@1presence/bridge 0.33.0 → 0.35.0

package/dist/claude.js

@@ -111,9 +111,59 @@ function renderHistoryMessage(msg) {
 }
 // ─── Active processes ─────────────────────────────────────────────────────────
 const active = new Map();
+// 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
+// to "Claude Code" here is intentional and consistent with Local Mode's other
+// operational errors (e.g. the "claude CLI not found" message): in Local Mode
+// the user is knowingly running their own Claude Code install.
+//
+// NOTE on the 400 tool-use case: this is an open Claude Code print-mode
+// regression (introduced in 2.1.19, still present in 2.1.146 — the current
+// latest), so upgrading does NOT fix it. We deliberately do not suggest an
+// upgrade; the automatic retry is the real mitigation and resending sometimes
+// gets through.
+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 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()}`;
+    }
+    return `Local Mode stopped unexpectedly (claude exited with code ${code ?? 'unknown'}). Please try again.`;
+}
 // ─── Spawn ────────────────────────────────────────────────────────────────────
 function spawnClaude(params) {
-    const { conversationId, presenceSessionId, text, uid, history, vaultFileOpen, clientCapabilities, syncedFolders, onEvent, onDone, onError } = params;
+    const { conversationId, presenceSessionId, text, uid, history, vaultFileOpen, clientCapabilities, syncedFolders, onEvent, onDone, onError, onNotice } = params;
+    const attemptIdx = params._attemptIdx ?? 0;
+    const firstAttemptAt = params._firstAttemptAt ?? Date.now();
     const systemPromptPath = (0, path_1.join)((0, os_1.tmpdir)(), `agent-${uid}.md`);
     const mcpConfigPath = (0, path_1.join)((0, os_1.tmpdir)(), `mcp-${uid}.json`);
     if (verbose) {
@@ -125,6 +175,13 @@ function spawnClaude(params) {
         process.stderr.write(paint('90', `[bridge:verbose] conversation:  ${conversationId}`) + '\n');
         process.stderr.write(paint('90', `[bridge:verbose] history turns: ${history.length}`) + '\n');
     }
+    // Surface the user's UID before the session line in every mode — it's the
+    // Firestore doc prefix (`sessions/<uid>_<conversationId>`), so logging it
+    // makes a reported bridge failure correlatable to the stored session without
+    // having to ask which account hit it. The CLI's own `--session-id` is
+    // ephemeral and is NOT the Firestore conversationId, so the uid is the key
+    // join column when debugging.
+    process.stderr.write(`[bridge] user ${uid}\n`);
     // Debug transcript: lead with the user prompt for this turn (the clean
     // message, before the gateway's ephemeral-context prefix), plus the session
     // id (correlates with the chat URL / Firestore session doc) and a hint at
@@ -162,6 +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` —
@@ -198,6 +268,10 @@ function spawnClaude(params) {
     //   across turns of a chat — even with --no-session-persistence. The
     //   bridge passes the per-spawn `conversationId` here; the presence
     //   sessionId is correlated separately via bridge logs and spool records.
+    // The CLI treats --session-id as "claim this new session ID" and rejects a
+    // reused id with "Session ID X is already in use". A retry is a fresh spawn,
+    // so it MUST use a new uuid; the first attempt keeps the correlation id.
+    const spawnSessionId = attemptIdx === 0 ? presenceSessionId : crypto.randomUUID();
     const args = [
         '--print',
         '--input-format', 'stream-json',
@@ -210,7 +284,7 @@ function spawnClaude(params) {
         '--mcp-config', mcpConfigPath,
         '--strict-mcp-config',
         '--no-session-persistence',
-        '--session-id', presenceSessionId,
+        '--session-id', spawnSessionId,
     ];
     const pinnedModel = (0, config_1.getBridgeModel)();
     if (pinnedModel) {
@@ -266,6 +340,15 @@ function spawnClaude(params) {
     let extractedModel = null;
     let buffer = '';
     let killedForViolation = false;
+    // Retry/error-surfacing tracking for this attempt:
+    //  - sawApiError: the CLI emitted an "API Error:" assistant text event (the
+    //    way Claude Code reports an underlying API failure mid-turn).
+    //  - apiErrorText: that text, captured for describeCliFailure().
+    //  - producedRealOutput: any real assistant text or tool_use was emitted, so
+    //    a later failure must NOT be retried (could double-run a side-effect).
+    let sawApiError = false;
+    let apiErrorText = '';
+    let producedRealOutput = false;
     proc.stdout.on('data', (chunk) => {
         buffer += chunk.toString('utf-8');
         const lines = buffer.split('\n');
@@ -282,6 +365,10 @@ function spawnClaude(params) {
                 continue;
             }
             const type = event['type'];
+            // Set when this event is the CLI's "API Error:" turn — we neither forward
+            // it to the PWA nor let it reach the accumulator (it carries no real
+            // content and would poison history / show a raw error mid-stream).
+            let suppressEvent = false;
             // Extract model + key source info from the first system/init event.
             // No session-id persistence — Firestore is the only source of truth
             // now, and we pin --session-id to presenceSessionId on every spawn.
@@ -322,6 +409,7 @@ function spawnClaude(params) {
                     let wroteText = false;
                     for (const block of content) {
                         if (block['type'] === 'tool_use') {
+                            producedRealOutput = true;
                             const toolName = block['name'];
                             const toolId = block['id'];
                             if (toolId)
@@ -370,13 +458,27 @@ function spawnClaude(params) {
                         else if (block['type'] === 'text') {
                             const text = block['text'];
                             if (text) {
-                                if (debug) {
-                                    // Full text, newlines intact — the readable transcript.
-                                    debugBlock('assistant', exports.SECTION_COLORS.assistant, text);
+                                if (/^API Error:/i.test(text.trimStart())) {
+                                    // The CLI is reporting an underlying API failure as assistant
+                                    // text. Capture it for the user-facing message, and suppress
+                                    // the whole event so the raw error never reaches the PWA or
+                                    // the accumulator (the gateway also blanks it via
+                                    // cleanTurnText — this is the upstream defense).
+                                    sawApiError = true;
+                                    apiErrorText = text.trim();
+                                    suppressEvent = true;
+                                    process.stderr.write(paint(exports.SECTION_COLORS.result, `[bridge] ${text.replace(/\n+/g, ' ')}`) + '\n');
                                 }
                                 else {
-                                    process.stderr.write(paint(exports.SECTION_COLORS.assistant, text.replace(/\n+/g, ' ')));
-                                    wroteText = true;
+                                    producedRealOutput = true;
+                                    if (debug) {
+                                        // Full text, newlines intact — the readable transcript.
+                                        debugBlock('assistant', exports.SECTION_COLORS.assistant, text);
+                                    }
+                                    else {
+                                        process.stderr.write(paint(exports.SECTION_COLORS.assistant, text.replace(/\n+/g, ' ')));
+                                        wroteText = true;
+                                    }
                                 }
                             }
                         }
@@ -412,7 +514,8 @@ function spawnClaude(params) {
                 if (typeof c === 'number')
                     costUsd = c;
             }
-            onEvent(event);
+            if (!suppressEvent)
+                onEvent(event);
         }
     });
     proc.stderr.on('data', (chunk) => {
@@ -434,9 +537,29 @@ function spawnClaude(params) {
             catch { /* ignore */ }
         }
         if (code !== 0 && code !== null) {
+            // 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
-            // PWA and the gateway's bridge usage store can still record it.
-            onError(`claude exited with code ${code}`, usage, extractedModel);
+            // PWA and the gateway's bridge usage store can still record it. Surface a
+            // classified, user-readable message instead of the opaque exit code.
+            onError(describeCliFailure(code, apiErrorText), usage, extractedModel);
         }
         else {
             onDone(messageCount, costUsd, usage, extractedModel);

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": "0.33.0",
+  "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"