enhance: align media modes and consolidate output negotiation

src/opencode_a2a/execution/executor.py

@@ -34,6 +34,7 @@
     UpstreamConcurrencyLimitError,
     UpstreamContractError,
 )
+from ..output_modes import accepts_output_mode, normalize_accepted_output_modes
 from ..parts.mapping import (
     UnsupportedA2AInputError,
     extract_text_from_a2a_parts,
@@ -90,6 +91,8 @@
 )
 
 logger = logging.getLogger(__name__)
+_TEXT_PLAIN_MEDIA_TYPE = "text/plain"
+_APPLICATION_JSON_MEDIA_TYPE = "application/json"
 
 __all__ = [
     "_build_output_metadata",
@@ -148,6 +151,7 @@ class _PreparedExecution:
     directory: str | None
     workspace_id: str | None
     session_binding_context_id: str
+    allow_structured_output: bool
 
 
 def _build_session_binding_context_id(
@@ -368,6 +372,7 @@ async def _bind_session(self) -> None:
                     directory=self._prepared.directory,
                     workspace_id=self._prepared.workspace_id,
                     terminal_signal=self._stream_terminal_signal,
+                    allow_structured_output=self._prepared.allow_structured_output,
                 )
             )
 
@@ -775,6 +780,7 @@ async def execute(self, context: RequestContext, event_queue: EventQueue) -> Non
         identity = (call_context.state.get("identity") if call_context else None) or "anonymous"
 
         streaming_request = self._should_stream(context)
+        accepted_output_modes = normalize_accepted_output_modes(context.configuration)
         message_parts = (
             getattr(context.message, "parts", None) if context.message is not None else None
         )
@@ -853,6 +859,22 @@ async def execute(self, context: RequestContext, event_queue: EventQueue) -> Non
             )
             return
 
+        if not accepts_output_mode(accepted_output_modes, _TEXT_PLAIN_MEDIA_TYPE):
+            await self._emit_error(
+                event_queue,
+                task_id=task_id,
+                context_id=context_id,
+                message="acceptedOutputModes must include text/plain for OpenCode chat responses.",
+                state=TaskState.failed,
+                streaming_request=streaming_request,
+            )
+            return
+
+        allow_structured_output = accepts_output_mode(
+            accepted_output_modes,
+            _APPLICATION_JSON_MEDIA_TYPE,
+        )
+
         logger.debug(
             (
                 "Received message identity=%s task_id=%s context_id=%s "
@@ -877,6 +899,7 @@ async def execute(self, context: RequestContext, event_queue: EventQueue) -> Non
             directory=directory,
             workspace_id=workspace_id,
             session_binding_context_id=session_binding_context_id,
+            allow_structured_output=allow_structured_output,
         )
         coordinator = _ExecutionCoordinator(
             self,
@@ -1097,6 +1120,7 @@ async def _consume_opencode_stream(
         terminal_signal: asyncio.Future[_StreamTerminalSignal],
         directory: str | None = None,
         workspace_id: str | None = None,
+        allow_structured_output: bool = True,
     ) -> None:
         await self._stream_runtime.consume(
             session_id=session_id,
@@ -1110,6 +1134,7 @@ async def _consume_opencode_stream(
             terminal_signal=terminal_signal,
             directory=directory,
             workspace_id=workspace_id,
+            allow_structured_output=allow_structured_output,
         )
 
 

src/opencode_a2a/execution/stream_runtime.py

@@ -77,6 +77,7 @@ async def consume(
         terminal_signal: asyncio.Future[_StreamTerminalSignal],
         directory: str | None = None,
         workspace_id: str | None = None,
+        allow_structured_output: bool = True,
     ) -> None:
         part_states: dict[str, _StreamPartState] = {}
         pending_deltas: defaultdict[str, list[_PendingDelta]] = defaultdict(list)
@@ -85,6 +86,8 @@ async def consume(
 
         async def _emit_chunks(chunks: list[_NormalizedStreamChunk]) -> None:
             for chunk in chunks:
+                if not allow_structured_output and getattr(chunk.part.root, "kind", None) == "data":
+                    continue
                 resolved_message_id = stream_state.resolve_message_id(chunk.message_id)
                 chunk_text = getattr(chunk.part.root, "text", "")
                 if stream_state.should_drop_initial_user_echo(

src/opencode_a2a/output_modes.py

@@ -0,0 +1,29 @@
+from __future__ import annotations
+
+from collections.abc import Collection
+from typing import Any
+
+
+def normalize_accepted_output_modes(source: Any) -> tuple[str, ...] | None:
+    accepted = getattr(source, "accepted_output_modes", None) or getattr(
+        source, "acceptedOutputModes", None
+    )
+    if not isinstance(accepted, list):
+        return None
+
+    normalized: list[str] = []
+    for value in accepted:
+        if not isinstance(value, str):
+            continue
+        mode = value.strip().lower()
+        if not mode or mode in normalized:
+            continue
+        normalized.append(mode)
+    return tuple(normalized) or None
+
+
+def accepts_output_mode(
+    accepted_output_modes: Collection[str] | None,
+    media_type: str,
+) -> bool:
+    return accepted_output_modes is None or media_type in accepted_output_modes

src/opencode_a2a/server/agent_card.py

@@ -42,6 +42,10 @@
 from ..jsonrpc.application import SESSION_CONTEXT_PREFIX
 from ..profile.runtime import RuntimeProfile, build_runtime_profile
 
+_CHAT_INPUT_MODES = ["text/plain", "application/octet-stream"]
+_CHAT_OUTPUT_MODES = ["text/plain", "application/json"]
+_JSON_RPC_MODES = ["application/json"]
+
 
 def _select_public_extension_params(
     params: dict[str, Any],
@@ -381,6 +385,8 @@ def _build_agent_skills(
                     "Handle core A2A chat turns with shared session binding and optional "
                     "request-scoped model selection."
                 ),
+                input_modes=list(_CHAT_INPUT_MODES),
+                output_modes=list(_CHAT_OUTPUT_MODES),
                 tags=["assistant", "coding", "opencode", "core-a2a", "portable"],
             ),
             AgentSkill(
@@ -390,6 +396,8 @@ def _build_agent_skills(
                     "Inspect OpenCode session status, history, and low-risk lifecycle actions "
                     "through provider-private JSON-RPC extensions."
                 ),
+                input_modes=list(_JSON_RPC_MODES),
+                output_modes=list(_JSON_RPC_MODES),
                 tags=["opencode", "sessions", "history", "provider-private"],
             ),
             AgentSkill(
@@ -399,6 +407,8 @@ def _build_agent_skills(
                     "Discover available upstream providers and models through provider-private "
                     "JSON-RPC extensions."
                 ),
+                input_modes=list(_JSON_RPC_MODES),
+                output_modes=list(_JSON_RPC_MODES),
                 tags=["opencode", "providers", "models", "provider-private"],
             ),
             AgentSkill(
@@ -408,6 +418,8 @@ def _build_agent_skills(
                     "Manage OpenCode projects, workspaces, and worktrees through "
                     "provider-private JSON-RPC extensions."
                 ),
+                input_modes=list(_JSON_RPC_MODES),
+                output_modes=list(_JSON_RPC_MODES),
                 tags=["opencode", "project", "workspace", "worktree", "provider-private"],
             ),
             AgentSkill(
@@ -417,6 +429,8 @@ def _build_agent_skills(
                     "Recover pending permission and question interrupts through "
                     "provider-private JSON-RPC extensions."
                 ),
+                input_modes=list(_JSON_RPC_MODES),
+                output_modes=list(_JSON_RPC_MODES),
                 tags=["interrupt", "permission", "question", "provider-private"],
             ),
             AgentSkill(
@@ -426,6 +440,8 @@ def _build_agent_skills(
                     "Reply to streaming permission and question interrupts through shared "
                     "JSON-RPC callbacks."
                 ),
+                input_modes=list(_JSON_RPC_MODES),
+                output_modes=list(_JSON_RPC_MODES),
                 tags=["interrupt", "permission", "question", "shared"],
             ),
         ]
@@ -439,6 +455,8 @@ def _build_agent_skills(
                 "TextPart and FilePart inputs to OpenCode sessions with shared session "
                 "binding and optional request-scoped model selection."
             ),
+            input_modes=list(_CHAT_INPUT_MODES),
+            output_modes=list(_CHAT_OUTPUT_MODES),
             tags=["assistant", "coding", "opencode", "core-a2a", "portable"],
             examples=_build_chat_examples(settings.a2a_project),
         ),
@@ -449,6 +467,8 @@ def _build_agent_skills(
                 "provider-private OpenCode session/history and session-control surface "
                 "exposed through JSON-RPC extensions."
             ),
+            input_modes=list(_JSON_RPC_MODES),
+            output_modes=list(_JSON_RPC_MODES),
             tags=["opencode", "sessions", "history", "provider-private"],
             examples=_build_session_query_skill_examples(
                 capability_snapshot=capability_snapshot,
@@ -461,6 +481,8 @@ def _build_agent_skills(
                 "provider-private OpenCode provider/model discovery surface exposed "
                 "through JSON-RPC extensions."
             ),
+            input_modes=list(_JSON_RPC_MODES),
+            output_modes=list(_JSON_RPC_MODES),
             tags=["opencode", "providers", "models", "provider-private"],
             examples=[
                 "List available providers (method opencode.providers.list).",
@@ -474,6 +496,8 @@ def _build_agent_skills(
                 "provider-private OpenCode project/workspace/worktree control surface "
                 "exposed through JSON-RPC extensions."
             ),
+            input_modes=list(_JSON_RPC_MODES),
+            output_modes=list(_JSON_RPC_MODES),
             tags=["opencode", "project", "workspace", "worktree", "provider-private"],
             examples=_build_workspace_control_skill_examples(),
         ),
@@ -484,6 +508,8 @@ def _build_agent_skills(
                 "provider-private OpenCode interrupt recovery surface exposed through "
                 "JSON-RPC extensions."
             ),
+            input_modes=list(_JSON_RPC_MODES),
+            output_modes=list(_JSON_RPC_MODES),
             tags=["interrupt", "permission", "question", "provider-private"],
             examples=_build_interrupt_recovery_skill_examples(),
         ),
@@ -495,6 +521,8 @@ def _build_agent_skills(
                 "JSON-RPC methods a2a.interrupt.permission.reply, "
                 "a2a.interrupt.question.reply, and a2a.interrupt.question.reject."
             ),
+            input_modes=list(_JSON_RPC_MODES),
+            output_modes=list(_JSON_RPC_MODES),
             tags=["interrupt", "permission", "question", "shared"],
             examples=[
                 "Reply once/always/reject to a permission request by request_id.",
@@ -535,8 +563,8 @@ def _build_agent_card(
         version=settings.a2a_version,
         protocol_version=settings.a2a_protocol_version,
         preferred_transport=TransportProtocol.http_json,
-        default_input_modes=["text/plain", "application/octet-stream"],
-        default_output_modes=["text/plain"],
+        default_input_modes=list(_CHAT_INPUT_MODES),
+        default_output_modes=list(_CHAT_OUTPUT_MODES),
         capabilities=AgentCapabilities(
             streaming=True,
             extensions=_build_agent_extensions(