fix(llm): omit parallel_tool_calls on tool-less calls (Azure OpenAI 400)

[herman-young-ai]

Problem

make_model_settings sets parallel_tool_calls=False unconditionally, but the parameter is only valid alongside a tool list. Azure OpenAI rejects it outright:

BadRequestError - Invalid value for 'parallel_tool_calls':
'parallel_tool_calls' is only allowed when 'tools' are specified.

The native OpenAI route silently tolerates it, so the divergence only surfaces on Azure deployments — where it currently makes Strix unusable end to end. warm_up_llm fails before any scan starts, so strix --target ... exits 1 with LLM CONNECTION FAILED.

Scope

Four call paths send no tools, and all four break on Azure:

Call site	When	Impact
warm_up_llm, main model	startup	blocks the scan (sys.exit(1))
warm_up_llm, dedupe model	startup	blocks the scan
compaction._summarize	context overflow, mid-scan	summarization fails
dedupe._dedupe_model_settings	per vulnerability report	dedupe fails

Only the agent loop in core/runner.py carries tools and is unaffected — so fixing the warm-up alone would still leave a scan that dies the first time context fills or a finding is recorded.

Change

Add an explicit with_tools keyword to make_model_settings, defaulting to the current behaviour, and set it False at the four tool-less call sites. ModelSettings drops None fields, so the parameter is simply not sent rather than sent as false.

No behaviour change on any existing route: every tool-carrying call keeps parallel_tool_calls=False.

Verification

Reproduced and confirmed fixed against a live Azure OpenAI deployment (azure/gpt-5.4, Entra ID auth):

• Before: warm-up 400s, scan never starts.
• After: warm-up succeeds, sandbox comes up, root agent registers, scan proceeds.

Two regression tests added in tests/test_inputs.py covering both branches.

ruff check / ruff format / mypy clean on all changed files. Full suite: 628 passed, with one pre-existing order-dependent failure in tests/test_execution.py::test_finish_scan_bypasses_active_agent_guard_after_reserve that reproduces on main without this change (it passes in isolation; tests/test_cost_tracking.py leaves an uninitialised ReportState in the module-level global).

[greptile-apps]

Greptile Summary

This PR prevents Azure OpenAI from receiving parallel_tool_calls on requests that do not provide tools.

• Adds a with_tools switch to model-settings construction while preserving existing behavior for agent calls that carry tools.
• Marks warm-up, deduplication, and compaction requests as tool-less.
• Adds regression coverage for both tool-bearing and tool-less settings.

Confidence Score: 5/5

The PR appears safe to merge, with every current tool-less model request updated consistently and the tool-bearing agent path preserving its existing setting.

All four changed production callers issue requests without tools and now omit the Azure-incompatible parameter, while the only remaining default path constructs agents with a nonempty tool list.

Important Files Changed

Filename	Overview
strix/core/inputs.py	Adds the with_tools setting and omits parallel_tool_calls for tool-less requests while retaining the existing default for agent calls.
strix/interface/main.py	Correctly marks both startup warm-up requests as tool-less, matching their empty tool lists.
strix/llm/compaction.py	Correctly omits the tool-specific setting from tool-less context summarization requests.
strix/report/dedupe.py	Correctly omits the tool-specific setting from the text-only deduplication judge.
tests/test_inputs.py	Covers both branches of the new settings behavior by asserting false for tool-bearing calls and None for tool-less calls.

_{Reviews (1): Last reviewed commit: "fix(llm): omit parallel_tool_calls on to..." | Re-trigger Greptile}