Hardened AI orchestration reliability and reduced non-converging edit loops

[AbhishekSinghDhadwal]

This PR hardens AI tool-calling reliability and loop termination behavior in CQL Studio with minimal runtime-focused changes (no API or settings schema changes). It also improves one-shot code-edit completion to reduce unnecessary post-edit self-correction loops.

What changed

1) Act-mode response contract hardening

• Enforced explicit structured act responses with required next_action: "tool" | "final".
• Added runtime validation for tool-call shape and conditional requirements.
• Kept backward compatibility for legacy structured responses by inferring action.

Files:

• src/app/services/ai.service.ts
• src/app/services/ai-planning.service.ts

2) Deterministic tool-phase decoding

• Added mode-specific generation options to reduce random behavior in tool orchestration:
  • Act+tools: lower temp/top_p
  • Plan+tools: constrained temp/top_p
  • Content-only path unchanged (more creative)

File:

• src/app/services/ai.service.ts

3) Bounded continuation + no-progress safeguards

• Added structured continuation outcome metadata (reason, progressDelta).
• Added guardrails:
  • max continuation rounds
  • max no-progress rounds
• Stop reason is explicit and user-visible.

Files:

• src/app/services/ai-stream-response-handler.service.ts
• src/app/components/cql-ide/tabs/ai-tab/ai-tab.component.ts

4) Tool execution resilience

• Increased default tool timeout from 5s to 25s.
• Added one retry for transient failures (timeout/network/429/5xx-like).
• Kept no-retry behavior for validation failures.
• Increased tool summary truncation limit to preserve more context for continuation.

File:

• src/app/services/ai-tool-execution-manager.service.ts

5) Token-aware context compaction

• Enforced approximate token budget in API message selection.
• Preserves system messages and most recent relevant turns under budget.
• Maintains deterministic behavior without adding semantic summarization complexity.

File:

• src/app/services/conversation-manager.service.ts

6) One-shot edit loop reduction

• insert_code / replace_code now validate non-empty code and return concise accepted metadata instead of verbose placeholder payloads.
• UI edit application now ignores failed edit tool results.
• Added explicit completion policy so simple “create/insert and done” requests finalize rather than entering unsolicited validation/refactor loops.

Files:

• src/app/services/tools/insert-code.tool.ts
• src/app/services/tools/replace-code.tool.ts
• src/app/components/cql-ide/tabs/ai-tab/ai-tab.component.ts
• src/app/services/ai.service.ts

Why

It was observed that there existed multiple tool chain orchestration issues like - flaky tool-call chaining, random loop continuation, and non-converging edit-repair cycles. These changes prioritize deterministic control flow, bounded retries/loops, and better completion behavior for common user intents.

Scope

• Runtime reliability and orchestration only.
• No public HTTP API changes.
• No new user-facing settings.
• No backend endpoint changes.

Sources I've referred to in order to implement these changes

• OpenAI Function Calling: https://platform.openai.com/docs/guides/function-calling
• Anthropic Tool Use Overview: https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/overview
• MCP Tools Specification: https://modelcontextprotocol.io/specification/2025-06-18/server/tools
• Lost in the Middle (TACL): https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00638/119630/Lost-in-the-Middle-How-Language-Models-Use-Long

NOTE - This is not a final fix, and is simply a step towards strengthening the application. Future fixes will further strengthen the tool calling chain and associated components, especially towards code completion.