Skills changelog
New versions of my Claude Code plugins, newest first. RSS
-
compost 0.34.3
The skill-drift mod's
missing_guidancequestion fired on nearly everycompost:buildturn (16 of 18 turns with no gap scored over 0.6). It now asks whether something the skill relies on failed to fit the case: a command returning an outcome the skill doesn't say how to handle, or a method it prefers that its own tools can't check, with no fallback given. The threshold is now 0.5. On the 23 labeled build turns, 5 hit a real gap (Jev find-test exiting 2, or duplicate-test not seeing a doctest build preferred). Over ten runs it flagged at most 1 of the other 18 and 4 or 5 of the 5; the margin is a few hundredths, so the live eval allows one false hit. Only Jev-tool gaps are labeled so far. -
compost 0.34.2
The skill-drift mod judges
deviatedstep by step: one Jev question per numbered step of the skill, and the turn deviates by its likeliest skipped step, which the hit's evidence names. The single whole-skill question scored everycompost:buildturn 0.75-0.95 whether or not a step was skipped, so it flagged all of them. On 23 hand-labeled build turns from two toy repos, step by step at 0.75 flags none of the 8 that did every step and 11 or 12 of the 15 that skipped one (it varies run to run). The threshold was chosen on 17 of those turns and held on the other 6, from a second repo. Steps are keyed by position, so a skill with two numbered lists (pause) has every step asked. A skill with no numbered steps keeps the whole-skill question, whosedeviatedthreshold also moves from 0.7 to 0.75; onlywait-whathas no steps. The labeled turns are compost-evals'skill-drift-steps.json, replayed bytests/test_evals_skill_drift.py. -
compost 0.34.1
The skill-drift mod judges a skill loaded through the Skill tool against the turn of the loop that loaded it. A skill a subagent loaded (an implement worker's
compost:build) was judged against the main session's turn, which never ran it, so every lane run logged a falsedeviatedhit for build. Subagent turns are now judged on their own transcript; only the main loop's turns get the next-message correction check. A skill an agent definition preloads (no Skill call) still counts as the main loop's. -
compost 0.34.0
compost:implementstep 2 defines done (closed, or ticked on the parent), blocked, taken (another session's branch for the issue, matched as<issue>-*, exists; it is pushed when the issue starts so other clones see it, and the claim expires after a day with no comment on the parent from any other session), and the frontier in one place, andcompost:buildchecks blockers against the same "done". On GitHub a reviewed, landed blocker stays open until the PR merges, so build's old "every blocker is closed" check reported it as blocking. -
understand 0.2.4
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
superfan 0.1.1
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
observability-harness 0.3.2
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
mochi-creator 1.5.1
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
learn-harness 0.1.1
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
hexagonal-agents 1.6.3
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
compound-knowledge 0.10.3
Add per-plugin CHANGELOG.md and require an entry on every version bump
-
compost 0.33.2
thin.pyno longer crashes on a test whose parametrize id contains|: it cuts the coverage context's phase off the last|, not the first, sotest_x[a|b]stays one test id. -
compost 0.33.1
compost:setupcatches compost enabled under two ids (compost@compostandcompost@oliphant-plugins):pile.py replacedoffers to disable the copy that isn't running, and never the one that is. -
compost 0.33.0
compost moved to its own repository, JoshuaOliphant/compost, with history;
compost@oliphant-pluginskeeps installing it from there. -
compost 0.32.5
-
compost 0.32.4
Maintenance release.
-
compost 0.32.3
Compost stop-guard: let a message's origin decide whether the user sent it (#206)
-
observability-harness 0.3.1
Verify.sh stage 2 no longer dies on an unbound variable (#203)
-
learn-harness 0.1.0
Initial release.
-
understand 0.2.3
Move directory normalization into the shared config loader (#106)
-
understand 0.2.2
Initial release.
-
superfan 0.1.0
Initial release.
-
observability-harness 0.3.0
Initial release.
-
mochi-creator 1.5.0
Retire the feedback skill in favour of auto memory
-
mochi-creator 1.4.3
Initial release.
-
hexagonal-agents 1.6.2
Name the eval datasets
-
hexagonal-agents 1.6.1
Carry eval seed state in case inputs
-
hexagonal-agents 1.6.0
Retire the feedback skill in favour of auto memory
-
hexagonal-agents 1.5.1
Hexagonal-agents: build the eval task function on the real Agent SDK API
-
hexagonal-agents 1.5.0
Initial release.
-
compound-knowledge 0.10.2
Maintenance release.
-
compound-knowledge 0.10.1
Maintenance release.
-
compound-knowledge 0.10.0
Add the garden face to compound-graduate (#60)
-
compound-knowledge 0.9.1
Move directory normalization into the shared config loader (#106)
-
compound-knowledge 0.9.0
Retire the feedback skill in favour of auto memory
-
compound-knowledge 0.8.8
Compound-knowledge: make the solution schema and the captured corpus agree (#73)
-
compound-knowledge 0.8.7
Initial release.
-
compound-knowledge 0.8.6
chore: bump the plugins the ready-for-agent sweep changed
-
understand 0.2.1
Merge main (plugin consolidation) into prune-repo-docs
-
understand 0.2.0
Merge main (plugin consolidation) into compost-prune
-
observability-harness 0.2.1
Merge main (plugin consolidation) into prune-repo-docs
-
observability-harness 0.2.0
Merge main (plugin consolidation) into compost-prune
-
hexagonal-agents 1.4.3
Merge main (plugin consolidation) into prune-repo-docs
-
hexagonal-agents 1.4.2
Merge main (plugin consolidation) into compost-prune
-
compound-knowledge 0.8.5
chore: bump the plugins the ready-for-agent sweep changed