anthropics / anthropics/claude-code

[MODEL] "Succeeded but empty" never triggers suspicion — HTTP 200 + [] propagated silently for days (271-incident retro, 2/5)

Abierto
#94,169 1 comentario 0 reacciones 0 asignados Ver en GitHub
Lenguaje dominante
Python
Estrellas
145k
Forks
23.1k
Métricas de merge de PR
Métricas de PR pendientes

Descripción

### Context

From a 90-day retrospective of **271 logged incidents** building two production SaaS apps with Claude Code (each record has root cause + prevention + detection channel). This is a recurring **pattern report**; sibling reports from the same dataset are linked at the bottom.

### Type of Behavior Issue

Claude made incorrect assumptions about my project (specifically: that success-shaped responses are trustworthy).

### What Happened

When an upstream dependency fails *politely* — **HTTP 200 with an empty array** — Claude (and the pipelines it writes) treat it as a valid result:

- Our scraping vendor, when blocked by the target site, returns success+empty. Five pipeline layers each reported "normal" while the product silently produced nothing **for days**. Every alert Claude had designed hung off `catch` blocks — and this failure never enters a catch. Our logged wording: *"all alerting hangs off catch — which is no alerting at all."*
- A separate fetch-window bug silently discarded 2/3 of candidates. Same signature: nothing thrown, nothing logged, output just quietly shrank.

When asked to investigate "no data" symptoms, Claude's default hypothesis was "there is genuinely no data," not "the source may be failing politely."

### Expected Behavior

In data-pipeline contexts, **empty-success from an external dependency should be treated as a hypothesis to investigate, not a result to propagate.** Concretely:

1. When writing pipeline code, default to recording *why* a result was empty (quota / blocked / genuinely none) and flagging consecutive-empty streaks.
2. When verifying its own work, Claude Code could surface "this verification returned 0 rows/items" as a distinct signal instead of letting it read as a pass.

### Reproducibility

Pattern-level; at least 3 major incidents in 90 days, one costing days of silent production downtime.

### Model / Version / Platform

Opus (most sessions; some Sonnet) · Claude Code 2.1.235 · Anthropic API · macOS

### Impact

High — silent product downtime plus real money burned on paid API calls that returned nothing.

Guía de contribución

No hay ninguna guía de contribución indexada para este repositorio

Línea de trabajo

The payload names no repository files, tests, or entry points; start by locating the response-handling and verification paths related to the reported pipeline behavior. Done means the scope is agreed and the relevant behavior distinguishes empty-success responses, consecutive-empty streaks, and zero-item verification results.

Escrito por el modelo de indexación a partir del texto del issue.

Evaluación

Área
ai, data-engineering, tooling
Tipo de issue
Nueva funcionalidad
Dificultad
5/5
Tiempo estimado
Más de una semana
Estado de actividad
Activo
Claridad
Necesita aclaración
Aptitud para principiantes
25/100

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.