fix: harden system message tools and wire toolOverrides to system message path

[shanevcantwell]

Summary

• Strengthened system message tool instructions for models on the non-native tools path (used by local models via capabilities: [] or experimental.onlyUseSystemMessageTools). The previous prose instructions were too vague — local models reverted to XML/JSON tool call formats from their training priors.
• Fixed toolOverrides being silently ignored on the system message path. applyToolOverrides() already existed and ran on the native tools path (in BaseLLM.streamChat()), but was never called when tools were injected via system message text. Config YAML toolOverrides for disabled and description now take effect on both paths.

Changes

1. core/tools/systemMessageTools/toolCodeblocks/index.ts — Replaced prose systemMessagePrefix/systemMessageSuffix with explicit numbered rules that prohibit XML tags, JSON format, and tool invention
2. core/index.d.ts — Added toolOverrides to ModelDescription interface
3. core/config/load.ts — Serialized toolOverrides in llmToSerializedModelDescription()
4. gui/src/redux/thunks/streamNormalInput.ts — Imported applyToolOverrides and applied model-level overrides to activeTools before both the native and system message paths

Test plan

• cd core && npm run vitest — 1654 passed, 0 failures
• Built vsix, tested with local models (nemotron-3-nano-30b, gpt-oss-20b, mirothinker-v1.5-30b) via LM Studio
• Verified view_diff with disabled: true is absent from system message
• Verified hardened numbered rules appear in system message prefix/suffix
• Verified local models follow codeblock format through multi-step tool call chains
• Verify native tools path unaffected (cloud models)

🤖 Generated with Claude Code

---

Summary by cubic

Improves reliability of system-message tool calls for local models and applies model-level toolOverrides on that path. Local models now follow the codeblock format, and disabled/description overrides from config are respected.

• Bug Fixes
  • Hardened system message with clear, numbered rules: use only tool code blocks; no XML/JSON; one tool per response; code block last; no invented tools.
  • Applied applyToolOverrides before both native and system-message paths and added toolOverrides to ModelDescription serialization, so per-model disabled/description overrides take effect.

^{Written for commit 1a59f1a. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 4 files