@1presence/bridge 0.34.0 → 0.35.0

package/dist/claude.js

@@ -111,15 +111,32 @@ function renderHistoryMessage(msg) {
 }
 // ─── Active processes ─────────────────────────────────────────────────────────
 const active = new Map();
-// Maximum automatic retries when the `claude` CLI exits non-zero BEFORE
-// producing any real output. This covers the known Claude Code print-mode 400
-// regression that surfaces as "API Error: 400 due to tool use concurrency
-// issues" (GitHub anthropics/claude-code#18131, still open) — it is
-// non-deterministic enough that a fresh spawn frequently succeeds. We retry
-// ONLY when the failed attempt produced no real assistant text and no tool
-// calls, so a failure that lands after real work (where retrying could
-// double-execute a side-effectful tool) is surfaced, never silently re-run.
-const MAX_TURN_RETRIES = 1;
+// conversationId → pending retry timer. A retry is scheduled with a backoff
+// delay, during which the conversation has NO entry in `active`. If a new user
+// message arrives in that window it must cancel the stale retry (otherwise the
+// retry would re-run the OLD turn's history and clobber the new one). The
+// supersede block clears any pending timer here before spawning.
+const pendingRetries = new Map();
+// Automatic retries when the `claude` CLI exits non-zero BEFORE producing any
+// real output. This covers the known Claude Code print-mode 400 regression that
+// surfaces as "API Error: 400 due to tool use concurrency issues" (GitHub
+// anthropics/claude-code#18131, still open) — it is non-deterministic enough
+// that a fresh spawn often succeeds. We retry ONLY when the failed attempt
+// produced no real assistant text and no tool calls, so a failure that lands
+// after real work (where retrying could double-execute a side-effectful tool)
+// is surfaced, never silently re-run.
+//
+// 2 retries = up to 3 attempts/turn. The first retry captures nearly all of the
+// transient wins; further attempts buy little on transient failures but add
+// latency and re-send the full (1M-context) history again on deterministic ones
+// — see vault Bugs.md. Retries use escalating backoff (avoids a subscription
+// rate-limit cascade from rapid re-spawns) and stop once total retry time
+// exceeds the wall-clock cap, so a slow-failing attempt can't strand the user.
+// All below the SSE boundary, so the user sees only a slightly longer
+// "thinking" gap, never an intermediate error.
+const MAX_TURN_RETRIES = 2;
+const RETRY_BACKOFF_BASE_MS = 750; // delay = base * attempt# → 750ms, 1500ms
+const RETRY_WALL_CLOCK_CAP_MS = 12_000; // stop retrying past this much elapsed
 // Map a non-zero CLI exit + any captured "API Error:" line to a concise,
 // user-facing Local Mode message. The raw upstream text stays in operator logs
 // only — we never echo a wall of provider error JSON into the chat. Referring
@@ -135,7 +152,7 @@ const MAX_TURN_RETRIES = 1;
 function describeCliFailure(code, apiErrorText) {
     const t = apiErrorText.trim();
     if (/API Error:\s*400/i.test(t) && /(tool use|concurren|parallel)/i.test(t)) {
-        return 'Local Mode hit a known Claude Code error (a print-mode bug that affects every current version). I retried once automatically — sending the message again sometimes gets through.';
+        return 'Local Mode hit a known Claude Code error (a print-mode bug that affects every current version). I retried a few times automatically — sending the message again sometimes gets through. See https://github.com/anthropics/claude-code/issues/18131';
     }
     if (/^API Error:/i.test(t)) {
         return `Local Mode error from Claude Code: ${t.replace(/^API Error:\s*/i, '').trim()}`;
@@ -144,8 +161,9 @@ function describeCliFailure(code, apiErrorText) {
 }
 // ─── Spawn ────────────────────────────────────────────────────────────────────
 function spawnClaude(params) {
-    const { conversationId, presenceSessionId, text, uid, history, vaultFileOpen, clientCapabilities, syncedFolders, onEvent, onDone, onError } = params;
+    const { conversationId, presenceSessionId, text, uid, history, vaultFileOpen, clientCapabilities, syncedFolders, onEvent, onDone, onError, onNotice } = params;
     const attemptIdx = params._attemptIdx ?? 0;
+    const firstAttemptAt = params._firstAttemptAt ?? Date.now();
     const systemPromptPath = (0, path_1.join)((0, os_1.tmpdir)(), `agent-${uid}.md`);
     const mcpConfigPath = (0, path_1.join)((0, os_1.tmpdir)(), `mcp-${uid}.json`);
     if (verbose) {
@@ -201,6 +219,19 @@ function spawnClaude(params) {
         existing.kill('SIGTERM');
         active.delete(conversationId);
     }
+    // Cancel any retry scheduled for this conversation that hasn't fired yet.
+    // Without this, a new user message arriving during a retry's backoff window
+    // would race the stale retry — which carries the OLD turn's history and would
+    // clobber the new turn. Skip when this call IS the retry firing (attemptIdx>0,
+    // the timer already deleted itself before invoking us).
+    if (attemptIdx === 0) {
+        const pending = pendingRetries.get(conversationId);
+        if (pending) {
+            clearTimeout(pending);
+            pendingRetries.delete(conversationId);
+            process.stderr.write(`[bridge] cancelled pending retry for ${conversationId} (superseded by new turn)\n`);
+        }
+    }
     // Note: ephemeral context (vault_file_open / client_capabilities / synced_folders)
     // is injected into the last user message by the gateway BEFORE history is
     // sent over the WS. The bridge no longer constructs `userMessageText` —
@@ -506,13 +537,23 @@ function spawnClaude(params) {
             catch { /* ignore */ }
         }
         if (code !== 0 && code !== null) {
-            // Auto-retry once when the CLI failed BEFORE producing any real output —
-            // the signature of the known print-mode 400 regression. A fresh spawn
-            // (new --session-id) usually succeeds. We never retry once real text or a
-            // tool call landed, to avoid double-running a side-effectful tool.
-            if (attemptIdx < MAX_TURN_RETRIES && sawApiError && !producedRealOutput) {
-                process.stderr.write(`[bridge] turn failed before output (${apiErrorText.replace(/\n+/g, ' ').slice(0, 120)}) — retrying once\n`);
-                spawnClaude({ ...params, _attemptIdx: attemptIdx + 1 });
+            // Auto-retry when the CLI failed BEFORE producing any real output — the
+            // signature of the known print-mode 400 regression. A fresh spawn (new
+            // --session-id) often succeeds. We never retry once real text or a tool
+            // call landed, to avoid double-running a side-effectful tool. Retries use
+            // escalating backoff and stop past the wall-clock cap (see consts above).
+            const elapsed = Date.now() - firstAttemptAt;
+            if (attemptIdx < MAX_TURN_RETRIES && sawApiError && !producedRealOutput && elapsed < RETRY_WALL_CLOCK_CAP_MS) {
+                const delay = RETRY_BACKOFF_BASE_MS * (attemptIdx + 1);
+                const nextAttempt = attemptIdx + 2;
+                process.stderr.write(`[bridge] turn failed before output (${apiErrorText.replace(/\n+/g, ' ').slice(0, 120)}) — retrying (${nextAttempt} of ${MAX_TURN_RETRIES + 1}) in ${delay}ms\n`);
+                // Admin-only ephemeral thread notice — jargon is fine in Local Mode.
+                onNotice?.(`Claude Code print-mode 400 (tool-use concurrency, anthropics/claude-code#18131) — respawning, attempt ${nextAttempt}/${MAX_TURN_RETRIES + 1}…`);
+                const timer = setTimeout(() => {
+                    pendingRetries.delete(conversationId);
+                    spawnClaude({ ...params, _attemptIdx: attemptIdx + 1, _firstAttemptAt: firstAttemptAt });
+                }, delay);
+                pendingRetries.set(conversationId, timer);
                 return;
             }
             // Pass any partial token usage we observed before the failure so the

package/dist/index.js

@@ -244,6 +244,14 @@ async function handleMessage(conversationId, text, sessionId, history, auth, vau
                 currentWs.send(JSON.stringify({ type: 'stream', conversationId, event }));
             }
         },
+        onNotice: (message) => {
+            // Ephemeral, non-persisted thread notice (admin-only Local Mode). Relayed
+            // by the gateway to the PWA SSE stream as a `notice` AgentEvent; it does
+            // NOT go through the turn accumulator, so it never lands in history.
+            if (currentWs?.readyState === ws_1.default.OPEN) {
+                currentWs.send(JSON.stringify({ type: 'notice', conversationId, message }));
+            }
+        },
         onDone: (messageCount, costUsd, usage, model) => {
             const parts = [];
             if (usage)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@1presence/bridge",
-  "version": "0.34.0",
+  "version": "0.35.0",
   "description": "Run 1Presence on your Mac and use your Claude.ai Pro subscription from any device",
   "bin": {
     "1presence-bridge": "dist/index.js"