<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j" -->

---
title: Debugger attached, upload speed tanked, and agents still...
description: This week's biggest story is a debugging tale with a twist: a file-storage engineer discovered that attaching a CDP debugger to Chromium throttles fetch...
canonical: https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Debugger attached, upload speed tanked, and agents still can't test right | daily.dev
og:description: This week's biggest story is a debugging tale with a twist: a file-storage engineer discovered that attaching a CDP debugger to Chromium throttles fetch...
og:url: https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j
og:image: https://api.daily.dev/og/posts/ErMOh8v9j.png
og:image:alt: Debugger attached, upload speed tanked, and agents still can't test right
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Debugger attached, upload speed tanked, and agents still can't test right

**[Testing Digest](https://daily.dev/sources/testing_digest)** · 5 min read · 0 upvotes · 0 comments

## Summary

This week's biggest story is a debugging tale with a twist: a file-storage engineer discovered that attaching a CDP debugger to Chromium throttles fetch uploads by up to 21x, meaning the measurement tool itself was the bug. A large empirical study found coding agents mostly ignore explicit instructions to use TDD, formal methods, or property-based testing, defaulting to shallow unit tests regardless of what you tell them. Webflow's MCP eval harness caught a production failure that only showed up on GPT models, not Claude, a reminder that single-model testing gives false confidence. And a CI pipeline post-mortem shows what happens when AI-assisted development quadruples merge volume without a matching bump in compute.

## Content

**TLDR:** This week's biggest story is a debugging tale with a twist: a file-storage engineer discovered that attaching a CDP debugger to Chromium throttles fetch uploads by up to 21x, meaning the measurement tool itself was the bug. A large empirical study found coding agents mostly ignore explicit instructions to use TDD, formal methods, or property-based testing, defaulting to shallow unit tests regardless of what you tell them. Webflow's MCP eval harness caught a production failure that only showed up on GPT models, not Claude, a reminder that single-model testing gives false confidence. And a CI pipeline post-mortem shows what happens when AI-assisted development quadruples merge volume without a matching bump in compute.

---

## Attaching a debugger to Chromium throttles fetch() uploads by up to 21x

A file-storage engineer spent time chasing a mysterious 5x gap between browser and .NET upload speeds, ruling out chunk count, pipelining, and concurrency as causes. The actual culprit: having a CDP debugger attached (via Playwright or Puppeteer) throttles Chromium's fetch() uploads that use Uint8Array/ArrayBuffer bodies by roughly 17-21x, while Blob and ReadableStream bodies are unaffected. Hand-driven tests without a debugger attached matched native .NET throughput, meaning the automated benchmark was measuring the debugger, not the browser. The team's actual fix, switching to Blob bodies, was a real ~1.3x win and sidesteps a separate Firefox bug where ReadableStream bodies get silently corrupted. [Read more](https://daily.dev/posts/IgiWe4Bwc)

## Coding agents told to use TDD or formal methods mostly ignore the instruction

A large study prompted coding agents (codex with GPT-5.6) across 26 conditions to use specific verification techniques, TDD, formal methods like Verus and TLA+, property-based testing, fuzzing, mutation testing, on a Zstd Rust implementation task. Almost none of the agents meaningfully adopted the named technique, instead defaulting to shallow unit tests no matter what the prompt asked for. Formal methods mostly produced trivial or irrelevant proofs, TDD underperformed as predicted, and the no-instructions baseline scored above average. Only a custom skill built to nudge agents away from default failure patterns did meaningfully better, and even that had real limits. [Read more](https://daily.dev/posts/SqBqS8j73)

## Webflow's MCP eval harness caught a bug Claude never triggered

Webflow built an eval harness running real agents (Claude Code and Codex CLI) against disposable sites to see how their MCP server behaves under actual agent use, not just whether tool calls return valid responses. Validating only against Claude gave false confidence: a ChatGPT app submission got rejected in production after a GPT-family model misused a Designer-canvas tool Claude had never touched. The team responded by adding a second agent-host driver and a deterministic semantic_html_ratio check to catch structural issues invisible to screenshot-based judges, and now runs the suite nightly. [Read more](https://daily.dev/posts/kYSFM4zeO)

## A CI pipeline crept back past 30 minutes as AI-assisted merges quadrupled

A team that had gotten its CI pipeline down to 22 minutes in 2025 watched it climb back past 30 as test suites doubled and merge volume quadrupled at peak, driven by AI-assisted development, while compute capacity grew only 7%. A three-person platform team spent six weeks splitting end-to-end tests into critical versus deprecated tiers, rebalancing parallel groups using real timing data, and adding an aggressive fail-fast kill switch. Runtime dropped to 15-16 minutes with a 95% green completion rate, and the team now tracks pipeline speed as an owned metric with explicit targets instead of treating it as a one-off fix. [Read more](https://daily.dev/posts/w7AssUsoY)

---

## Also notable

- **CodeRabbit couldn't prove it read a rules file; Qodo could, until its citations broke:** In a controlled test with a planted arbitrary rule, Qodo cited specific rule numbers and lines for every finding while CodeRabbit never named a source, but a follow-up perturbation test showed Qodo's citations didn't update when the rule moved, revealing they were reconstructed from matched text rather than a live re-read. [Read more](https://daily.dev/posts/IWC50a5Qy)
- **Mirrord's chaos testing lets independent agents fault-inject the same staging deployment:** Mirrord 3.250.0 attaches chaos rules to a single session instead of the whole cluster, so multiple agents can inject latency or connection resets against shared staging simultaneously without affecting each other or anyone else. [Read more](https://daily.dev/posts/sBiZ4bFSh)
- **tsgolint hits stable v7, now covering 59 of 61 type-aware ESLint rules:** tsgolint's Go-based type-aware TypeScript linting runs 12-18x faster than ESLint with typescript-eslint per the team's benchmarks, up from 43 covered rules in the December alpha to 59 of 61 now, though migration requires TypeScript 7.0+. [Read more](https://daily.dev/posts/ggPDJTuR6)
- **PHPUnit 13.3 adds a --repeat flag to surface flaky tests directly:** The revived --repeat option (and #[Repeat] attribute) runs individual tests multiple times to catch flakiness, alongside a new #[Retry] attribute for automatic retries that get transparently reported in the test summary. [Read more](https://daily.dev/posts/goqQXP1lK)
- **WordPress core is cleaning up 31,000 PHPUnit tests before 7.2:** The coordinated cleanup targets 501 loose assertEquals() calls to convert to assertSame() and 101 assertion-free test methods to remove, ahead of a PHPUnit 12 prerequisite that requires splitting oversized test classes. [Read more](https://daily.dev/posts/pNxHPJ9it)

## Similar posts on daily.dev

- [The Bug Was in My Benchmark: Attaching a Debugger to Chromium Throttles fetch\(\) Uploads 20x](https://daily.dev/posts/the-bug-was-in-my-benchmark-attaching-a-debugger-to-chromium-throttles-fetch-uploads-20x-igiwe4bwc) · Medium · 0 upvotes · 0 comments
- [Why your coding agent shouldn't be your QA system](https://daily.dev/posts/why-your-coding-agent-shouldn-t-be-your-qa-system-p3wg9nmqm) · QA Wolf · 2 upvotes · 0 comments

---

Tags: [#testing](https://daily.dev/tags/testing), [#ai-agents](https://daily.dev/tags/ai-agents), [#cicd](https://daily.dev/tags/cicd), [#mcp](https://daily.dev/tags/mcp), [#chromium](https://daily.dev/tags/chromium)

[View this post on daily.dev](https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j","headline":"Debugger attached, upload speed tanked, and agents still can't test right","text":"This week's biggest story is a debugging tale with a twist: a file-storage engineer discovered that attaching a CDP debugger to Chromium throttles fetch uploads by up to 21x, meaning the measurement tool itself was the bug. A large empirical study found coding agents mostly ignore explicit instructions to use TDD, formal methods, or property-based testing, defaulting to shallow unit tests regardless of what you tell them. Webflow's MCP eval harness caught a production failure that only showed up on GPT models, not Claude, a reminder that single-model testing gives false confidence. And a CI pipeline post-mortem shows what happens when AI-assisted development quadruples merge volume without a matching bump in compute.","url":"https://daily.dev/posts/debugger-attached-upload-speed-tanked-and-agents-still-can-t-test-right-ermoh8v9j","datePublished":"2026-09-14T04:18:22.005Z","dateModified":"2026-09-14T04:18:42.661Z","author":{"@type":"Organization","name":"Testing Digest","logo":"https://media.daily.dev/image/upload/s--1rQKIOqD--/f_auto,q_auto/v1773839410/logos/testing_digest?_a=BAMAMiiu0","url":"https://daily.dev/sources/testing_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/testing_digest","name":"Testing Digest"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Testing Digest","item":"https://daily.dev/sources/testing_digest"},{"@type":"ListItem","position":3,"name":"Debugger attached, upload speed tanked, and agents still can't test right"}]}
```

