ALPHA-GAP-ANALYSIS.md	Updates gap analysis to include #399 and status/details for #368.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.