What does this PR do?

When agent.tool_use_enforcement is left at its default "auto", agent/system_prompt.py only injects TOOL_USE_ENFORCEMENT_GUIDANCE if the active model name contains a substring from TOOL_USE_ENFORCEMENT_MODELS in agent/prompt_builder.py:271. The tuple was ("gpt", "codex", "gemini", "gemma", "grok", "glm") — both qwen and deepseek were missing, even though both families exhibit the exact failure mode the enforcement prompt was written for (describing intended actions instead of calling tools, hallucinating execution, ignoring existing context/memory, silently stopping mid-task).

This PR adds "qwen" and "deepseek" to the tuple, mirroring the established additive pattern in this area (merged #5595 added grok, #24715 added glm, #27797 widened grok to xai-oauth).

Mirrors agent/system_prompt.py::_build_system_prompt matching logic (substring in model_lower against the tuple) — this is the single source of truth for tool_use_enforcement="auto", so no other input precedence chain needs covering.

The "robust" alternative from the issue (default-true for all models) is intentionally not taken — it would silently flip behavior for users who currently rely on auto leaving Claude and other non-listed families unsteered, and every prior merged change in this area has been additive.

Related Issue

Fixes #28079

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• agent/prompt_builder.py — append "qwen" and "deepseek" to TOOL_USE_ENFORCEMENT_MODELS. 1-line tuple edit.
• tests/agent/test_prompt_builder.py — add test_enforcement_models_includes_qwen and test_enforcement_models_includes_deepseek, mirroring the existing _includes_gpt / _includes_codex / _includes_grok pattern.
• tests/run_agent/test_run_agent.py — add test_auto_injects_for_qwen (model qwen/qwen3.6-plus) and test_auto_injects_for_deepseek (model deepseek/deepseek-r1) in TestToolUseEnforcementConfig, confirming that tool_use_enforcement="auto" now causes the guidance string to appear in the system prompt for both families.

How to Test

1. Regression guard (proves the fix is necessary):

   git stash push -- agent/prompt_builder.py
   uv run --with pytest --with pytest-xdist --with pytest-asyncio python3 -m pytest \
     tests/run_agent/test_run_agent.py::TestToolUseEnforcementConfig::test_auto_injects_for_qwen \
     tests/run_agent/test_run_agent.py::TestToolUseEnforcementConfig::test_auto_injects_for_deepseek \
     tests/agent/test_prompt_builder.py::TestToolUseEnforcementGuidance::test_enforcement_models_includes_qwen \
     tests/agent/test_prompt_builder.py::TestToolUseEnforcementGuidance::test_enforcement_models_includes_deepseek -v
   # 4 failures (regression proved)
   git stash pop

2. With the production fix:

   uv run --with pytest --with pytest-xdist --with pytest-asyncio python3 -m pytest \
     tests/agent/test_prompt_builder.py::TestToolUseEnforcementGuidance \
     tests/run_agent/test_run_agent.py::TestToolUseEnforcementConfig -v
   # 9 + 18 = 27 passed

3. Real-world: set model.default: qwen/qwen3.6-plus (or any deepseek model), leave agent.tool_use_enforcement at its default "auto", run hermes chat -q "list files in this directory". Before the fix the agent often replies with a narrated plan and no tool call; after the fix TOOL_USE_ENFORCEMENT_GUIDANCE appears in the system prompt and tool use is enforced.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(agent):)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix (no unrelated commits)
• I've run focused tests for the touched code and all pass (9 unit + 18 integration assertions in the two test classes)
• I've added tests for my changes — 4 new tests that all fail without the production change
• I've tested on my platform: macOS 26.x, Python 3.11

Documentation & Housekeeping

• I've updated relevant documentation — N/A (the tuple has an inline # Add new patterns here when a model family needs explicit steering. comment that already documents the maintenance pattern)
• I've updated cli-config.yaml.example if I added/changed config keys — N/A (no config keys changed; the existing agent.tool_use_enforcement key behaviour is unchanged for explicit values, only the auto default is widened)
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact — N/A (pure string-substring match, no platform-specific code paths)
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Sibling code paths that may need the same fix

The issue body explicitly flags mistral and llama as potentially affected (no concrete repro provided). I left them out of this PR's scope to keep the diff minimal and the additions evidence-driven — happy to widen to ("mistral", "llama") during cherry-pick if the same chatty-narration failure mode is confirmed on those families.

Screenshots / Logs

N/A — guidance-string injection has no UI surface; the change is observable only via system-prompt assembly, which is exercised by the new tests.