Kevin Veen-BirkenbachandClaude Opus 5.5
0343a2f3ce
kvb-check-efficiency measures a change against its base: tokens.py counts skill text tokens and listing words, measure.py runs one command in a base worktree and the working tree (one run per side, the mean of five over 100, an estimate marked ~ when other processes load the CPU). kvb-check-speed reads wall time from the same run and is not relevant for skills. kvb-check-score prints fix, break, efficiency, speed and the product score with emojis plus a gate line: below 90/5 it reports, over 100 it requires asking the operator before the commit. Every skill that reported confidence and break risk now shows the score block; the 90/5 bar adds efficiency and speed at most 100. Commit blocks and the commit call's approval prompt name worktree, branch, diff stat and the score values, also for merges. 24 descriptions were shortened to keep the skill listing within its budget. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
3.9 KiB
name, description
| name | description |
|---|---|
| kvb-dream | Process the last 24 hours of every repository and session: memory, improvements, kvb-optimize, cleanups. Trigger on a request to learn from recent work. |
Replay what happened, separate what held from what broke, keep only conclusions that survive an attack, then optimize and clean up.
Objective
Make the next run measurably less likely to repeat each failure found; what
changes no future behaviour is not recorded. An error class is closed only by
something structural (test, lint rule, CI check, hook, written rule in the
agent instructions, skill), never by "be more careful". Per closed class,
report the kvb-check-score block of its structural fix, calibrated as the
kvb-check-confidence skill requires, and name what is still open.
Scope
Every repository and session this agent touched in the window: its
transcripts' working directories and worktrees, its kvb-org-rasci task
workdirs, the current one. Window: 24 hours, or wider on request (7d after
the command). State both in one table first. Recall and score each repository
on its own; a finding goes only into its own project's memory. Run only when
the operator asks; read hints.md next to this file when a hint appears or
before changing the reminder.
Step 1: recall
The actual record per repository, not your recollection:
- History:
git log --sinceover all branches plus--reflog(resets, rebases, amends), with--stat. - Working state:
git status, staged and unstaged diff, stashes. - CI: the window's runs, conclusions, and each red run's failing log line, not the job title.
- Reviews: pull requests and issues touched, with review comments.
- Register:
rasci.py toolsranks the tools the tasks ran (refactoring candidates);get <task>shows objective, roles, workdir, artifacts. - Memory: each project's memory directory and
MEMORY.md, to extend recorded findings instead of repeating them. - Instructions:
CLAUDE.md,AGENTS.md, the project's skills: which rule a mistake violated. - Sessions: readable transcripts, for operator corrections that never reached a commit.
Step 2: score
Every action, with evidence (sha, file:line, run id):
- Held: shipped and still proven: green CI, a passing test covering it, no revert, no follow-up fix.
- Broke: reverted, amended, hot-fixed later, red CI, or corrected by the operator; ending green does not undo the failed first attempt.
- Unknown: no verification either way, so no win; a finding of its own, never a quiet pass.
Steps 3 to 6: findings
When the record holds a pattern or an expensive failure, read
hypotheses.md next to this file. Offer each improvement, ranked by
evidence, in its own AskUserQuestion with multiSelect: false: Apply now, Record as a task, Skip.
Step 7: optimize
Run the kvb-optimize skill with this run's findings as its evidence, so none
of its skills re-collects them; its kvb-transcend pass carries every
in-scope session's scripts, tests and skills into their repositories.
Step 8: clean up
The kvb-cleanup skill for the disk, /tmp, Docker and agent directories the
window filled, with the finished tasks' artifacts (rasci.py artifacts --ended) as candidates; then the kvb-git-branch-gc skill for every
in-scope repository's merged branches and worktrees. Each keeps its own
confirmations; the dream never writes, deletes or commits for them. Close
with date +%s > ~/.claude/.dream-stamps/"$CLAUDE_CODE_SESSION_ID", so the
next hint is twelve hours out.
Rules
- Read-only until step 6: the dream writes memory; repository code changes only on an explicit radio answer.
- An empty window is valid: report "nothing in the last 24 hours" per repository and skip it; never invent activity.
- Report the whole ledger: held, broke, unknown, surviving findings, refuted hypotheses with why they fell.