Promote model delivery to a ResolvedModel capability

[mpfaffenberger]

Summary

Sixth in the capability series (#828 SteerInjection, #829 HistoryCompaction, #830 PluginMessageTransform, #831 PerModelSettings, #832 AssembledInstructions): promotes model delivery from the Agent(model=...) constructor kwarg to a first-class pydantic-ai capability, ResolvedModel, on the dedicated get_model() seam.

Both construction sites (agents/_builder.py and tools/subagent_invocation.py) previously handed their fully resolved Model instance to the constructor kwarg while everything else about the agent's behavior already composes through capabilities=[...].

What changed

• New code_puppy/agents/_resolved_model.py — @dataclass ResolvedModel(AbstractCapability[Any]) with a single model: Model field, delivering via get_model().
• _builder.build_pydantic_agent drops model= and hoists one shared ResolvedModel instance across the probe + final construction passes (mirrors the single load_model_with_fallback call; safe because the capability is a static snapshot, default for_run returns self, and it carries no id).
• subagent_invocation drops model= and constructs ResolvedModel(model) inline.
• Resolution logic untouched (same restraint as Promote per-model settings to a PerModelSettings capability #831/refactor: promote system-prompt delivery to an AssembledInstructions capability #832): load_model_with_fallback still owns the fallback chain + per-conversation warning dedup; sub-agent pin/explicit-override resolution unchanged; BaseAgent.cur_model still set. build_tool_probe_for_agent keeps its model= kwarg deliberately (stripped probe, per the refactor: promote system-prompt delivery to an AssembledInstructions capability #832 precedent for instructions="").

Parity (verified against pydantic-ai 2.31.0 source)

• Static contribution, resolved once per run. A non-callable get_model() return is used directly as the run's model (Agent._evaluate_model_contribution), exactly like the agent-slot model. _check_dynamic_model_resume is a no-op for static contributions — no selector semantics introduced.
• run(model=...) and Agent.override(model=...) stay authoritative. Both set model_is_explicit, which short-circuits capability contribution evaluation entirely — same precedence as the old kwarg. Pinned by tests.
• Enter/exit lifecycle preserved. Agent.__aenter__ enters a static capability model's context identically (line ~3736: static_selection = capability_model), so provider HTTP clients open/close the same.
• One observable divergence: the built agent's .model property now reads None (the model no longer occupies the agent slot). Verified: zero code_puppy readers — everything goes through BaseAgent.cur_model. One plugin reader exists: wiggum's _resolve_judges probes get_pydantic_agent().model.model_name for its default-judge fallback and now takes its own next fallback, BaseAgent.get_model_name() — the configured model name, which is what judge configs resolve against anyway (arguably more correct than the provider model id it used to get). Degrades gracefully by design (getattr(None, "model_name", None)); documented in the module docstring and pinned by test_agent_model_slot_reads_none.
• Not spec-constructible by design: get_serialization_name() -> None because the capability holds a live, provider-configured Model instance (HTTP clients and all) — same precedent as SteerInjection/HistoryCompaction.
• List position inert: get_model is a configuration seam, not a request hook; placed first in both blocks for readability.

Tests

7 contract tests in tests/agents/test_resolved_model_capability.py:

1. direct seam identity (get_model() is model)
2. spec-construction opt-out
3. kwarg-vs-capability wire parity via FunctionModel (model-visible content compared; timestamps excluded)
4. run(model=...) authoritative over the capability
5. override(model=...) authoritative in scope, capability restored outside
6. pinned .model is None divergence
7. end-to-end build_pydantic_agent drive: built agent runs on the capability-delivered model, cur_model tracks it, agent slot empty

Full suite: 7583 passed; the 3 reds are the documented pre-existing main failures (test_render_version_check_current, plus the two full-run-order FieldInfo has no attribute 'strip' flakes noted in #832) — both subagent_invocation-adjacent ones pass in isolation on this branch.

Merge-order note

Like its five siblings, this touches the shared capabilities=[...] blocks in _builder.py + subagent_invocation.py. Whichever PRs land last eat trivial rebases.

[mpfaffenberger]

Review trail: code-puppy clone did two passes. Pass 1: clean APPROVE with three optional test-hardening findings (pin the Agent.aenter/aexit lifecycle of the capability model, prove static delivery across a multi-step tool-call run, cover the sub-agent construction site end-to-end). All three applied in c511377. Pass 2: APPROVE, confirmed the new tests are non-vacuous (lifecycle counters, forced two-step run, subagent test fails if the capability is dropped). 10/10 contract tests green; full suite matches main's failure profile (3 documented pre-existing reds, all reproduce on clean main).