Home / Articles / Re-anchoring / Versions / v1.1.0
Re-anchoring: what changed in v1.1.0
From v1.0.0 (2026-10-10, 0335818e4) to v1.1.0 (2026-10-10, 1a69fcfaf), paragraph by paragraph.
3 paragraphs added, 1 removed, 0 changed in place, 67 unchanged. About 286 words added and 39 removed. Insertions are marked like this, deletions like this; unchanged runs are folded to one line; figures appear as their file names.
52 unchanged paragraphs, under In short, What summarisation does to a session, Three ways a rule fails…
• The re-anchor after a compaction has not been observed yet. There has been no compaction since the hooks were installed. When there is, the anchor line will show it, and this article will be updated with what happened.
• The re-anchor after a compaction has now been observed. Updated 10 October 2026: at 12:03:36 UTC the session was summarised for the nineteenth time, automatically, from 789,144 tokens to 12,092. The SessionStart hook fired in the same second, printed the policy back from disk, and logged the event; the session resumed with the full policy in view, and the anchor line moved to 1 re-anchor for 1 compaction. The release that followed ran its leak scan before the commit, as rule 2 asks.
• Then the canary did its other job. At about 15:16 UTC the session's container was restarted, and everything outside the repository and the transcript went with it: the policy file, the script and the hooks in the user settings. The next prompt arrived with no anchor line. What the agent noticed first was its missing working files; the absent anchor confirmed that the hooks had gone too, and the first report due after the restart would have been missing, which is the signal the person watches for. The policy was restored from the copy the last re-anchor had printed into the conversation, and its hash came back as the same #a0d6d2, byte for byte. The script was rebuilt, carrying the last report's totals as a baseline. Lesson: a policy that lives on an ephemeral machine needs a copy somewhere that survives it, and the anchor line is worth keeping even when no report is due.
• The leak scan counter was undercounting scans. It recognised one way of writing the scan, so commits that were scanned in a different way were counted as unscanned. The scan is now one script that every commit runs, and the counter looks for that script.
15 unchanged paragraphs, under The recipe changes with the platform, What it does not do, Build your own