Skip to content

Troubleshooting

Start with:

Terminal window
centaur doctor

Use centaur doctor --redact-paths when sharing output. It replaces workspace, Centaur home, history, and integration paths with safe labels. Review the result before posting it.

Error: no search/replace blocks found in response

Copy the complete model response. Confirm it contains File:, <<<<<<< SEARCH, =======, and >>>>>>> REPLACE delimiter lines. The default prompt tells the model not to wrap the blocks in an outer code fence.

If the AI answered an informational or architectural question without proposing file edits, there are no patches to apply. An exact NO_CHANGES response is a successful no-op. Extra prose around it is not.

Error: search text matches N times in <file>

The Search text appears more than once. Ask the model to include more unchanged lines around the target. Centaur refuses to guess.

Error: search text not found in <file>

The source may have changed since it was exported. Export fresh context, then ask the model to regenerate its response. Use --llm auto only as an optional local fallback.

The changed or staged export is unexpected

Section titled “The changed or staged export is unexpected”

Run git status or git diff --cached --name-only in the same directory. Use explicit paths with a full export when Git state is not the desired scope.

Save the complete model response to a UTF-8 text file:

Terminal window
centaur --file response.txt

Scripts can also use standard input. Preview first or provide explicit approval:

Terminal window
generate-centaur-patch | centaur --stdin --dry-run
generate-centaur-patch | centaur --stdin --yes

Restart the AI client after centaur install, then run centaur doctor. For one client, check the exact MCP and command IDs in Client setup.

Codex installation updates the current project’s .codex/config.toml; the project must be trusted before Codex loads it. Claude Code installation uses its private local scope for the current project. centaur doctor checks the entry and workspace binding, but only the restarted client can confirm that the tools were discovered.

Provider browser choices are independent in [chat.browsers]. For example, Qwen can use Firefox while Gemini uses Chrome. An explicit --browser overrides the saved browser for that handoff. A failed named-browser launch no longer opens a different browser.

In the development build, centaur browser accounts prints non-secret account memory from accounts.json in the Centaur home. Remember the intended account explicitly with:

Terminal window
centaur browser accounts --to gemini --browser chrome --select intended@example.com
centaur browser accounts --to gemini --browser chrome --check

Selection remembers your intent and clears the previous verification. It does not sign in, switch accounts, or prove a subscription. Native browser sessions use the profile label browser-session because the OS profile cannot be identified reliably. Use --profile centaur only for Centaur’s separate Chrome profile. Account memory never contains cookies, tokens, or passwords. The last check is historical evidence, and sending checks the current account again.

--check inspects selected existing browser pages and records signed_in, account_mismatch, or unknown in Centaur. It never opens or selects a tab, focuses a window, copies context, or sends a request. Unknown and mismatched checks exit with an error. The active browser address must match the inspected document, and its visible account control must identify the intended account. A sign-in tab or another logged-in account cannot satisfy this check.

If the intended chat is in an inactive tab, explicitly add --select-tab to the check. Centaur inspects existing provider tabs, selects the one whose account matches, and checks it again. It never opens a page, changes accounts, inserts context, or sends. A failed search restores the original tab selections.

For Gemini, add --url 'https://gemini.google.com/u/2/app?pageId=none' to an explicit selection if that is your observed app URL. This is only a route hint: numeric account indices can change. Handoffs inspect existing provider tabs for the intended email, use the hint only among verified matches, and recheck after focus before inserting context and before submission.

Windows handoffs inspect existing provider pages before opening a new window. A separate sign-in tab does not establish that the app is signed out. Ambiguous tabs, an unreadable identity, or a different account stop automatic sending. Select the intended provider tab and account, then retry. Claude and Qwen use Copy task with their built-in entries. Centaur only sends automatically after verifying the intended account. Normal browser handoffs and Centaur’s separate ChatGPT profile have their own account memory.

For MCP, inspect get_settings.workspace before exporting files. A connection bound to another project stays bound to that project; global account memory is not permission to use its files. Reconnect the intended workspace through client setup. get_settings also exposes saved browser choices and account observations without treating them as current login proof.

If an account check or browser confirmation stops the handoff before sending, fix the browser or account selection and review the task again. You can also use another provider or a manual handoff.

If submission was attempted and Centaur cannot confirm the outcome, check the original conversation before sending again. On Windows, open Menu > Inspect send recovery. For an agent job, check the sent field in get_job. A value of null means submission is uncertain, true confirms submission, and false means Centaur did not send. If a reply already exists, copy it and choose Menu > Save reply to a task.

Centaur asks a clarification only when the outcome is too vague to select safe, useful context. Answer the one focused question in the terminal. For JSON or another non-interactive caller, repeat the request with the missing detail from status: "clarification_needed".

centaur check --run --sandbox requires macOS Seatbelt or Linux Bubblewrap. Centaur refuses the command when no backend is available. On Windows, run the displayed check without --sandbox only after reviewing it; Centaur still supervises the process, timeout, environment, and child cleanup, but that is not OS filesystem isolation.

Error: target file modified since patch was applied

Centaur refuses undo when a target changed after the recorded patch. It also refuses history records that predate drift-safe metadata.

Inspect centaur history. Recover manually from the named snapshot instead of overwriting later work.

For less common behavior involving encodings, multiple blocks, large exports, symlinks, or MCP pagination, see Edge cases.

Windows build fails with “Access is denied (os error 5)”

Section titled “Windows build fails with “Access is denied (os error 5)””

Error: failed to remove target\debug\centaur.exe: Access is denied. (os error 5)

On Windows, running executables are locked by the OS and cannot be overwritten by cargo build or cargo test. This commonly happens if a background tool (such as Claude Code or an MCP client) was configured to run target\debug\centaur.exe directly rather than an installed release binary.

To resolve:

  1. Install the binary to Cargo’s global bin directory: cargo install --path . --locked
  2. Run centaur mcp install or configure clients to invoke centaur (or %USERPROFILE%\.cargo\bin\centaur.exe) instead of the debug target directory.
  3. Run centaur doctor to inspect for active binary locks and ensure client configurations point to the installed executable.

Run centaur doctor --redact-paths, review the output for private information, and report the problem. GitHub sign-in is required. For a suspected vulnerability, use the private security form.

Website privacy