Request
Please expose the app-server turn ID and its start acknowledgment through the ACP interface, correlated with the originating session/prompt submission.
In installed @agentclientprotocol/codex-acp 1.10.0:
CodexAppServerClient.runTurn() receives turnStarted.turn.id from turn/start, calls onTurnStarted, then awaits completion.CodexAcpAgent.prompt() retains that ID in activePrompt.currentTurn and sessionState.currentTurnId.CodexEventHandler handles turn/started by retaining the ID and flushing pending errors, but returns no ACP update for that event.
This is an exposure request for information already received and retained internally, not a request to add provider plumbing. A supported, opt-in ACP extension notification or equivalent metadata with submission correlation would help clients retain this evidence before text/tool output arrives. It should not manufacture visible model output or require changing ACP v1 completion semantics.
Why
We are building durable per-submission evidence in Seam: jbulpitt/seam-acp#536. When a prompt fails before the client observes output, the client cannot tell whether the attempt reached the agent. Replaying the original prompt can duplicate work; continuing can omit a task that never arrived. We are recording explicit unknown rather than changing the retry default from absence of output.
Scope of the evidence
The signal must be labeled app-server turn started, not provider acceptance, billing, completion, or proof of tool side effects. We understand that a lost acknowledgment remains uncertain and that a turn ID is not an idempotency key. Documentation of this distinction, including any replay/correlation guarantees, would be valuable.
No prompt text, tool arguments, credentials, or output bodies are needed. We have not patched the dependency locally.
Request
Please expose the app-server turn ID and its start acknowledgment through the ACP interface, correlated with the originating
session/promptsubmission.In installed
@agentclientprotocol/codex-acp1.10.0:CodexAppServerClient.runTurn()receivesturnStarted.turn.idfromturn/start, callsonTurnStarted, then awaits completion.CodexAcpAgent.prompt()retains that ID inactivePrompt.currentTurnandsessionState.currentTurnId.CodexEventHandlerhandlesturn/startedby retaining the ID and flushing pending errors, but returns no ACP update for that event.This is an exposure request for information already received and retained internally, not a request to add provider plumbing. A supported, opt-in ACP extension notification or equivalent metadata with submission correlation would help clients retain this evidence before text/tool output arrives. It should not manufacture visible model output or require changing ACP v1 completion semantics.
Why
We are building durable per-submission evidence in Seam: jbulpitt/seam-acp#536. When a prompt fails before the client observes output, the client cannot tell whether the attempt reached the agent. Replaying the original prompt can duplicate work; continuing can omit a task that never arrived. We are recording explicit
unknownrather than changing the retry default from absence of output.Scope of the evidence
The signal must be labeled app-server turn started, not provider acceptance, billing, completion, or proof of tool side effects. We understand that a lost acknowledgment remains uncertain and that a turn ID is not an idempotency key. Documentation of this distinction, including any replay/correlation guarantees, would be valuable.
No prompt text, tool arguments, credentials, or output bodies are needed. We have not patched the dependency locally.