Return session ID in NewSessionResponse _meta for ACP session identity

[t2wei]

Problem

When codex-acp creates a new session via new_session(), the Codex internal thread ID is used as the ACP sessionId in NewSessionResponse::new(session_id). However, ACP session managers like acpx extract the agent's internal session identifier from _meta.agentSessionId or _meta.sessionId — not from the top-level sessionId field.

This means the agent's internal thread ID is available but not surfaced through _meta, so ACP orchestrators cannot link back to the Codex session logs on disk.

Context

The ACP session identity model (as implemented by acpx v0.1.15) defines three distinct IDs:

Field	Semantics
acpxRecordId	Stable session manager record key
acpSessionId	ACP wire session ID (the sessionId from NewSessionResponse)
agentSessionId	Agent's internal session/thread ID — extracted from _meta

acpx extracts agentSessionId via:

const AGENT_SESSION_ID_META_KEYS = ["agentSessionId", "sessionId"];
function extractAgentSessionId(meta) {
  // looks at result._meta.agentSessionId or result._meta.sessionId
}

Currently codex-acp does not set _meta on NewSessionResponse, so agentSessionId is always undefined.

Proposed Fix

In src/codex_agent.rs, add .meta() to the NewSessionResponse:

// current (line ~371):
Ok(NewSessionResponse::new(session_id)
    .modes(load.modes)
    .models(load.models)
    .config_options(load.config_options))

// proposed:
Ok(NewSessionResponse::new(session_id.clone())
    .meta(serde_json::json!({"sessionId": session_id.0}))
    .modes(load.modes)
    .models(load.models)
    .config_options(load.config_options))

Similarly for load_session().

Use Case

We use OpenClaw + acpx to orchestrate Codex as an ACP sub-agent. With agentSessionId populated, we can:

• Link ACP session records to Codex's native session logs (~/.codex/sessions/)
• Provide end-to-end session debugging and conversation history transparency
• Correlate Codex thread IDs in SQLite with ACP session metadata

Without this, the only linking method is UUIDv7 timestamp heuristics between acpxRecordId and Codex thread.id, which is unreliable in retry scenarios.

Version

Tested against @zed-industries/codex-acp@0.9.5 + acpx@0.1.15.

[benbrandt]

I don't understand... Why can't use just use the standard session id field?