0 Fork 0

Files

T

Kevin Veen-BirkenbachandClaude Opus 5.5 3c2f6c44b5 refactor(skills): cap every SKILL.md at 600 words and cut agent messages

Every loaded SKILL.md and every listed description is paid on every turn of every session, and every delivered message costs its recipient a full turn. scripts/budget.py gains SKILL_WORDS = 600, enforced by test_lock, and test_naming forbids a skill naming its own slash command; 53 descriptions and 10 body lines drop it. The 18 longest skills move details into 31 side files read only in their situation (organize's capacity, succession, peers, offices, relay and lineage; cleanup's docker, worktrees, execute and ask; code-tune, changelog, triage, dream, split and others), so all SKILL.md files shrink from 33866 to 26291 words without dropping a rule; an independent review of HEAD against the new text found no loss in 60 of 64 skills and three weakenings, now restored (WSJF re-estimates after an objective change, cleanup's volume checklist, essence's absolute dates).

The kvb-org-organize Peer protocol sends status that asks nothing to the news board, answers nothing that asks nothing, keeps messages terse, and lets light runs admit themselves in the operator's wording; the skills follow it (operator reads the register, relax and wakeup reply through the news, evening collects them there, wsjf asks only for missing estimates, cleanup and triage batch their questions). Skills that explore code ask the codebase-memory-mcp graph first, subagent briefs scale with the stakes and cap their answers, and the skills that need the full register or council tables call show --wide or list --long.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2026-11-04 17:07:23 +01:00

3.4 KiB

Raw Blame History

Tune: working a candidate

Before forming it

Search every branch (the kvb-git-merge-fix skill's find_fix.py -S <symbol> --path <path>) for a change that already removes or reshapes the step. A fix on a sibling branch can make the whole candidate worth zero.
A step that is long by design (a backup and restore drill, an end-to-end suite) is not a candidate. Its cost buys its guarantee.
Apply the session's saving floor: a smaller saving is listed as "below floor" and dropped.
Pruning build caches to free disk costs a cold rebuild of every custom image on the next run. Name that cost before anyone prunes.

Levers that usually pay:

Batch per-item process spawns: N calls that each boot an interpreter or an application console become one process running N commands (one app console boot, one module, one batched API request). Keep each item's failure named and the exit code honest.
Skip no-op work a previous step already did, such as a separate enable after an install that already enabled.
Idempotency flip-flops: any changed on a second consecutive run.
Image pulls and builds: pull only what the host lacks; keep build caches warm.
Overlap waits with work only where concurrent execution is proven safe. Schema migrations and lock-holding steps are usually not.

Delegating it

Hand every optimization to an agent through the kvb-org-split skill: one sidetree per candidate, branched from an exact commit, with a time cap and "measure first, report the time split before implementing". One agent per candidate, never more; the brief names the step, its log line and number, the files in scope and the claim to prove, and caps the report (for example 15 lines: time split, change, before and after). The agent probes live systems read-only and never starts gates or deploys. A broad brief without a cap can run for hours without a word; at its cap, collect the agent's interim findings and let it run on.

Before merging, run independent skeptics through the kvb-think-dialectic skill. The antithesis must try to show that the candidate removes coverage, hides a failure, or only moves the cost somewhere the measurement does not look. Merge what reaches the bar; submit the rest as proposals.

Re-measuring it

Re-measure from the same source under the same conditions and report a before and an after number. Discard a before/after pair, not just its noise, when:

Another gate, deploy or heavy job shared the machine during either run.
The two runs differ in image or cache state. A cold build after a cache prune or a Dockerfile edit is expected: before calling a cold build a defect, compare the instruction text and the base image digest across both logs.
One run pays a one-time cost (first install, migration, purge, model download) the other does not.
A long wait hides retries or a failure; count retry lines before calling it slow.
A fresh-install pass is compared with an update pass. Analyse each pass of a multi-pass run separately.

Reporting it

Hand the prototypes and before/after harnesses of every kept candidate to the kvb-code-transcend skill, so the measurement stays in the repository as a benchmark or regression test. A rejected candidate stays in the output table with the reason it fell: below floor, coverage cut, outside the session's limits, invalid comparison, made moot by another branch, or cap reached without a measured result.

3.4 KiB Raw Blame History