Drop the ACP layer

Follow-up to opencode-serve-migration.md. Two phases — Phase 1 (transport) and Phase 2 (vocabulary cleanup) shipped together. Phase 3 (inline the schema types + drop the PyPI dep) is deferred indefinitely behind a wrapper.

What shipped

The opencode acp transport is gone. OpencodeServeClient is the only path behind KubernetesSandboxManager.send_message and DockerSandboxManager.send_message. AGENT_TRANSPORT env var, the AgentTransport enum, the per-message ACP exec clients (ACPExecClient, DockerACPExecClient), ACPExecClientBase, and the entrypoint idle branch are deleted. Per-session opencode.json writes are gone — provider config lives at pod/container scope via OPENCODE_CONFIG_CONTENT. Port 8081 (AGENT_PORT) cleared from the pod/service/Dockerfile.

The "ACP" vocabulary is gone from production code, tests, deployment, and non-historical docs. Internal symbols renamed: ACPEvent → SandboxEvent, _yield_acp_events → _yield_sandbox_events, _persist_acp_event → _persist_sandbox_event, [ACP-EVENT] → [SANDBOX-EVENT], ACPError → SandboxError, _merge_acp_with_announces → _merge_events_with_announces. Config: ACP_MESSAGE_TIMEOUT → SANDBOX_TURN_TIMEOUT_SECONDS. Frontend: ACPEvent → SandboxEvent, ACPErrorEvent → SandboxErrorEvent, ACPBaseEvent → SandboxEventBase, ACPErrorPacket → SandboxErrorPacket. Every from acp.schema import … now routes through backend/onyx/server/features/build/sandbox/event_schema.py — a thin re-export wrapper, the only place acp.schema is mentioned in the tree.

What's deferred

Inlining the schema types into Onyx-owned Pydantic models and dropping the agent-client-protocol>=0.7.1 PyPI dep. The wrapper makes this a single-file change — replace the re-exports in event_schema.py with local Pydantic definitions, drop the dep, done.

Why deferred:

• The wrapper hides the dep from every consumer. Nothing reads "ACP" anywhere in the codebase, so the dep doesn't leak conceptually.
• Inlining means copying ~9 Pydantic models field-for-field with byte-identical model_dump(by_alias=True) output. Tractable but a real code review burden.
• The schema is the abstraction boundary for a future in-house agent harness. When that work starts and we need to evolve fields the upstream doesn't ship, that's the trigger — replace the wrapper's contents, drop the dep, evolve freely.

DB-shape consumers worth noting before any inline

If/when you do swap the wrapper for inlined types:

• tool_call_progress and agent_plan_update rows in BuildMessage.message_metadata are model_dump(by_alias=True) JSON. Old rows have whatever aliases the upstream package shipped; new rows would have whatever the inlined types ship. The frontend already defensively reads both raw_input/rawInput and tool_call_id/toolCallId in packetTypes.ts, so minor drift is survivable — but identical aliases are the safe play.
• All other persisted message types (user_message, agent_message, agent_thought, task output) are custom-shaped synthesis, not raw schema output. Orthogonal to the inline.

Reference plan (when Phase 3 becomes worth doing)

1. Inline the types. In sandbox/event_schema.py, replace the from acp.schema import … lines with local Pydantic v2 model definitions. Same field names, same validators, same model_dump(by_alias=True) output. Add a round-trip test that asserts the JSON shape matches a canned reference per type — locks the wire contract.

2. Drop the dep. Remove agent-client-protocol>=0.7.1 from pyproject.toml. Re-lock with uv sync.

3. Sweep ee/. Enterprise Edition code might re-import acp.schema directly. Grep before merging.

Risks specific to Phase 3

• Pydantic field-shape divergence. Easy to miss a Field(alias=…) or model_config = ConfigDict(populate_by_name=True). The round-trip test catches the structural case; manual Playwright pass on at least one Craft session catches the wire-format case.
• DB row-shape divergence for tool_call_progress and agent_plan_update. Keep aliases byte-identical or backfill.

Out of scope even for Phase 3

• Replacing the inlined types with opencode-native types on the SSE wire.
• Renaming the class names themselves (AgentMessageChunk → AssistantTextDelta etc.) — cosmetic, touches every consumer.