i-need-help
I need help
Purpose
Save time, tokens and user attention by recognizing when continued task work has stopped producing value. This is an automatic process skill, not a host watchdog; follow it when the evidence meets its trigger.
Check for progress
At the start of a continuation or handoff, compare current task history with the blocker before doing more work. If the history already records two unproductive attempts on that blocker, the threshold is met; do not take another task action to confirm it. Otherwise, before another attempt, compare its proposed approach and expected result with the previous attempt. Maintain only a short note in the task's existing state: blocker, approaches tried, observed results and last useful progress. Do not create tracking or evidence files for this process unless the task itself requires them.
An attempt is a purposeful approach or experiment and may contain several tool calls. Useful progress is relevant new evidence, resolved uncertainty or a verified step toward the requested outcome. A useful negative result counts. Repeated reads, cosmetic edits, changed command wording, reassuring status, unrelated discoveries and cycling between failed approaches do not.
Stop when two attempts at the same blocker add no useful progress. Count equivalent retries and oscillating approaches together, including attempts carried forward by a supervisor or earlier agent. When current task history records two matching attempts, treat the threshold as met; do not rerun them or search to reconstruct their results. Reset the count only when relevant evidence or conditions materially change. Stop sooner if an established missing capability, access or user decision leaves no useful authorized next step. Elapsed time alone is not a trigger.