Branch: refs/heads/misc-conversion-session-cost Home: https://github.com/xwiki/xwiki-dev-llm Commit: d763630e4e3892977e29a56f43f21b3288ca6c52 https://github.com/xwiki/xwiki-dev-llm/commit/d763630e4e3892977e29a56f43f21b... Author: Vincent Massol <vincent@massol.net> Date: 2026-08-30 (Sun, 30 Aug 2026) Changed paths: M .claude-plugin/marketplace.json M kimi.plugin.json M opencode.jsonc M xwiki/.claude-plugin/plugin.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 A xwiki/skills/xwiki-doc-writing/tools/docplan.py M xwiki/skills/xwiki-doc-writing/tools/docshot.sh Log Message: ----------- [Misc] Cut the per-session cost of a documentation conversion Measured across the 58 past sessions that converted or wrote xwiki.org documentation (the LaTeX, Release Notes and AntiSpam conversions): 1.44B billed context tokens, of which 96-99% is cache reads — a session costs roughly turns x context, so the levers are fewer turns and a smaller context. 22% of all those tokens were spent re-orienting before the first authoring action of the session, and ~17% on the screenshot shoot/look/adjust loop. * Add `docplan.py` to the doc-writing tools: `status` prints a conversion's whole orientation in one call — the current task's full brief, the Setup answers, the recent Decisions and Open questions, and what comes next — replacing PLAN.md read by hand in five to thirteen `sed` chunks per session (76 KB re-read, against a 19 KB digest). `start` / `done` / `block` write the status and the Outcome to both PLAN.md's row and the task file, which is also the pair that otherwise drifts apart. Decisions and Open questions are trimmed to their most recent *whole* entries with a count of what was dropped: a real plan accumulates tens of thousands of characters of them, and the recent ones are the ones a session would otherwise re-litigate. * Make `docshot.sh` run `checkredbox.py` on the shot it just took whenever a box selector was given, and fail if the box is broken. A capture now reports its own verdict, so there is never a reason to open the PNG to find out whether it worked — an image read into a session stays in context for every later turn, which is what turns a screenshot into a shoot/look/adjust loop. `CHECK=0` opts out; a shot with no box is skipped. * Point the `xwiki-doc-convert` resume protocol at `docplan.py` in both the skill and the plan reference, and add the three habits that decide whether a task fits in one session: one call rather than a conversation for anything mechanical, never open a screenshot to check a capture, and read the inventory rather than the raw legacy source. * Add a rule to keep Decisions and Open questions to entries that still bind a future session, since the digest shows the recent ones and a decision buried under fifty later ones is one a session will not see. Version bumped 1.2.1 -> 1.3.0 (minor — the skills now depend on a newly shipped tool). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> To unsubscribe from these emails, change your notification settings at https://github.com/xwiki/xwiki-dev-llm/settings/notifications