When `agent.tool_use_enforcement` is `"auto"` (the default), the
runtime checks the active model name against `TOOL_USE_ENFORCEMENT_MODELS`
in `agent/prompt_builder.py` and only injects `TOOL_USE_ENFORCEMENT_GUIDANCE`
if a substring matches. Qwen and DeepSeek hit the same chatty/hallucinatory
failure mode as GPT, Codex, Grok, and GLM (describing intended actions
instead of calling tools, ignoring memory, silently stopping mid-execution),
but neither substring was in the tuple — so the enforcement prompt was
never injected for users on those families, even with `auto` left at
its default.
Add `"qwen"` and `"deepseek"` to the tuple, matching the established
additive pattern (#5595 added grok, #24715 added glm, #27797 widened
grok to xai-oauth). Add four regression-guard tests that fail before
the production change and pass after: two unit assertions in
`test_prompt_builder.py` mirroring the existing grok/gpt checks, and
two integration tests in `test_run_agent.py` confirming that a
qwen/deepseek model under `tool_use_enforcement="auto"` now gets the
guidance string in its system prompt.
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 / non-listed families
unsteered, and the maintainer's prior merged work in this area is
uniformly additive.
Fixes#28079
briandevans
403e567cec