Branch: refs/heads/vmassol/doc-conversion-performances Home: https://github.com/xwiki/xwiki-dev-llm Commit: 32ea962723b63c86568aa8d1fce6a95250af855f https://github.com/xwiki/xwiki-dev-llm/commit/32ea962723b63c86568aa8d1fce6a9... Author: Vincent Massol <[email protected]> Date: 2026-10-15 (Thu, 15 Oct 2026) Changed paths: M scripts/size-budgets.json M xwiki/skills/xwiki-doc-convert/SKILL.md M xwiki/skills/xwiki-doc-convert/references/conversion-plan.md M xwiki/skills/xwiki-doc-writing/tools/README.md M xwiki/skills/xwiki-doc-writing/tools/docplan.py Log Message: ----------- [Misc] Chain a documentation conversion's tasks instead of running one per session * Run tasks back to back, stopping only at the gates (after target-map, before publish, a blocked task) and handing off to a fresh session once the context passes ~200k tokens. * Add `docplan.py brief`: the current task's brief alone, for chaining on without reprinting the setup and decisions. * Merge the standard task set's small tasks: one `extract` for all legacy pages, one `originals` for the original pages and the deletions. * Keep the resume steps in SKILL.md only; the reference keeps what they rely on and the measured rationale. Lower its size budget accordingly. Measured on the transcripts of the LaTeX and Markdown conversions: one task per session spent 55-82% of most sessions on warm-up. Replayed turn by turn and priced, chaining with a ~200k hand-off costs 15-27% less than one task per session, and 150-200k is the cheapest threshold; a sub-agent per task costs the same or more once its cache writes are priced. Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]> To unsubscribe from these emails, change your notification settings at https://github.com/xwiki/xwiki-dev-llm/settings/notifications