<span id="a2a-方法执行与回执边界"></span>

# A2A methods, execution, and acknowledgement boundaries

The existing public A2A pages receive error corrections only.

<span id="模型与受理"></span>

## Models and acceptance

Ordinary messages resolve the model from explicit params.model, the agent default, or model configuration. If none provides a usable binding, resolution tries the current platform default policy and the authorized agent_chat model catalog. Missing or unavailable default policies and ambiguous same-name backends all fail; omission does not guarantee success. Top-level model does not participate in this parameter selection.

Direct message/send returns after acceptance and does not support configuration.blocking:true. The Go SDK follows submitted/working tasks by default; client NoWait returns only acceptance. CLI --no-wait corresponds to this client behavior and is not sent to the server. Python A2ARequest currently has no no_wait option and follows by default. Its existing general resource_any API can issue raw requests; there is no dedicated Python parameter to invent.

<span id="六个方法"></span>

## Six methods

| Method | Key input | Response and limitations |
| --- | --- | --- |
| message/send | User message, actual messageId/contextId, and model and metadata chosen for the task. | An ordinary task or acknowledgement of a control/feedback/interaction message. |
| message/stream | Same as sending, with SSE enabled. | JSON-RPC frames and heartbeat. Stream termination or disconnection does not mean task success or cancellation. |
| tasks/get | id takes precedence over taskId; afterSeq defaults to 0. | A2A Task projection. afterSeq is passed to event reading rather than fixed at 0. |
| tasks/resubscribe | id/taskId and afterSeq | Subscribe to events. Waiting for input or authorization can end the current stream; continue handling the task. |
| tasks/cancel | id/taskId | Reread after requesting cancellation. metadata.cancellationAccepted:true can coexist with an active state; wait for a terminal state. |
| matrixflow/input.submit | taskId, current requestId, and an answers object | Continue the original task. requestId differs from the outer JSON-RPC id and is not turnId. |

Input answer example:

```json
{"agent_id":"agent-example","jsonrpc":"2.0","id":"rpc-example","method":"matrixflow/input.submit","params":{"taskId":"task-example","requestId":"input-example","answers":{"question-example":{"answers":["实际答案"]}}}}
```

The current Go/Python AgentInputSubmission helpers still construct turnId/callId without the inner requestId. Use verified raw A2A params. The outer RequestID/request_id does not repair the inner field.

<span id="重新生成"></span>

## Regeneration

Only completed source tasks belonging to the current account, agent, workspace, and context can be replayed. The source manifest/input must be readable. Keep controlId fixed. action currently supports only regenerate; rerunMode may be omitted or same_manifest; conversationMode may be omitted, replace_view, or branch. Nonempty override and checkpointId are currently unsupported. Their presence in structures does not establish override support.

branch additionally requires recovery of the original user question and source answer lineage; not every historical task supports branching. guidance supplies modification requirements for regeneration rather than PATCHing saved text. Repeated-control task identifiers depend on the mode: branch uses workspace, context, and controlId; other modes also depend on source task/message identifiers. Preserve the original JSON and related identifiers on retries. Reusing only the outer RPC id is insufficient.

Success may return a Task or a Message containing moi.run.control.result. Obtain the actual new taskId for the same controlId, then verify it with tasks/get. Acceptance, created, and an embedded task label do not establish the actual new task state.

<span id="错误"></span>

## Errors

Both public paths use JSON-RPC after entering the same handler. Authentication, role, or workspace checks can return product errors before that point. Parameters, permissions, missing objects, conflicts, unsupported behavior, dependencies, and internal failures have distinct mappings. Not all tasks/get errors are 404, and the general entry point is not the old forwarding process.

An additional local owner fixture injected a failed execution result and failed stored state, while the control acknowledgement's embedded task.status.state remained fixed at completed; see ../regenerate-receipt-evidence.json. This is a candidate consistency issue, not verified against a real Provider or online deployment. The public guide still requires reading the new task to confirm its state; the fixture is not evidence of a production failure.
