Promote per-model settings to a PerModelSettings capability

[mpfaffenberger]

What

Fourth entry in the capability-conversion series (#828 SteerInjection, #829 HistoryCompaction, #830 PluginMessageTransform): the per-model ModelSettings payload — auto-derived max_tokens budget, GPT-5 reasoning effort/verbosity, Anthropic extended-thinking, GLM extra_body riders, Copilot API translation, and the yolo-mode parallel_tool_calls gate — previously rode in on the Agent(model_settings=...) constructor kwarg at both construction sites. It now travels through pydantic-ai's first-class get_model_settings capability seam as PerModelSettings in code_puppy/agents/_model_settings.py, joining the rest of the agent's behavior in the capabilities=[...] block.

Upstream precedent: pydantic-ai's own Thinking capability (capabilities/thinking.py) is exactly this shape — a @dataclass subclass returning a static ModelSettings from get_model_settings().

Feature parity

• Identical merged payload on the standard path. pydantic-ai layers settings agent → capability → run over the model's base settings (_layer_model_settings, verified in the installed 2.31.0 source). Nothing else contributes capability-level settings and per-run settings (agent.run(model_settings=...)) still merge last, so the payload the model receives is unchanged everywhere code_puppy builds agents. Proven by a FunctionModel wire-parity test against the old kwarg wiring, plus a full three-layer merge test (model base < capability < per-run).
• One deliberate, reviewer-flagged divergence: Agent.override(model_settings=...) replaces the agent-slot settings — previously ours, now empty — so the capability snapshot merges after it and wins conflicting keys; the built agent's model_settings attribute also reads None. No code_puppy call site uses either seam (verified: zero .override( callers, zero readers of the built agent's model_settings). The trade-off is documented in the module docstring and pinned by a dedicated regression test so it stays visible rather than accidental.
• Identical snapshot timing. The kwarg froze settings at agent-build time; get_model_settings is re-extracted on every run (for_run re-resolution), which would silently pick up config edits mid-session. __post_init__ therefore snapshots make_model_settings() once — config changes keep applying on agent rebuild, exactly as before.
• Builder still computes once. build_pydantic_agent constructs a single PerModelSettings shared by both construction passes, mirroring the old single make_model_settings() call. Sub-agents construct theirs inline (single pass).
• List position is inert — get_model_settings is a configuration seam, not a request hook; documented at both call sites.
• Spec-constructible. Like Promote plugin message transform to a first-class capability #830 (and unlike refactor: promote steer injection to a first-class pydantic-ai capability #828/refactor: promote history compaction to a first-class pydantic-ai capability #829), the inherited get_serialization_name() default is kept: the only fields are plain data (model_name/max_tokens), no live agent references. The spec path is tested end-to-end via from_spec.

make_model_settings itself is untouched and stays exported from model_factory — plugins that import it directly are unaffected.

Tests

• tests/agents/test_model_settings_capability.py: snapshot-freeze contract, make_model_settings equivalence, FunctionModel wire-parity vs the old kwarg, three-layer merge precedence, the pinned Agent.override divergence, and from_spec construction.
• Two existing tests stubbed _builder.make_model_settings; retargeted to code_puppy.agents._model_settings.make_model_settings and their stubs now return ModelSettings() instead of an annotation-invalid None.
• Full suite: 7494 passed, ruff clean. The lone red (test_render_version_check_current) is pre-existing on main.

Heads-up

Like #828/#829/#830, this grazes the shared capabilities=[...] blocks in _builder.py and subagent_invocation.py — whichever of the four lands last eats a trivial rebase.

Do not merge — awaiting human review.