pydantic-acp 0.9.7

ACP adapter for pydantic-ai agents.

pydantic-acp

pydantic-acp adapts pydantic_ai.Agent instances to the ACP agent interface without rewriting the underlying agent.

The core contract is simple:

1. keep the existing pydantic_ai.Agent
2. expose it through ACP
3. only publish ACP-visible state the runtime can actually honor

Entry Points

• run_acp(...)
• create_acp_agent(...)
• AdapterConfig
• AcpSessionContext
• StaticAgentSource
• FactoryAgentSource
• MemorySessionStore
• FileSessionStore

What It Covers

pydantic-acp includes:

• ACP session lifecycle, replay, resume, and persistence
• session-local model selection
• mode and slash-command control
• native ACP plan state with structured TaskPlan
• approval bridging
• prompt resources including files, embedded resources, images, and audio
• projection maps for filesystem, hooks, web tools, and builtin tool families
• capability bridges for upstream Pydantic AI capabilities
• client-backed filesystem and terminal helpers

Quick Start

from pydantic_ai import Agent
from pydantic_acp import run_acp

agent = Agent("openai:gpt-5", name="demo-agent")
run_acp(agent=agent)

If another runtime should own transport lifecycle:

from acp import run_agent
from pydantic_ai import Agent
from pydantic_acp import AdapterConfig, MemorySessionStore, create_acp_agent

agent = Agent("openai:gpt-5", name="composable-agent")

acp_agent = create_acp_agent(
    agent=agent,
    config=AdapterConfig(session_store=MemorySessionStore()),
)

run_agent(acp_agent)

If you are using Codex-backed Pydantic models through codex-auth-helper, pass explicit instructions when building the model. That is the preferred seam for Codex-specific system behavior:

from codex_auth_helper import create_codex_responses_model
from pydantic_ai import Agent

model = create_codex_responses_model(
    "gpt-5.4",
    instructions="You are a careful coding assistant.",
)
agent = Agent(model, name="codex-agent")

On the Pydantic path, Agent(instructions=...) is also valid and may still be useful for agent-specific behavior layered on top of the model:

from codex_auth_helper import create_codex_responses_model
from pydantic_ai import Agent

model = create_codex_responses_model(
    "gpt-5.4",
    instructions="You are a careful coding assistant.",
)
agent = Agent(
    model,
    name="codex-agent",
    instructions="Ask for clarification when the task is underspecified.",
)

In short: Codex-backed Pydantic models should not rely on an implicit default instruction string. Set instructions explicitly at the factory level, and add Agent(instructions=...) when you want extra agent-owned behavior.

Native Plan Mode

TaskPlan is the structured native plan output surface.

Use PrepareToolsBridge to expose plan mode:

from pydantic_ai import Agent
from pydantic_ai.tools import RunContext, ToolDefinition
from pydantic_acp import (
    AdapterConfig,
    PrepareToolsBridge,
    PrepareToolsMode,
    run_acp,
)


def read_only_tools(
    ctx: RunContext[None],
    tool_defs: list[ToolDefinition],
) -> list[ToolDefinition]:
    del ctx
    return list(tool_defs)


agent = Agent("openai:gpt-5", name="plan-agent")

run_acp(
    agent=agent,
    config=AdapterConfig(
        capability_bridges=[
            PrepareToolsBridge(
                default_mode_id="plan",
                default_plan_generation_type="structured",
                modes=[
                    PrepareToolsMode(
                        id="plan",
                        name="Plan",
                        description="Return a structured ACP task plan.",
                        prepare_func=read_only_tools,
                        plan_mode=True,
                    ),
                ],
            ),
        ],
    ),
)

Important behavior:

• plan_generation_type="structured" is the default plan-mode behavior
• structured mode expects structured TaskPlan output instead of exposing acp_set_plan
• switch to plan_generation_type="tools" when you explicitly want tool-based native plan recording
• keep plan_tools=True when you also want progress tools such as acp_update_plan_entry

Projection Maps

Projection maps decide how known tool families render into ACP-visible updates.

Built-in projection helpers:

• FileSystemProjectionMap
• HookProjectionMap
• WebToolProjectionMap
• BuiltinToolProjectionMap

Example:

from pydantic_acp import (
    AdapterConfig,
    BuiltinToolProjectionMap,
    FileSystemProjectionMap,
    HookProjectionMap,
    run_acp,
)

run_acp(
    agent=agent,
    config=AdapterConfig(
        projection_maps=[
            FileSystemProjectionMap(
                default_read_tool="read_file",
                default_write_tool="write_file",
            ),
            HookProjectionMap(
                hidden_event_ids=frozenset({"after_model_request"}),
                event_labels={"before_model_request": "Preparing Request"},
            ),
            BuiltinToolProjectionMap(),
        ],
    ),
)

Capability Bridges

Current built-in bridges include:

• ThinkingBridge
• PrepareToolsBridge
• ThreadExecutorBridge
• SetToolMetadataBridge
• IncludeToolReturnSchemasBridge
• WebSearchBridge
• WebFetchBridge
• ImageGenerationBridge
• McpCapabilityBridge
• ToolsetBridge
• PrefixToolsBridge
• OpenAICompactionBridge
• AnthropicCompactionBridge

Use bridges when the runtime should gain upstream Pydantic AI capabilities and ACP-visible metadata without rewriting the adapter core.

Harness-backed Capabilities

pydantic-acp also ships a maintained bridge and projection layer for pydantic-ai-harness.

Public seams:

• HarnessFileSystemBridge
• HarnessShellBridge
• HarnessCodeModeBridge
• HarnessFileSystemProjectionMap
• HarnessShellProjectionMap
• HarnessCodeModeProjectionMap

Minimal example:

from pathlib import Path

from pydantic_ai import Agent
from pydantic_acp import (
    AdapterConfig,
    HarnessFileSystemBridge,
    HarnessFileSystemProjectionMap,
    HarnessShellBridge,
    HarnessShellProjectionMap,
    MemorySessionStore,
    run_acp,
)

workspace_root = Path(".harness-agent")

agent = Agent(
    "openai:gpt-5",
    name="harness-agent",
    instructions="Use the harness filesystem and shell tools inside the workspace only.",
)

run_acp(
    agent=agent,
    config=AdapterConfig(
        session_store=MemorySessionStore(),
        capability_bridges=[
            HarnessFileSystemBridge(root_dir=workspace_root),
            HarnessShellBridge(cwd=workspace_root),
        ],
        projection_maps=[
            HarnessFileSystemProjectionMap(),
            HarnessShellProjectionMap(),
        ],
    ),
)

Use HarnessCodeModeBridge only when the run should expose CodeMode. The maintained example keeps that bridge opt-in so the native ACP target stays limited to filesystem and shell by default:

• example source
• detailed guide

The harness filesystem projection now renders read_file as a read-specific preview instead of a fake diff, which makes ACP transcript output much more truthful for inspection-only tool calls.

Factories, Sources, And Host-owned State

Use agent_factory= when the ACP session should influence which agent gets built:

from pydantic_ai import Agent
from pydantic_acp import AcpSessionContext, AdapterConfig, MemorySessionStore, run_acp


def build_agent(session: AcpSessionContext) -> Agent[None, str]:
    workspace_name = session.cwd.name
    model_name = "openai:gpt-5.4-mini"
    if workspace_name.endswith("-deep"):
        model_name = "openai:gpt-5.4"
    return Agent(model_name, name=f"workspace-{workspace_name}")


run_acp(
    agent_factory=build_agent,
    config=AdapterConfig(session_store=MemorySessionStore()),
)

Use AgentSource when the agent and its dependencies should be built separately. Use providers when models, modes, config values, plans, or approvals belong to the host layer instead of the adapter.

Session Store Notes

Use MemorySessionStore for ephemeral local runs and FileSessionStore when ACP sessions should survive process restarts. FileSessionStore is a local durable store, not a distributed coordination layer.

File-backed session ids are constrained before they become filenames:

• allowed characters are ASCII letters, digits, _, and -
• maximum length is 128 characters
• path separators, dot-prefixed ids, whitespace, and shell metacharacters are rejected

The file store writes through a temp file, fsync, and atomic replace. Malformed or partially written session files are skipped by public load/list flows.

Maintained Examples

Maintained runnable examples:

• finance_agent.py
• travel_agent.py

Focused docs recipes:

• Dynamic Factory Agents

Documentation

• Pydantic ACP Overview
• AdapterConfig
• Plans, Thinking, and Approvals
• Models, Modes, and Slash Commands
• Prompt Resources and Context
• Session State and Lifecycle
• Bridges
• Providers
• Security Guidance
• Host Backends and Projections
• API Reference

Compatibility Policy

pydantic-acp currently pins pydantic-ai-slim==1.106.0.

That pin is deliberate. The adapter is tested against a specific Pydantic AI surface and should still be upgraded deliberately, but the hook-compatibility seam is isolated behind ACP Kit’s own compatibility layer instead of scattering private upstream imports through the runtime. The 1.92 surface includes function-tool preparation, output-tool preparation, output validation/processing hooks, deferred-tool-call hooks, run metadata, and conversation IDs.