🤖 fix: Remove premature commitment from todo_write description

[ammar-agent]

Fixes premature commitment issues in both todo_write and status_set tool descriptions.

Problem

Both tool descriptions contained prescriptive language that pressured agents to commit to narratives before validating outcomes:

todo_write issues:

• "Use this for ALL complex, multi-step plans" - forced usage before validation
• "Before finishing your response, ensure all todos are marked as completed" - pressured false completions
• "Update frequently as work progresses" - created distraction from errors
• Complex structural guidance about old/recent/current/immediate/far future work - added cognitive load

status_set issue:

• "Set a final status before completing that reflects the outcome" - ambiguous timing allowed premature success claims

Changes

todo_write:

Removed:

• ❌ "Use this for ALL complex, multi-step plans"
• ❌ "Before finishing your response, ensure all todos are marked as completed"
• ❌ "Update frequently"
• ❌ Complex structural guidance

Added:

• ✅ "The TODO list is displayed to the user at all times" - contextualizes importance
• ✅ "If you hit the 7-item limit, summarize older completed items into one line" - brings back useful guidance without rigid structure
• ✅ "If work fails or approach changes, update the list to reflect reality" - explicit permission to show failures
• ✅ "Only mark tasks complete when they actually succeed" - clear expectation

Kept:

• ONE in_progress at a time
• Ordering (completed first, in_progress, pending last)
• Tense guidelines (past/present progressive/imperative)

status_set:

Changed:

• ✅ "Set a final status after completion, only claim success when certain (e.g., after confirming checks passed)"

This makes the sequencing explicit: do work → verify outcome → set final status.

Impact

Agents should:

• Validate approach before creating TODOs
• Update TODOs to reflect failures, not hide them
• Only claim success after verifying outcomes
• Focus on actual work rather than tool maintenance

Testing

No code logic changes - these are prompt/description updates. Will observe agent behavior in real use.

Generated with cmux