…ins SystemPromptParts

Include next_message in the system prompt existence check so that when
message_history ends with a ModelRequest that already has SystemPromptParts,
fresh ones aren't injected again.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

…uests, not just the first

The code iterates over every ModelRequest in the history and skips injection
if any SystemPromptPart is present anywhere, but the doc copy said 'first
request'. Align the wording with the actual behavior.

Previous approaches to UIAdapter's 'server mode' system-prompt reinjection
were architecturally unsatisfying: a ContextVar with mutable state (subtle
leakage into nested tool agents), then generalized auto-injection in
UserPromptNode (silently changed the contract of agent.run(message_history=[...])
that the docs explicitly documented).

Shift to a capability-based design:

- New `ReinjectSystemPrompt` capability in `pydantic_ai.capabilities`, hooked
  on `before_model_request`. If no SystemPromptPart is present in history,
  it resolves the agent's configured system_prompt and prepends it at the
  head of the first ModelRequest. No-op if any sys_prompt is already there.
- New public `AbstractAgent.system_prompts(...)` method that resolves
  configured static strings and runners (static + dynamic) into
  SystemPromptParts, using a minimal RunContext built from kwargs. Reusable
  by anyone who needs to produce the same parts outside an agent run.
- New `capabilities=` run-time kwarg on `Agent.run` / `iter` / `run_stream`
  / `run_stream_events` — additive with the agent's configured capabilities,
  matching the `toolsets=` / `builtin_tools=` pattern.
- UIAdapter in `manage_system_prompt='server'` mode adds ReinjectSystemPrompt
  via the new kwarg. Client mode is unchanged (agent's configured sys_prompt
  is not injected; add the capability explicitly for fallback behavior).
- Shared `resolve_system_prompts` helper moved into `_system_prompt.py`;
  `UserPromptNode._sys_parts` delegates to it, removing duplication.
- UserPromptNode's own injection logic is restored to the pre-PR
  behavior (only injects on the first turn of a fresh conversation) — no
  surprise behavior change for `agent.run(message_history=[...])` callers.

A `# TODO (v2):` on the capability notes the option of enabling it by
default in v2 to close #1646 — but as a deliberate breaking change, not
smuggled in.

…snapshot

Two small CI fixes for the ReinjectSystemPrompt capability work:

1. The `capabilities=` run-time parameter was added to `AbstractAgent`'s
   run/run_sync/run_stream/run_stream_events/iter overloads, but not
   propagated to the DBOS/Prefect/Temporal durable_exec agent overrides.
   Pyright flagged these as incompatible overrides.

2. `ReinjectSystemPrompt.get_serialization_name` returned snake_case, which
   broke the `CAPABILITY_TYPES` snapshot and deviated from the other
   capabilities' PascalCase convention. Remove the override so the default
   (class name) is used.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

When an agent is constructed without a configured model and receives it at
run time via `agent.run(model=...)`, `ReinjectSystemPrompt.before_model_request`
needs to call `agent.system_prompts()` to resolve the prompts. Previously
`system_prompts()` unconditionally called `self._get_model(None)` to build a
`RunContext`, which raised `UserError` in this case — so UI adapters in the
default `manage_system_prompt='server'` mode broke for model-less agents.

Accept an optional `model` parameter on `system_prompts()` and forward
`ctx.model` from the capability's `before_model_request` hook. The parameter
propagates through `AbstractAgent` → `Agent` → `WrapperAgent`.

Caught by Devin review on PR #4087.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

- Drop `system_prompt='Be fun!'` from the `'client'` mode examples in
  `docs/ui/ag-ui.md` and `docs/ui/vercel-ai.md`. In client mode the agent's
  configured `system_prompt` is intentionally not injected, so the example
  was silently demonstrating an inert setting. The prose above already covers
  the `ReinjectSystemPrompt` fallback pattern for callers who want it.
- Link `SystemPromptPart` and `ModelRequest` to their API reference pages in
  the `ReinjectSystemPrompt` section of `docs/capabilities.md` and the two
  UI docs, per `docs/CLAUDE.md` guidance.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

…tSystemPrompt

`request_context.messages` is a shallow copy of `ctx.state.message_history`,
so `ModelRequest` objects are shared between the capability's view and the
user's input `message_history=[...]` list. The previous implementation
mutated `msg.parts = [...]` on shared objects, leaking edits back into the
caller's list (they'd see `SystemPromptPart`s stripped from their own
`ModelRequest` instances).

Use `dataclasses.replace` to create new `ModelRequest` objects when we need
to change their parts, and swap them into the copy list without touching
the originals. Adds a regression test exercising `replace_existing=True`
with a caller-owned `ModelRequest` containing a `SystemPromptPart`.

Caught by Devin review on PR #4087.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The new `AbstractAgent` method returns `list[SystemPromptPart]`, not a list
of strings. `system_prompt_parts` makes that explicit and leaves the
unqualified `system_prompt` name free for future additions (for example, a
matching `instruction_parts()` that resolves the agent's configured
[`instructions`][pydantic_ai.Agent.instructions] into structured parts —
the `@agent.instructions` decorator already occupies the singular name).

Also adds `test_reinject_system_prompt_capability_reaches_model_and_all_messages`
using `FunctionModel` to capture what the model actually received, asserting
both that the injected `SystemPromptPart` reaches the model AND that it
persists in `result.all_messages()`. This directly verifies that the hook's
mutations to `request_context.messages` are synced back to canonical state
via `_agent_graph.py:835`, and addresses the ambiguity flagged by Devin.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>