## Summary
- Extends `expected_error_kind` matcher ladder with 11 new substring arms across two existing buckets — no new variants, no new emit sites, no behavior change beyond reclassifying known wire shapes that were leaking past Wave 1-3 anchors.
- Closes 13 Sentry IDs (~366 events) — **all deterministic user-environment / user-config errors with no remediation path** (offline, DNS fail, captive portal, bad API key, out of credits, wrong model id, missing OAuth scope, region block).
- Pure classifier — no UI, retry, or fallback logic touched. The reliable-provider stack still falls back to OpenHuman's hosted tier for the Lane O bodies; users keep their chat. Lane N socket failures already gated through `report_error_or_expected` at `ws_loop.rs:191`, so the threshold-escalation event simply demotes to a `warn` breadcrumb instead of paging.
- Two micro-commits — Lane O (`is_provider_config_rejection_message` PHRASES const) and Lane N (`is_network_unreachable_message` substring arms) — each independently revertible.
## Problem
After the Wave 1-3 sweep landed, fresh Sentry triage surfaced **13 unresolved IDs** whose bodies match the *spirit* of an existing classifier bucket but use wire-shape variants the current substring anchors miss:
### Lane O — custom-provider config rejection (~250 events)
The `ProviderConfigRejection` variant + `is_provider_config_rejection_message` exist precisely for "user pointed OpenHuman at a custom_openai endpoint with a model / temperature / region / credential that provider doesn't accept." But the 7 phrases shipped in Wave 1-3 (#2079 / #2076 / #2202) only cover the DeepSeek / OpenRouter abstract-tier-leak and Moonshot temperature shapes. New surfaces:
| Sentry ID | Events | Real wire body |
|---|---|---|
| `R1` | 44 | `{"error":{"message":"This model is not available in your region.","code":403}}` |
| `R4` | 14 | `{"code":403,"reason":"ModelNotAllowed","message":"模型不允许访问"…}` (Doubao/ChatGLM) |
| `YC` | 16 | `{"error":{"type":"invalid_authentication_error"…}}` |
| `S5` | 14 | `{"error":{"message":"This request requires more credits, or fewer max_tokens…"}}` (OpenRouter 402) |
| `Y0` | 13 | `{'error': '/chat/completions: Invalid model name passed in model=reasoning-v1…'}` |
| `JN` | 14 | `{"error":{"message":"No active credentials for provider: openai"…}}` |
| `KB` | 16 | Same `No active credentials for provider` shape from OpenHuman backend re-emit |
| `JK` | 17 | `litellm.BadRequestError: Github_copilotException - Bad Request…` |
| `J2` / `J5` / `J4` | 62 | `{"error":{"message":"model 'llama3.3' not found","type":"not_found_error"…}}` |
### Lane N — socket WebSocket-connect transport (~116 events)
`is_network_unreachable_message` already catches `connection refused` / `dns error` / `network is unreachable` but two real shapes escape:
| Sentry ID | Events | Real wire body |
|---|---|---|
| `44` | 50 | `[socket] Connection failed: WebSocket connect: IO error: failed to lookup address information: nodename nor servname provided, or not known` |
| `4P` | 66 | `[socket] Connection failed: WebSocket connect: HTTP error: 200 OK` (captive portal / corporate proxy intercepting the WS upgrade handshake) |
Every one of these is deterministic **user-environment** / **user-configuration** state — the maintainers have no remediation. Sentry has no signal to act on. Every event was pure noise.
## Solution
### Lane O — `src/openhuman/inference/provider/config_rejection.rs:55-79`
Append 8 new phrases to the `PHRASES` const (case-insensitive substring match, same precedence as existing 7):
```rust
"not available in your region", // R1
"modelnotallowed", // R4
"invalid_authentication_error", // YC
"requires more credits", // S5
"invalid model name passed in model=", // Y0
"no active credentials for provider", // JN + KB
"litellm.badrequesterror", // JK
"not_found_error", // J2 + J5 + J4
```
Each anchor is intentionally narrow (e.g. `passed in model=` not bare `invalid model name`; `litellm.badrequesterror` not bare `litellm`) so a stray log line elsewhere can't accidentally demote a real provider/backend bug. The HTTP-layer wrapper (`is_provider_config_rejection_http`) still guards on `provider != openhuman_backend::PROVIDER_LABEL`, so a model rejection from **our own** backend (which would be a real regression) still reaches Sentry.
### Lane N — `src/core/observability.rs:299-319`
Three new substring arms appended to `is_network_unreachable_message`:
- `failed to lookup address` + `nodename nor servname` — libc `getaddrinfo()` failure renderings on macOS / BSD / POSIX resolvers when tungstenite wraps as `IO error` without the reqwest `dns error` prefix.
- `http error: 200 ok` — tungstenite-only render. Reqwest renders HTTP 200 as `"HTTP status server error (200)"`, so no collision with the regular HTTP call path. A negative precedence test (`http_200_classifier_does_not_silence_unrelated_log_lines`) pins this against benign `HTTP/1.1 200 OK` / `status: 200 OK` prose so a future broadening cannot silence success traces.
### Design trade-off — classifier vs. root-fix validation layer
The Lane O bugs *do* have a more durable root fix: a pre-save provider validation layer (test the API key, validate the model id against `/v1/models`, surface region-blocks at config-save time). That's a real product initiative requiring UX design, per-provider model-list infrastructure, and meaningful spec work — out of scope for a triage sweep. This PR follows the Wave 1-3 maintainer precedent (classifier-first noise suppression) so the Sentry signal/noise ratio improves immediately; the validation layer remains tracked as a separate follow-up. If/when it lands, these classifier arms become redundant belt-and-suspenders and can be deleted without conflict.
Lane N has no root-fix alternative — offline / firewall / captive-portal / DNS failures are pure user-environment state. Classifier-demote is the correct disposition.
## Submission Checklist
> If a section does not apply to this change, mark the item as `N/A` with a one-line reason. Do not delete items.
- [x] Tests added or updated (happy path + at least one failure / edge case) per [Testing Strategy](../gitbooks/developing/testing-strategy.md#failure-path-requirement)
- [x] **Diff coverage ≥ 80%** — changed lines (Vitest + cargo-llvm-cov merged via `diff-cover`) meet the gate enforced by [`.github/workflows/coverage.yml`](../.github/workflows/coverage.yml). 11 positive tests pinned to real Sentry bodies + 1 negative precedence test cover every new substring arm; only the rustdoc comments and the negative `unrelated_*` test body are non-executable lines.
- [x] N/A: behaviour-only change — Coverage matrix updated — added/removed/renamed feature rows in [`docs/TEST-COVERAGE-MATRIX.md`](../docs/TEST-COVERAGE-MATRIX.md) reflect this change
- [x] All affected feature IDs from the matrix are listed in the PR description under `## Related`
- [x] No new external network dependencies introduced (mock backend used per [Testing Strategy](../gitbooks/developing/testing-strategy.md#mock-policy))
- [x] N/A: classifier-only change with no UI / release-cut surface touched — Manual smoke checklist updated if this touches release-cut surfaces ([`docs/RELEASE-MANUAL-SMOKE.md`](../docs/RELEASE-MANUAL-SMOKE.md))
- [x] Linked issue closed via `Closes #NNN` in the `## Related` section
## Impact
- **Runtime**: no behavior change for users — chat / agent / web-channel / socket paths continue to fall back / retry exactly as before. The only delta is that `report_error_or_expected` now classifies these 13 wire shapes as `ExpectedErrorKind::*` instead of escalating as `tracing::error`, so Sentry stops receiving the events.
- **Performance**: negligible — `is_provider_config_rejection_message` lowercases once and runs 15 substring scans (was 7); `is_network_unreachable_message` runs 11 substring scans (was 8). Both already on the error-path, never on the hot path.
- **Security**: no new attack surface. Substring matchers run on already-rendered error messages — no parsing, no allocation beyond the existing `to_ascii_lowercase()`.
- **Migration**: none.
- **Compatibility**: forward-compatible. New phrases are append-only — no existing matcher behavior changes. If a future maintainer ships the root-fix validation layer, deleting these arms is a clean local edit.
## Related
- Closes OPENHUMAN-TAURI-R1
- Closes OPENHUMAN-TAURI-R4
- Closes OPENHUMAN-TAURI-YC
- Closes OPENHUMAN-TAURI-S5
- Closes OPENHUMAN-TAURI-Y0
- Closes OPENHUMAN-TAURI-JN
- Closes OPENHUMAN-TAURI-KB
- Closes OPENHUMAN-TAURI-JK
- Closes OPENHUMAN-TAURI-J2
- Closes OPENHUMAN-TAURI-J5
- Closes OPENHUMAN-TAURI-J4
- Closes OPENHUMAN-TAURI-44
- Closes OPENHUMAN-TAURI-4P
- Follow-up PR(s)/TODOs: pre-save provider-validation layer (validate API key + model id + region at config-save time) tracked as separate product initiative; would supersede the Lane O arms.
---
## AI Authored PR Metadata (required for Codex/Linear PRs)
> Keep this section for AI-authored PRs. For human-only PRs, mark each field `N/A`.
### Linear Issue
- Key: N/A (Sentry-driven triage; no Linear ticket)
- URL: N/A
### Commit & Branch
- Branch: fix/sentry-wave4-classifier-no
- Commit SHA: e16414ac (tip)
### Agent
- Claude Code (Sonnet 4.6) running locally
<!-- This is an auto-generated comment: release notes by coderabbit.ai -->
## Summary by CodeRabbit
* **Bug Fixes**
* Enhanced detection of network connectivity failures and provider configuration rejection scenarios for more accurate error classification.
* **Tests**
* Added comprehensive test coverage for network-related transport failures and provider configuration rejection cases.
<!-- review_stack_entry_start -->
[](https://app.coderabbit.ai/change-stack/tinyhumansai/openhuman/pull/2309?utm_source=github_walkthrough&utm_medium=github&utm_campaign=change_stack)
<!-- review_stack_entry_end -->
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
Co-authored-by: oxoxDev <nikhil@tinyhumans.ai>
Co-authored-by: Steven Enamakel <enamakel@tinyhumans.ai>
Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Steven Enamakel <enamakel@tinyhumans.ai>
Co-authored-by: google-labs-jules[bot] <161369871+google-labs-jules[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Steven Enamakel <enamakel@tinyhumans.ai>
