[BUG] App fails to load when self-hosted in a subpath with a reverse proxy

### Problem Description

Feature request: LangWatch Evaluation Nodes not visible in n8n Cloud

### **Problem**

fix(traces): count the span type facet only when its section opens

## For humans

fix(instant-evals): own decision models run without the hosted flag, and Evals Preview hides where it can't run

## For humans

Langy shows the same vague 'model provider rejected that' for not-deployed and unsupported-parameter errors

## Bug Every upstream rejection shows the same Langy message: **"The model provider rejected that. Try again, or pick a different model."** Two very different failures look identical:

Langy sends reasoning.effort to non-reasoning OpenAI models (gpt-4.1 → 400)

## Bug Langy sends a reasoning effort setting (`reasoning.effort`, from `ThinkingLevel: "medium"`) on **every** `openai/*` model. Models without reasoning support reject the whole call.

Model picker offers the full OpenAI catalog when the OpenAI provider has a custom base URL

## Bug When the OpenAI provider has a custom base URL (for example an Azure OpenAI v1 endpoint), the model pickers still show the **full OpenAI catalog** (100+ models). The endpoint can only serve the models it actually has. Picking any other model fails at runtime.

fix(analytics): read the stock trace widgets from trace_metrics

Closes langwatch/tasks#1117

Dashboards: per-dashboard access, and concurrent widget queries on small installs

Thanks for the self-hosted option. The custom dashboards in v3 were a big part of why we chose LangWatch.

Bulk imports on small self-hosted installs: eval-trigger retries dominate ClickHouse load

Thanks for the self-hosted option and for v3; it's been solid for us. Sharing what we saw during a backfill, in case it's useful.

Analytics > Topics: 'Threads Per Topic' and 'Most Discussed Topics' label series with raw topic ids instead of topic names

Hey Team, GREAT PLATFORM! really enjoying it. The following issue seems to be consistent theme across all charts honestly, a lot of these charts show ids and non-human tags/labels/legends kind of thing, not helpful for a human eye. Not sure if i've configured wrong stuff or wrong data schema, i've pointed my coding age

First trace of a new project can get no trace summary

On a fresh local stack, the first trace sent to a new personal project got no trace summary, so it never showed in the traces list. Its spans were stored and its coding-agent session folded fine.

pi: one trace per process, so a multi-prompt session shows one turn

The pi extension picks one trace id per pi process (`traceId()` in `sdks/typescript/src/cli/utils/governance/pi-extension-source.ts`). An interactive session with several prompts therefore lands as one trace with several `pi.harness.run` roots.

Coding agent traces: say when message content was not captured

When a coding agent runs with message content capture off (`OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT` unset), its trace has no input, output or conversation. The drawer shows the generic empty panes, which read as a bug.

Coding agent traces are named after their root span

A coding agent's trace is named after its root span, verbatim. So a pi trace reads `pi.harness.run`, a Copilot one `invoke_agent`, and so on. The user never typed those.

Remove the "Checked as" line from the scenario run drawer

<!-- convoz:@langwatch-assistant --> ## What

Experiment workbench shows stale saved-dataset rows after the dataset changes elsewhere (e.g. version restore)

## Bug

Instant Evals explorer says "Nothing matches" when every row was skipped

## For humans

feat(coding-agent): capture Grok CLI sessions through its native OpenTelemetry

## For humans

feat(aigateway): serve OpenAI Decisions API on /v1/decisions

## Why

feat: agent and evaluator versioning with deploy labels, plus experiment and dependency fixes

## Summary

Scenario setup and teardown hooks for HTTP agents

<!-- convoz:@langwatch-langy --> ## Why

Dashboards: batch a board's widget queries with tRPC httpBatchStreamLink

## Problem

fix(sso): self-hosted sign-in sends colleagues straight to the sole live connection

## SSO: self-hosted sign-in goes straight to the sole live connection

Trace renderings sent to judges: six defects (span typing, bounded digest, system prompt, RAG contexts, relative time, legacy digest)

Found while building the judge benchmark (langwatch/tasks#905). The benchmark renders traces with LangWatch's own code at main `88566e69`, and these are the renderings judges receive today. Each one changes what a judge or Instant Eval sees.

LLM Judge Quality Analysis: Measure the Reliability of Your LLM Evaluators

**Problem** LangWatch has a powerful LLM-as-judge evaluation system, but there is currently no way to measure whether those judges are actually reliable. Every evaluation score in the platform rests on an untested assumption: that the judge is correct.

Feature request: Improve resiliency at trace endpoint

We would like to improve the resilience of message delivery in the trace reporting system. Currently, messages are sent via HTTP to the `api/collector` endpoint. However, this approach is prone to failures — if the request fails, the message is lost.