Summary

When the model returns multiple tool calls in a single response, they are now executed concurrently using a thread pool instead of sequentially. This significantly reduces wall-clock time when multiple independent tools are batched (e.g. parallel web_search, read_file, terminal calls).

Architecture

• _execute_tool_calls() now dispatches between sequential and concurrent paths
• Sequential path (renamed to _execute_tool_calls_sequential): unchanged original behavior, used for:
  • Single tool calls (no overhead needed)
  • Batches containing clarify (requires interactive user input)
• Concurrent path (_execute_tool_calls_concurrent): uses ThreadPoolExecutor with up to 8 workers
• _invoke_tool() extracted as shared tool invocation helper (handles both agent-level tools and registry-dispatched tools)

Safety Considerations

• Pre-flight interrupt check: skips all tools if interrupted before starting
• Per-tool exception handling: one failure doesn't crash the batch
• Result truncation (100k char limit) applied per tool
• Budget pressure injection after all tools complete
• Checkpoints taken before file-mutating tools (before submitting to pool)
• CLI spinner shows batch progress, then per-tool completion messages
• _NEVER_PARALLEL_TOOLS frozenset controls which tools force sequential execution
• _MAX_TOOL_WORKERS = 8 caps thread pool size

Display/UX

• quiet_mode (CLI): Single spinner "running N tools concurrently" while executing, then individual cute messages per tool
• verbose mode: Prints all tool calls upfront, then completion messages with timing

Tests

10 new tests in TestConcurrentToolExecution:

• Dispatch routing (single → sequential, clarify → sequential, multi → concurrent)
• All tools executed and results in correct order
• Order preserved despite different tool execution times
• Error handling (one tool fails, others succeed)
• Interrupt before start skips all tools
• Result truncation in concurrent path
• _invoke_tool routing for regular and agent-level tools

Fixed 1 existing test (test_interrupt.py) that used MagicMock and needed real dispatch methods bound.

3374 tests passing, 0 new failures.