Summary creation and AI chat blocked with "error too many failed attempts" after single successful summary

[hyper-charlie]

Description

User successfully created a summary for a meeting (~90 min, completed in 10-15 seconds). Minutes later, when attempting to create a summary for a different meeting, both summary creation and AI chat are blocked with "error too many failed attempts" error. User reports this is unusual as they regularly create multiple summaries in succession without issues.

Reproduction Steps

1. Create a summary for a ~90 min meeting (succeeds, takes ~10-15 seconds)
2. A couple minutes later, attempt to create a summary for a different ~90 min meeting → fails with "error too many failed attempts"
3. Attempt to chat with AI → same error
4. Restart app → issue persists

Expected Behavior

Creating a second summary minutes after a successful first summary should work normally.

Actual Behavior

Both summary creation and AI chat return "error too many failed attempts" immediately. User can retry right away with no cooldown, but error persists. This blocks access to both features.

Environment

• Platform: macOS (aarch64)
• OS Version: macOS 26.3.1
• App Version: 1.0.14
• Build: 57054b4
• Locale: en-CA
• User ID: 1da53afd-bd6f-4a7f-9b40-fd09857782d9

Notes

• User has stable internet connection
• No other activity between successful summary and failed attempt
• User has never encountered this before despite regularly creating multiple summaries in succession
• Error affects both summary creation and AI chat simultaneously
• User does not want to retry for fear of losing the successful summary

[devin-ai-integration]

Codebase Exploration Findings

Key Finding: Error string not in codebase

The exact error message "too many failed attempts" does not appear anywhere in the Char codebase (desktop app, API server, Rust crates, or npm dependencies including the Vercel AI SDK, OpenRouter SDK, and Supabase SDK). This means the error originates from an external service — most likely the upstream LLM provider (via OpenRouter) or Supabase GoTrue's auth rate limiting.

---

Relevant Files & Components

Error display path (how users see errors):

• apps/desktop/src/session/components/note-input/enhanced/enhance-error.tsx — Displays error.message for enhance/summary errors
• apps/desktop/src/chat/components/message/error.tsx — Displays error.message for chat errors
• apps/desktop/src/store/zustand/ai-task/tasks.ts (lines 280-311) — extractUnderlyingError() unwraps AI_RetryError from the Vercel AI SDK to surface the underlying error message

AI task system (summary creation & chat share the same LLM path):

• apps/desktop/src/store/zustand/ai-task/task-configs/enhance-workflow.ts — Enhance/summary workflow; makes 2-4 LLM API calls per summary (template generation + summary generation with up to 2 validation retries)
• apps/desktop/src/chat/transport/index.ts — Chat transport using ToolLoopAgent, also goes through the same LLM provider
• apps/desktop/src/ai/hooks/useLLMConnection.ts — Both features use the same createLanguageModel() → same provider, same API key, same endpoint

API server rate limiting:

• apps/api/src/rate_limit.rs — Governor-based per-user rate limiter. Returns HTTP 429 with body "rate limit exceeded"
• apps/api/src/main.rs (lines 100-111) — LLM rate limits:
  • Pro: 30 burst + 1/sec replenish (very generous)
  • Free: 5 burst + 1 per 12 hours (very restrictive — a single summary can consume 2-4 of those 5 tokens)

Auth/token flow (relevant because access_token is the API key for hosted LLM):

• apps/desktop/src/auth/context.tsx — Supabase auth with autoRefreshToken: true and continuous auto-refresh ticker
• apps/desktop/src/auth/client.ts — Supabase client configuration
• apps/desktop/src/auth/errors.ts — Fatal session error detection (handles refresh_token_not_found, refresh_token_already_used)

Vercel AI SDK retry mechanism:

• node_modules/ai@6.0.116/src/util/retry-with-exponential-backoff.ts — Retries retryable errors (including 429) up to maxRetries (default 2), then throws RetryError with message "Failed after N attempts. Last error: ..."

---

Most Likely Root Causes (ranked by likelihood)

1. API server rate limiting (free tier):
If the user is on the free tier, the LLM rate limit is 5 requests per 12 hours. A single summary uses 2-4 LLM requests (template generation + summary stream + possible retries). Two summaries back-to-back would easily exhaust the quota. The 429 response from the API would be retried by the AI SDK, eventually surfacing as an error. Both summary and chat would be blocked because they share the same rate limiter.

2. Upstream LLM provider rate limiting:
For the hyprnote provider, requests are proxied through the API server to an upstream LLM provider (OpenRouter → model provider). The upstream provider could return its own rate limit error, which gets proxied back. The error text "too many failed attempts" could originate from the upstream provider's error response.

3. Supabase auth token refresh failure:
If the Supabase access_token expired and auto-refresh was rate-limited by GoTrue, all API requests would fail with auth errors. This would affect both summary and chat simultaneously and persist across restarts (since the auth state is persisted). There's extensive recent work on auth edge cases (commits b76e32ee5, 060c0646d, e94f51bb8, f90c08af5, ee8fc929a, 91b902a40, 66fe86fe6).

---

Why Both Features Break Simultaneously

Summary creation and AI chat both use useLLMConnection() → createLanguageModel() which routes through the same provider, API key, and API endpoint. They share the same server-side rate limiter (keyed by auth.claims.sub). So if one feature exhausts the rate limit or the auth token is invalid, both features fail.

---

Recent Related Changes

• a22522ad5 — "proper ratelimit for api usage, both free and pro" — introduced the current rate limiting tiers
• 344fa6c5c — "fix: gracefully yield output instead of throwing when validation exhausts retries (fix: gracefully yield output instead of throwing when validation exhausts retries #4192)" — changed retry exhaustion behavior
• 66fe86fe6 — "fix: prevent analytics event spam from Supabase token refreshes (fix: prevent analytics event spam from Supabase token refreshes #4221)"
• b76e32ee5 — "fix: prevent permanent auto-refresh death on singleton Supabase client (fix: prevent permanent auto-refresh death on singleton Supabase client #3773)"
• e94f51bb8 — "fix: bypass broken Tauri visibilitychange for Supabase auth session refresh (fix: bypass broken Tauri visibilitychange for Supabase auth session refresh #3911)"

---

Suggested Investigation Steps

1. Check the user's subscription tier (user ID: 1da53afd-bd6f-4a7f-9b40-fd09857782d9) — if free, this is almost certainly the API rate limit
2. Check API server logs for this user around the time of the error — look for 429 responses on LLM endpoints
3. Check Supabase auth logs for token refresh failures for this user
4. Consider improving the error message — the current flow surfaces raw upstream errors via error.message. Adding user-friendly messages for rate limit scenarios (e.g., "You've reached your usage limit") would help users understand the issue

[hyper-charlie]

User Report: Jules Bordat (ea46357a-26e4-492a-9105-b1c23b558981)

Environment:

• Platform: macOS
• OS Version: macOS 26.3.1
• App Version: 1.0.17
• Build: 866a098
• Locale: en-GB

Issue: Chat stopped generating notes with "Too many requests" error. Unable to generate summaries and AI responses are blocked.

This appears to be a recurring issue affecting multiple users. Additional context or reproduction steps would help narrow down the root cause.