CCooScope.

Checklist

Write an agent bug report that names the failing layer

A recovery checklist for separating the model, harness, tools and application state when an agent behaves differently than expected.

4 min estimated readStart hereEvidence baseline

THE SHORT ANSWER

Preserve the failing input and environment before attributing an agent problem to a model.

Keep the incident attributed

An archived ClaudeDevs post attributes a particular incident to Claude Code and the Agent SDK harness, while distinguishing those from the models and API. That is the publisher’s historical explanation; it was not reproduced here. Current Claude Code guidance emphasizes concrete verification evidence. Together they suggest a useful reporting habit: name the component and preserve what actually failed.

Capture a minimal case

Write the expected result, the observed result and the earliest point they diverged. Save the agent version, model identifier, tool configuration, relevant instruction and a redacted transcript. Note whether a tool returned an error or the agent misused a successful response. Keep screenshots or artifacts when they show behavior the transcript alone cannot establish. Remove credentials and unrelated private content.

Recover without losing the clue

Before retrying, record any partial file changes or external actions so a second attempt cannot quietly duplicate them. Then choose a narrow diagnostic change, such as replaying the same harmless input with the same tool response. Report uncertainty if you cannot isolate the layer. This proposed checklist is a bug-reporting aid, not an analysis of a current outage or proof that one component is blameless.

Put it to work.

  • Record expected and observed behavior.
  • Save versions and redacted evidence.
  • Inspect partial effects.
  • Change one diagnostic condition.

Go to the source.

Original material, with its own scope and publication date. This note is an explanation, not a CooScope performance test.

Source discovery: @ClaudeDevs

Source published · Editorial review

Archived English post text; primary sources reviewed separately. Generated annotations, media and missing replies are outside the evidence scope.

  1. @ClaudeDevs — archived X post (2026-04-23)
  2. Anthropic — Claude Code best practicesLink / source review: 2026-10-15
Evidence baseline and source-review dates are different. A recent link check does not make an older study a current product test. Read our method →

Keep the thread going.

Explore this topic →