[codex] recover OpenClaw prompt metadata sessions

docs/past-failures.md

@@ -4,7 +4,7 @@
 
 - Symptom: production OpenClaw sessions failed with `ACP_SESSION_INIT_FAILED` and `ACP metadata is missing for agent:main:acp:<id>`, instructing the operator to recreate the ACP session with `/acp spawn`.
 - Cause: the node daemon reused the server session id as the ACPX session name, but a stale ACPX session could exist without the backing ACP metadata OpenClaw needs to initialize.
-- Mitigation: the Go ACPX runner now treats that exact metadata-missing failure as recoverable, recreates the named ACPX session once with `sessions new --name <session>`, and retries `sessions ensure`. A regression test covers the failure, recreate, and retry sequence.
+- Mitigation: the Go ACPX runner now treats that exact metadata-missing failure as recoverable both during `sessions ensure` and during the first prompt run, recreates the named ACPX session once with `sessions new --name <session>`, and retries. Regression tests cover the ensure path and the OpenClaw prompt-time failure shape.
 
 ## 2026-05-12: Installer rejected valid Node 24 runtimes
 

internal/acpx/runner.go

@@ -74,6 +74,13 @@ func (Runner) Run(
 		if err := (Runner{}).Ensure(ctx, request); err != nil {
 			return nil, fmt.Errorf("ensure acpx session: %w", err)
 		}
+		output, err := runStreamingCommand(ctx, command, buildPromptArgs(request), request.WorkingDir, request.Env, request.Stdin, onStdoutLine)
+		if err == nil || request.Session == "" || !isMissingACPMetadataError(err) {
+			return output, err
+		}
+		if newErr := runCommand(ctx, command, buildSessionNewArgs(request), request.WorkingDir, request.Env, ""); newErr != nil {
+			return output, fmt.Errorf("%w; recreate acpx session after missing ACP metadata: %v", err, newErr)
+		}
 		return runStreamingCommand(ctx, command, buildPromptArgs(request), request.WorkingDir, request.Env, request.Stdin, onStdoutLine)
 	}
 
@@ -223,6 +230,7 @@ func runStreamingCommand(
 
 	var (
 		stdoutBuf bytes.Buffer
+		stderrBuf bytes.Buffer
 		mu        sync.Mutex
 		wg        sync.WaitGroup
 	)
@@ -246,16 +254,19 @@ func runStreamingCommand(
 	go func() {
 		defer wg.Done()
 		// Drain stderr so the child process never blocks on a full pipe.
-		// We intentionally discard the contents: acpx stderr is operator noise
-		// (banners, [acpx] tokens summary, http debug) and never carries protocol
-		// state. Surfacing it would let it leak into chat transcripts again.
-		_, _ = io.Copy(io.Discard, stderr)
+		// Keep it out of stdout/chat events, but retain it for errors because
+		// provider init failures such as OpenClaw's ACP metadata error arrive
+		// there.
+		_, _ = io.Copy(&stderrBuf, stderr)
 	}()
 
 	err = cmd.Wait()
 	wg.Wait()
 	output := stdoutBuf.Bytes()
 	if err != nil {
+		if stderrText := strings.TrimSpace(stderrBuf.String()); stderrText != "" {
+			return output, fmt.Errorf("run acpx: %w: %s", err, stderrText)
+		}
 		return output, fmt.Errorf("run acpx: %w", err)
 	}
 	return output, nil

internal/acpx/runner_test.go

@@ -65,6 +65,38 @@ func TestRunnerRecreatesSessionWhenACPMetadataIsMissing(t *testing.T) {
 	}
 }
 
+func TestRunnerRecreatesSessionWhenOpenClawPromptACPMetadataIsMissing(t *testing.T) {
+	t.Parallel()
+
+	statePath := filepath.Join(t.TempDir(), "state")
+	command, args := helperCommand(t, "recover-openclaw-prompt-missing-acp-metadata")
+	output, err := (Runner{}).Run(context.Background(), RunRequest{
+		Command: command,
+		Args:    args,
+		Agent:   "openclaw",
+		Session: "agent-main-acp-d9d29d47",
+		Stdin:   "review this",
+		Env: []string{
+			"GO_WANT_HELPER_PROCESS=1",
+			"ACP_METADATA_RECOVERY_STATE=" + statePath,
+		},
+	}, nil)
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if !strings.Contains(string(output), "prompt-ok") {
+		t.Fatalf("output = %q, want prompt-ok", output)
+	}
+
+	bytes, err := os.ReadFile(statePath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got, want := strings.TrimSpace(string(bytes)), "ensure\nprompt\nnew\nprompt"; got != want {
+		t.Fatalf("commands = %q, want %q", got, want)
+	}
+}
+
 func TestRunnerCancelsProcess(t *testing.T) {
 	t.Parallel()
 
@@ -146,6 +178,8 @@ func TestRunnerHelperProcess(t *testing.T) {
 		os.Exit(0)
 	case "recover-missing-acp-metadata":
 		helperRecoverMissingACPMetadata()
+	case "recover-openclaw-prompt-missing-acp-metadata":
+		helperRecoverOpenClawPromptMissingACPMetadata()
 	default:
 		fmt.Fprintf(os.Stderr, "unknown helper mode %q", mode)
 		os.Exit(2)
@@ -185,6 +219,50 @@ func helperRecoverMissingACPMetadata() {
 	os.Exit(0)
 }
 
+func helperRecoverOpenClawPromptMissingACPMetadata() {
+	statePath := os.Getenv("ACP_METADATA_RECOVERY_STATE")
+	if statePath == "" {
+		fmt.Fprintln(os.Stderr, "missing ACP_METADATA_RECOVERY_STATE")
+		os.Exit(2)
+	}
+
+	args := os.Args
+	command := ""
+	for index := 0; index+2 < len(args); index++ {
+		if args[index] != "openclaw" {
+			continue
+		}
+		switch {
+		case args[index+1] == "sessions":
+			command = args[index+2]
+		case args[index+1] == "--session" || (index+2 < len(args) && args[index+1] == "agent-main-acp-d9d29d47"):
+			command = "prompt"
+		default:
+			command = "prompt"
+		}
+		break
+	}
+	if command == "" {
+		fmt.Fprintf(os.Stderr, "missing openclaw command in %q", args)
+		os.Exit(2)
+	}
+
+	previous, _ := os.ReadFile(statePath)
+	if err := os.WriteFile(statePath, append(previous, []byte(command+"\n")...), 0o600); err != nil {
+		fmt.Fprintf(os.Stderr, "write state: %v", err)
+		os.Exit(2)
+	}
+
+	if command == "prompt" && !strings.Contains(string(previous), "new\n") {
+		fmt.Fprintln(os.Stderr, "ACP error (ACP_SESSION_INIT_FAILED): ACP metadata is missing for agent:main:acp:d9d29d47-c7c5-40d0-9773-e464d4430352. Recreate this ACP session with /acp spawn and rebind the thread. next: If this session is stale, recreate it with /acp spawn and rebind the thread.")
+		os.Exit(1)
+	}
+	if command == "prompt" {
+		fmt.Fprintln(os.Stdout, "prompt-ok")
+	}
+	os.Exit(0)
+}
+
 func helperCommand(t *testing.T, mode string) (string, []string) {
 	t.Helper()