cavekit · claude code plugin

One file survives the reset. Everything else is a command.

Spec-driven development over a single durable SPEC.md, written in caveman encoding — about 75% fewer tokens than the same spec in prose. Three commands run the loop. Failures land in §B; the bug classes that repeat harden into §V.

Frozen since August 2026 · still installable and working

Source

MIT · JavaScript · JuliusBrussee/cavekit

01Problem

The context resets and takes the plan with it. So the plan lives in a file.

Halfway through a change, the window fills. The agent forgets the constraint you agreed on two hours ago, re-litigates a decision, and reintroduces the bug you already fixed.

Cavekit keeps the whole agreement in one file the agent re-reads on every pass — and lets each failure make that file stricter.

  • /ck:specwrite the spec, name the invariants
  • /ck:buildexecute it, run the tests
  • /ck:checkdrift report, failures back to §B
02Loop

A failure enters as a bug. If it keeps coming back, it leaves as a rule.

SPEC.mdillustrative · four beats of one loop
./SPEC.mdspec written
§I intent
  • ship signup that never trusts the request body
§B bugs
§V invariants
  • 1org id comes from the verified claim, never the body
  • 2email is validated before any db write
no run yet

One file, three commands. A failure enters §B with the case that produced it; a bug class that keeps returning is written into §V and checked from then on.

03Inside

What shipped in the box. All of it still runs.

  1. 01

    One durable SPEC.md that survives every context reset. The agent re-reads it, not you.

  2. 02

    Caveman encoding for the spec itself — about 75% fewer tokens than the same thing in prose.

  3. 03

    Every test failure becomes a §B bug entry, with the case that produced it.

  4. 04

    Bug classes that keep coming back harden into §V invariants, checked from then on.

  5. 05

    Four optional commands when a change earns them: /ck:grill, /ck:research, /ck:review, /ck:deepen.

  6. 06

    Nine installed skills. No sub-agents, no dashboards, no orchestration binaries.

Frozen, not deleted.
The loop still runs.

Frozen since August 2026 · still installable and working

No longer in active development. Nine skills, three commands and one file — installed the same way it always was.