[ian-at-strix]

What

Adds an opt-in STRIX_PARALLEL_TOOL_CALLS setting (llm.parallel_tool_calls, default false). When on:

• requests are sent with parallel_tool_calls=True instead of False;
• the interactive prompt stops asking for exactly one tool call per turn;
• the autonomous prompt lets independent read-only calls (file reads, searches) share a turn, while anything that changes state, sends requests, spawns/messages agents, reports, or finishes stays alone.

The tool-required nudge now says "call a tool" instead of "call exactly one tool" in both modes. Otherwise, with the flag off, requests and the rendered system prompt are byte-identical to main.

Enabling

STRIX_PARALLEL_TOOL_CALLS=true (env or cli-config.json). The existing LLM_MAX_TOOL_CALLS_PER_TURN cap still applies.

Notes

• Evals showed that, on its own, this did not make a big difference to PR-review time. It is not an eval candidate.
• This is a clean, opt-in port of the experiment in feat(core): allow parallel tool calls and stop asking for exactly one tool call per turn #1546 / feat/parallel-reads.
• The runner tests (test_runner_*) already fail on main because their settings stubs have no telemetry (since aa2036f). This PR adds parallel_tool_calls=False to those stubs but does not fix the telemetry failures.
• Related: fix(llm): index Ollama's streamed tool calls so parallel calls stay separate #1420 (Ollama parallel tool-call indexing), feat(tools): batch_view_files / batch_terminal_execute — sandbox-side concurrency #791 (batch read/exec tools).