fix(claude-code): disable built-in tools to prevent MCP tool deferral in call_with_tools

[mkremnev]

Summary

Fixes flaky tool calling in call_with_tools() when using the claude-code LLM provider. Follow-up to #732 and PR #733.

Root Cause

When ClaudeAgentOptions doesn't specify tools, Claude Code CLI loads all built-in tools (~30+ tools: Read, Write, Bash, Grep, ToolSearch, Agent, etc.) alongside the MCP tools from hindsight_tools server.

With this many tools loaded, the CLI defers MCP tools to save context space. Deferred tools are registered by name only — their schemas aren't loaded into the context. To use a deferred tool, Claude must first call the built-in ToolSearch tool to load its schema.

This creates a fatal interaction with max_turns=1:

1. Claude sees MCP tools are deferred → calls ToolSearch to load schemas (1 turn spent)
2. max_turns=1 exhausted → ResultMessage(subtype="error_max_turns")
3. No actual MCP tool call happens → tool_calls=[] → reflect agent gets tools=[none]

The behavior is flaky because tool deferral depends on total context size. Sometimes MCP tools load eagerly (works), sometimes they're deferred (fails). Evidence from logs:

# Same query, same session — sometimes works, sometimes doesn't:
17:17:38 tools=[search_mental_models, search_observations]  ← WORKS (eager)
17:32:05 tools=[none]                                       ← FAILS (deferred)
17:38:09 tools=[search_mental_models, search_observations, recall]  ← WORKS
17:38:30 tools=[none]                                       ← FAILS

When deferred, stream logs show Claude calling ToolSearch instead of MCP tools directly:

[claude-code-stream] ToolUseBlock: name=ToolSearch,
  input={'query': 'select:mcp__hindsight_tools__search_mental_models,...'}
[claude-code-stream] ResultMessage: subtype=error_max_turns, result=None

Fix

Two changes in call_with_tools():

1. tools=[] — Disables all built-in CLI tools. With only MCP tools registered, the CLI loads them eagerly (schemas in context). Claude calls them directly without needing ToolSearch.

2. max_turns=2 — Allows the tool call + tool result round-trip. With max_turns=1, even when tools loaded eagerly, the single turn was consumed by the tool execution, leaving no room for additional processing.

Testing

Before fix (with debug logging):

[claude-code-stream] ToolUseBlock: name=ToolSearch  ← wasted on schema lookup
[claude-code-stream] ResultMessage: error_max_turns

After fix:

[claude-code-stream] ToolUseBlock: name=mcp__hindsight_tools__search_mental_models  ← direct call!
[claude-code-stream] ToolUseBlock: name=mcp__hindsight_tools__search_observations
[claude-code-stream] ToolUseBlock: name=mcp__hindsight_tools__recall
[claude-code-stream] ToolUseBlock: name=mcp__hindsight_tools__done  ← success
[claude-code-stream] ResultMessage: subtype=success

Tested with budget: "low" (previously always failed) — now consistently returns real data from memory bank.

Environment

• hindsight-api-slim v0.4.21
• Claude Code CLI v2.1.x
• claude-agent-sdk (Python)
• macOS arm64

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.6 (1M context) noreply@anthropic.com

[nicoloboschi]

LGTM