Sentry noise reduction: downgrade/classify non-actionable errors across core + tauri

[senamakel]

Problem

52 of 75 unresolved Sentry issues across tauri-react, tauri-rust, and core-rust are not actionable code defects — they are user configuration errors, third-party provider failures, connectivity flakiness, or informational messages logged at error level. Together they account for the vast majority of event volume, drowning out real bugs.

Categories of noise (with representative shortIds)

1. User/provider config errors (~22 issues, ~26k events)

These are the user's own API keys, model choices, or subscriptions failing — not something we can fix:

• Provider 4xx: huoshan 400 subscription (TAURI-RUST-Q4H, 4.8k), switchbot 401 token (TAURI-RUST-NTY, 3.8k), openai gpt-5.6 not supported (TAURI-RUST-KCN, 2.5k), openrouter 404 owl-alpha (TAURI-RUST-P98, 2k), ElevenLabs 401 payment (TAURI-RUST-QBN, 286), ollama context size exceeded (TAURI-RUST-Q8K, 281), Embedding API 401/404 (TAURI-RUST-GF3/CORE-RUST-1FF, 2.4k combined), ollama embed unauthorized/invalid model (CORE-RUST-GG/1GP, 348), llama-swap no router (CORE-RUST-1GQ, 56), deepseek-sensenova 400 (CORE-RUST-1PK, 51), qwen3:8b not available (CORE-RUST-CX/Y1, 1k), cloud API 401 (CORE-RUST-1A6, 175)
• Email config: IMAP "Unsafe Login" (TAURI-RUST-Q2T, 1.1k)
• Skill fetch 404s: hermes-agent SKILL.md (CORE-RUST-1NK, 14)

2. Third-party provider 5xx / internal errors (~6 issues, ~6k events)

Failures in the user's local or external provider — out of our control:

• ollama 500 EOF (TAURI-RUST-4R4, 2.7k), lmstudio 400/502 (TAURI-RUST-56Q/CORE-RUST-1DF/1NG, 3.2k combined)

3. Connectivity / network flakiness (~8 issues)

Transient failures that are expected and self-resolving:

• CoreRpcError: Failed to fetch (TAURI-REACT-32/33/37, 4e)
• UnhandledRejection: Failed to fetch (TAURI-REACT-34, 2e)
• CoreRpcError HTTP 401 (TAURI-REACT-35, 1e) — auth token expiry
• error sending request to api.tinyhumans.ai (CORE-RUST-EA, 52e)
• OpenHuman API 504 Gateway Timeout (CORE-RUST-18G, 32e)

4. Informational messages logged at error (~4 issues, ~8.4k events)

These are not errors at all — they are operational states or configuration choices:

• approval gate DISABLED (CORE-RUST-1MF, 8,031 events): OPENHUMAN_APPROVAL_GATE=0 is a supported config, not an error. Should be info! or debug!.
• No backend session (CORE-RUST-5J, 209e) / backend session expired (CORE-RUST-1MG, 23e): expected state for offline use or session expiry. Should be warn! or debug!.

5. Expected/self-resolving conditions (~5 issues)

• auth store failed: other (TAURI-REACT-1S, 95e/0u) — 0 users, likely CI/bot
• Failed to create thread (TAURI-REACT-3A, 1e) — one-off user action
• Aborted() WASM (TAURI-REACT-2Z/2Y, 2e) — WebAssembly internal abort, not actionable
• Rive mascot memory OOB (TAURI-REACT-2X, 2e) — WASM memory, may self-resolve

6. Stale/ancient clients (~4 issues)

Errors from very old releases (v0.54.0) still hitting the same Sentry project:

• composio triage failed (CORE-RUST-CW, 565e, v0.54.0)
• OpenHuman API 504 (CORE-RUST-18G, 32e, v0.54.0)
• skill_registry_search timeout (CORE-RUST-1MV, 14e)
• steering 404 (CORE-RUST-1PG, 5.7k — same as already-tracked TAURI-RUST-PNK/orchestration steering read floods Sentry: GET /orchestration/v1/steering 404s every 20s (TAURI-RUST-PNK) #5149)

Proposed fixes

A. error! → warn! / info! downgrade

Audit every Sentry-captured log in the categories above and downgrade:

• User config / provider errors: warn! — the user needs to fix their config, but it's not an application bug
• Third-party 5xx: warn! — out of our control
• Connectivity: debug! or warn! at most — transient, self-resolving
• Informational: info! or debug! — not errors
• Expected state: debug! — normal operation

B. before_send filter / event sampler

Add a Sentry before_send hook (or equivalent in the Rust sentry crate config) that drops or samples:

• Error messages matching known user-config patterns (provider 4xx, model not found, subscription errors)
• Connectivity errors (Failed to fetch, timeout, 502/504)
• Events from releases older than N versions (stale clients)
• Events with 0 users after a grace period (bot/CI noise)

C. Rate-limiting for non-actionable categories

For errors we keep at warn! but still want to see occasionally: rate-limit to 1 event per category per user per hour/day so a single misconfigured user doesn't generate thousands of events.

D. Release-based filtering

Drop or route events from releases older than (say) 5 minor versions. A user still on v0.54.0 generating 500+ composio errors is not actionable for the current codebase.

Impact

After these fixes, the Sentry feed would shrink from ~75 unresolved issues to ~15 genuinely actionable ones, making real regressions immediately visible.

Source: Sentry triage of tauri-react, tauri-rust, core-rust — 2026-07-23.

[YellowSnnowmann]

Closing as fixed.

All 52 noise categories from the issue body were addressed in PR #5171 (merged 2026-07-24). Three before_send filter functions — is_user_config_provider_event, is_connectivity_event, is_stale_release_event — now exist in src/core/observability.rs and are wired into both src/main.rs and app/src-tauri/src/lib.rs. The approval-gate DISABLED log was downgraded to log::info!. Session-expired events are filtered via is_session_expired_event. The KNOWN_PROBE_METHODS allowlist in src/core/dispatch.rs handles method-not-found noise.