SKILL.md carries two self-estimated context-percentage handoff triggers that are mutually inconsistent (>40% at line ~191, >50% at line ~207), in a genuinely long-running resumable pipeline. Models cannot measure their own context occupancy reliably; sibling skills (session-flow handoff) already disclaim fixed token counts, and context-guard exists as the measured mechanism. Redesign the handoff triggers around qualitative/measured signals and reconcile the two thresholds; mis-editing casually would break the resume protocol, hence a dedicated change.
From the prompting-guide alignment audit (PR #2018, finding V-absence-008 / KES-003/004).
SKILL.md carries two self-estimated context-percentage handoff triggers that are mutually inconsistent (>40% at line ~191, >50% at line ~207), in a genuinely long-running resumable pipeline. Models cannot measure their own context occupancy reliably; sibling skills (session-flow handoff) already disclaim fixed token counts, and context-guard exists as the measured mechanism. Redesign the handoff triggers around qualitative/measured signals and reconcile the two thresholds; mis-editing casually would break the resume protocol, hence a dedicated change.
From the prompting-guide alignment audit (PR #2018, finding V-absence-008 / KES-003/004).