How Do You Write a Useful DeepSeek Harness Bug Report Without Secrets?

1. Give the DeepSeek Harness bug report one question

Choose a title naming the action and visible failure. “New session cannot send after model selection” is clearer than “Harness broken.” Include an exact error code when available; keep suspected causes separate.

Keep each report focused. When failures occur at unrelated steps, describe them separately or explain the observed sequence connecting them. Combining startup, authentication and unexpected edits without context makes investigation harder.

2. Record the environment you can verify

Include the operating system, installed Harness version, installation route and relevant runtime or plugin versions. Record the selected provider and model when they relate to the problem, using consistent placeholders for private names. Mark unavailable version information as unknown.

Say whether the task ever worked and what changed. Include a last known working version only with evidence. A date alone cannot identify the executable or configuration used.

Keep the inventory proportional. A model-selection problem rarely needs your full project history. Include detail because it helps reproduce the failure, not to make the report longer.

3. Reduce the task to a shareable reproduction

Create a small synthetic example that contains no private code or real customer information. Write numbered actions someone else could follow, starting from a clear state. Keep the actual input when it is harmless, or replace it with an equivalent synthetic input and test whether the same failure remains.

A minimal reproduction must be observed. If you cannot reproduce outside the original project, say so and describe the relevant condition. Do not upload the project to fill the gap.

• State the initial condition, such as a newly created session and selected sample workspace.
• List the actions in order, including the provider or model choice if relevant.
• Give the expected result and the actual visible result separately.
• Say whether the failure occurred once, repeatedly or intermittently in your own attempts.

4. Prepare a redacted error log, not a raw archive

Start with the shortest excerpt that includes the error and necessary context. Review it locally before sharing. Remove API keys, authorization headers, cookies, access tokens, private repository URLs, usernames and sensitive paths. Check query strings and command arguments as well as obvious key fields.

Use clear replacements such as [REDACTED_TOKEN], [WORKSPACE] and [PRIVATE_HOST]. Preserve the same replacement wherever the same value appears. Retain non-sensitive error codes and structure so the report remains interpretable. Never include part of a credential as proof that you used the right one.

Open the final screenshot after cropping or covering private material. Terminal scrollback, browser tabs and paths can disclose unrelated information. A text excerpt is often easier to review than a full-screen recording.

5. Use a compact example without pretending it is a real incident

The following is an illustrative report skeleton. Replace the labels with your verified observations; no failure or version is claimed here.

• Title: [specific action] produces [exact error code].
• Environment: [OS], [Harness version], [installation route], [relevant plugin or runtime].
• Steps: open [sample workspace]; choose [provider/model]; start [session type]; submit [synthetic input].
• Expected: [observable result]. Actual: [observed result and redacted excerpt].
• Checks already tried: [one change per attempt and its outcome]. Unknowns: [unverified details].

6. Review the destination before sharing

The upstream repository (opens in a new tab) directs feedback and bug reports to GitHub Discussions. Check its current instructions and relevant existing reports before deciding where to post. This guide prepares a report; it does not submit anything.

For an error in our community article, the DeepSeekDSH about page (opens in a new tab) explains the site’s identity and correction guidance. Identify the page and outdated statement. The community site does not represent upstream support.

Conclusion: a DeepSeek Harness bug report is ready when a reader can understand the failure, inspect the evidence and identify the remaining question. Share only what that purpose requires, after a final privacy review.