Ponytail Problems, Limitations and Troubleshooting

Registration, loading, and behavior are different observations. Record which check failed instead of reporting every unexpected output as “Ponytail not working.”

Trigger: a request such as “make this simpler” leaves behavior, compatibility, and error handling unstated. An agent may choose a shorter interpretation that was never acceptable to the user.

Preserve: the observable contract, supported inputs, data integrity, and required failure behavior. Compare the patch against a written acceptance list, not just the old line count. Include at least one relevant invalid or boundary input.

Useful instruction: “Keep the current public interface and all documented behavior. Reuse existing code when suitable. Ask before dropping a requirement; explain any change that alters error handling.” This is an editorial example, not an upstream test result. Revert the patch if the contract is unclear, refine the task, and try again from a clean state.

Trigger: a new implementation needs separate responsibilities, an extension boundary, or a compatibility adapter, but the prompt describes only the immediate happy path. A smaller implementation may be easier to read today while violating the actual contract.

Preserve: structure justified by a concrete requirement. State the required callers, deployment boundaries, or extension behavior. Ask the agent to explain why a layer is needed and which simpler alternatives it considered; an architectural diagram by itself does not establish necessity.

Useful instruction: “The adapter is required because both supported backends must remain compatible. Keep that boundary. Simplify inside it, and do not replace the contract with a special case for the current example.” Review both supported paths, then use the same acceptance tests after refactoring.

Trigger: a proposed cleanup removes validation, authorization checks, safe error handling, accessible labels, keyboard behavior, or cleanup after a failed operation. A small diff can still create a serious defect.

Preserve: trust-boundary validation, data-protection behavior, required permissions, and accessible operation. Ponytail's upstream guidance does not license deleting these safeguards to obtain a smaller result.

Useful instruction: “Preserve authorization and server-side validation. Keep error paths that protect data. The interaction must remain keyboard accessible with a meaningful label and focus state. Reduce duplication only after those conditions pass.” Check forbidden as well as permitted inputs. If a proposed deletion lacks an equivalent safeguard, reject it rather than counting it as saved code.

Trigger: a readable sequence becomes a dense one-liner, or a named concept becomes an expression repeated across callers. Line count falls while comprehension or maintenance becomes harder.

Preserve: useful names, visible failure paths, and the ability to inspect intermediate states. Ask a reviewer to explain the new code without relying on the agent's description. Run boundary tests and examine whether duplication has simply moved elsewhere.

Useful instruction: “Prefer the clearest complete implementation. Keep named steps where they express distinct decisions. A one-liner is appropriate only if its behavior is immediately readable.” The checked skill's decision ladder is contextual; it is not an unconditional rule to compress every function.

• Claude Code: ordinary Claude web chat is not the same plugin surface. Some features have additional host-version requirements.
• Codex: installing a plugin does not eliminate the need to inspect hook trust. Skill invocations use the plugin namespace.
• Cursor: the hooks contract does not inject into subagents; cloud agents do not trigger its documented session-start path. A retained checkout and the right scope matter.
• Rules-only paths: guidance in a file does not provide plugin commands or a dependable plugin mode toggle.
• All paths: instructions can be misunderstood or compete with other applicable instructions. The model still needs a clear task and human review.

These boundaries are described in the installation documentation and Cursor hooks contract. They are not evidence of reproduced failures in every current host version.

For an upstream defect, use the official issue tracker after checking existing reports. Include the Ponytail commit, host version, operating system, installation method and scope, applicable rules, expected behavior, actual behavior, and the smallest nonprivate task that demonstrates it.

Remove secrets, private paths, credentials, and proprietary code from logs. A code example should establish the problem without uploading a real production repository. For an error in this guide, use the editorial contact and identify the page and source that need correction.

To recover, stop applying the proposed changes, restore the known-good project state, and use the platform's documented removal or rule-restoration steps. Off is not uninstall. Broad cleanup can affect shared state across hosts, so match the operation to the scope you intend to change.