Summary

isCliSessionExpiredErrorMessage() in pi-embedded-helpers does not match Claude CLI's "No conversation found with session ID: ..." error. The existing recovery logic (clear stale session + retry fresh) never triggers, causing an infinite retry loop with the same dead session ID across all fallback models.

Root Cause

The pattern list in isCliSessionExpiredErrorMessage() includes "conversation not found" but not "no conversation found":

// pi-embedded-helpers — isCliSessionExpiredErrorMessage()
lower.includes("conversation not found")  // ← in the list
// but Claude CLI returns:
// "No conversation found with session ID: dfee08ff-..."
// lower: "no conversation found with session id: dfee08ff-..."
//
// "conversation not found" !== "no conversation found"
// ↑ missing "no " prefix — off by one word

Because the pattern doesn't match, classifyFailoverSignal() returns "unknown" instead of "session_expired", and the recovery path at agent-command line ~269 is never entered:

// agent-command — the recovery that SHOULD fire but doesn't:
if (err instanceof FailoverError && err.reason === "session_expired" && ...) {
    clearCliSession(updatedEntry, params.providerOverride);
    return runCliWithSession(void 0); // ← retry with fresh session
}

Observed Behavior

1. A Claude CLI session expires or becomes unreachable (e.g., after context exhaustion, rate limit kills the session, or gateway restart)
2. Gateway tries --resume <stale-id> → Claude CLI returns "No conversation found with session ID: <id>"
3. Pattern doesn't match → reason = "unknown" → no session cleanup
4. Model fallback kicks in, but all fallback models reuse the same stale session ID
5. All 4 models fail → "All models failed (4)" → user sees "Something went wrong"
6. Next message → same stale ID → same failure → permanent loop until manual intervention

Gateway error log shows 30+ consecutive failures over hours, all hitting the same dead session ID:

[model-fallback/decision] candidate_failed reason=unknown  ← should be session_expired
Embedded agent failed: All models failed (4):
  claude-cli/claude-opus-4-6: "No conversation found with session ID: 40c32f08-..."
  claude-cli/claude-sonnet-4-6: "No conversation found with session ID: 40c32f08-..."
  claude-cli/sonnet-4.6: "No conversation found with session ID: 40c32f08-..."
  claude-cli/sonnet-4.5: "No conversation found with session ID: 40c32f08-..."

Fix

Add "no conversation found" to the pattern list:

 function isCliSessionExpiredErrorMessage(raw) {
     if (!raw) return false;
     const lower = raw.toLowerCase();
     return lower.includes("session not found") ||
            lower.includes("session does not exist") ||
            lower.includes("session expired") ||
            lower.includes("session invalid") ||
            lower.includes("conversation not found") ||
+           lower.includes("no conversation found") ||
            lower.includes("conversation does not exist") ||
            ...
 }

This enables the existing recovery in agent-command to fire: clear the stale cliSessionIds/claudeCliSessionId from the session store entry and retry with a fresh session.

Environment

• OpenClaw: 2026.4.2 (d74a122)
• Claude Code CLI: 2.1.92
• Provider: claude-cli (OAuth, firstParty)
• OS: macOS Darwin 25.3.0

Workaround

Manually patch pi-embedded-helpers-*.js in the installed openclaw dist to add the missing pattern, then restart the gateway.