v3.1.0 — Kit-level file scoping for team claims

Highlights

Kit-level file footprints. Build-site tables can now declare an optional trailing Files column listing the globs a task is expected to touch. The scheduler uses it for real path-overlap detection, and cavekit team claim auto-defaults --paths to those globs — no more silent "user forgot --paths" failure mode.

Install fix. The bash cavekit wrapper now delegates team, version, debug, reset, status, and monitor subcommands to the Go binary (cavekit-bin), which install.sh now builds into ~/.cavekit/bin. Previously cavekit team init failed on fresh installs.

Using Files

## Tier 1 — Depends on Tier 0

| Task  | Title    | Spec    | Requirement | blockedBy | Effort | Files                         |
|-------|----------|---------|-------------|-----------|--------|-------------------------------|
| T-003 | DB layer | spec.md | R2          | T-001     | M      | internal/db/**, migrations/** |

Then:

cavekit team claim T-003   # paths default from the kit's Files column

Comma-separated globs (or semicolon-separated if any pattern contains a comma). Kits without the column keep working — the scheduler falls back to the prior heuristics.

Changes

• internal/site: Task.Files plumbed through parser; trailing Files column parsed for tier 0 and tier N.
• internal/team/scheduler: real PathsOverlap check when the task declares Files; legacy substring fallback retained.
• internal/team/manager: Claim defaults --paths to Task.Files when omitted.
• install.sh + scripts/cavekit: build and delegate to cavekit-bin so cavekit team … works after install.
• README: new Teams section covering setup, the Files column, and day-to-day commands.
• Tests: parser coverage for the new column; Claim_DefaultsPathsFromSiteFiles asserts the ledger records kit-derived paths.

Requires Go 1.21+ at install time to build the team CLI.

v3.0.0 — Simplified Hunt cycle (25 → 12 commands)

Summary

Cavekit v3.0.0 consolidates the command surface from 25 commands to 12 while preserving every capability. The core cycle is unchanged and now fits on one line:

/ck:sketch → /ck:map → /ck:make → /ck:check

One shortcut: /ck:ship "<description>" runs all four end-to-end for tiny features.

Command Consolidation

Deleted (14): architect, build, draft, inspect (legacy aliases), backprop, scan, judge, watch, progress, setup-tools, quick, review-branch, codex-review, gap-analysis.

New (3):

• /ck:ship — one-shot sketch → map → make → check, no gates. For throwaways.
• /ck:review — unified branch review with --mode gap / --codex / --tier / --strict flags. Absorbs scan, judge, review-branch, codex-review, gap-analysis.
• /ck:status — live + point-in-time progress with --watch. Absorbs watch, progress.

Extended with flags:

• /ck:revise --trace — single-failure six-step backpropagation protocol (absorbs /ck:backprop).
• /ck:init --tools-only — capability re-detection with recommendations (absorbs /ck:setup-tools).

All 14 deleted commands were hard-deleted — no deprecation shims. Migrations are noted in /ck:help.

Retained commands polished

Every retained core command (sketch, map, make, check) gained a two-line "What this does / When to use it" header and a Next: pointer to the following phase. Deprecation notes from earlier versions stripped throughout.

Skill Consolidation

23 → 21 skills:

• skills/backpropagation folded into skills/revision (new §8 "Automated Backpropagation (Single-Failure Trace)").
• skills/peer-review-loop folded into skills/peer-review (new "Codex Loop Mode" section).

No capability lost — the revision skill now describes both the multi-commit sweep and the single-failure trace, and peer-review describes both the six review modes and the Cavekit+Ralph+Codex loop pattern.

Documentation

• README.md fully rewritten. First 30 lines show the mental model, install, and a greenfield first-run narrative. Runtime internals, Codex review modes, skills, and methodology moved into collapsed <details> blocks.
• help.md rewritten for the 12-command surface with a migration table.
• All /bp:* aliases removed from prose.

Reference Audit

Every dangling command and skill reference cleaned up across scripts/, hooks/, agents/, install.sh, skills/, .claude-plugin/marketplace.json, .claude-plugin/plugin.json, and the dog-food context/ kits, plans, and impl files.

Version

plugin.json bumped 2.2.0 → 3.0.0. Marketplace metadata updated. Keywords backpropagation → revision.

Upgrade

cd ~/.cavekit && git pull

Restart Claude Code afterwards. Existing projects keep working — the Hunt lifecycle (sketch/map/make/check) is untouched. If you had muscle memory for a deleted command, /ck:help shows the new location.

v2.0.0 — CaveKit

🪨 Blueprint is now CaveKit

Complete rebrand from Blueprint (formerly SDD) to CaveKit — part of the Caveman ecosystem.

What changed

• Project identity: Blueprint → CaveKit. New name, new README, caveman-themed tone throughout
• CLI namespace: all /bp:* commands now available as /ck:* (/ck:sketch, /ck:map, /ck:make, /ck:check, etc.)
• Agent namespace: all bp:* subagent types renamed to ck:* (ck:drafter, ck:architect, ck:builder, etc.)
• Caveman token compression: integrated the Caveman skill for ultra-compressed communication that cuts token usage ~75%
• Draft mode: added draft permission mode for spec-writing phases
• Ecosystem links: cross-linked with Caveman and Revu projects

Migration

If you had /bp:* commands in workflows, switch to /ck:*. Both prefixes currently work — bp will be removed in a future release.

Install

claude mcp add cavekit -- npx -y github:JuliusBrussee/cavekit

v1.0.0

SDD — Spec-Driven Development
Claude Code plugin + tmux dashboard for spec-driven development with automated iteration loops.