dsh-plan-checkup

LIMITATIONS

Known limitations

- It looks for specific slips: dangerous commands, unrequested work, vague steps, missing tests. It does not judge whether the approach is right or whether the facts a plan relies on are true. - The questions are written for coding plans. For research or writing plans, most of them don't apply. - Steps are split from the plan's top-level list or headings; nested items fold into their step, a step longer than 1,200 characters is cut, and sections that aren't steps (goals, background) are not checked. - If the judgment engine is the same model that wrote the plan, it shares that model's blind spots. - Thresholds were calibrated on synthetic data with Qwen3.8-27B. Other engines and models need their own check; the probabilities and votes in `ledger.jsonl` can be used for that. - *May miss part of the request* is the weakest check; treat it as a soft reminder. - Plans from sub-agents never get a review card (sub-agents can't ask the user), so they are not checked. - **Request changes** cancels the review, so the feedback is sent by you, not by the plugin. - It is an advisory tool, not a security boundary: dsh's sandbox and approvals work as before.