caveman-internal
Internal token-compression protocol for Cavekit agent artifacts — handoff notes, artifact summaries, context bundles, review notes, status lines. NOT the user-facing /caveman skill; that one compresses assistant replies to the user. This skill compresses machine-to-machine prose so the next agent in the loop reads fewer tokens. Three intensities (lite / full / ultra) with automatic selection based on budget pressure. Used by task-builder, builder, inspector, verifier, convergence-monitor, and the stop-hook status block.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- caveman-internal
- description
- Internal token-compression protocol for Cavekit agent artifacts — handoff notes, artifact summaries, context bundles, review notes, status lines. NOT the user-facing /caveman skill; that one compresses assistant replies to the user. This skill compresses machine-to-machine prose so the next agent in the loop reads fewer tokens. Three intensities (lite / full / ultra) with automatic selection based on budget pressure. Used by task-builder, builder, inspector, verifier, convergence-monitor, and the stop-hook status block.
02 资源文件 1
- 参考文档references/budget-thresholds.md1.8 KB
03 SKILL.md 正文
Caveman Internal Protocol
Scope
IN-scope (compress)
- Artifact summaries written to
.cavekit/artifacts/ - Context bundles between tasks (
.cavekit/context-bundles/) - Review notes and gap lists (non-user-facing)
state.mdnotes:field andloop-log.mdentries- Status-block body (dashboard injected by the stop hook)
- Inter-agent handoff memos
OUT-of-scope (never compress)
- Source code (any language, any file)
- Git commit messages and PR descriptions
- Kits (specs), build sites, DESIGN.md
- Error messages that a human will act on
- Security warnings and deprecation notices
- Structured findings tables (P0/P1/P2/P3, coverage matrices)
- Regression test names and assertion messages
Compressing out-of-scope content is a bug. If a reader cannot act on a compressed artifact, regenerate it verbose and log a caveman fallback — see "Fallback protocol" below.
Intensity levels
Lite — trim the fat (~20% savings)
Professional tone. Grammar intact. Drop filler, pleasantries, hedging. Keep articles and full sentences.
Rate-limit middleware passes unit tests. Integration failed on 429 header case sensitivity. Fix in progress.
Full — default (~40% savings)
Classic caveman. Drop articles, filler, pleasantries. Fragments fine.
Technical terms stay exact. Pattern: [thing] [action] [reason]. [next].
Rate-limit mw pass unit. Integration fail on 429 header case. Fixing.
Ultra — maximum grunt (~65% savings)
Telegraphic. Abbreviate common terms (DB, auth, config, req, res, fn, impl, mw, mcp). Arrow notation for causality. One-word answers when enough.
mw ok, int fail header case → fix
Automatic intensity selection
Consult the session budget ledger before writing an internal artifact:
| Budget remaining | Intensity |
|---|---|
| > 50 % | lite |
| 20 – 50 % | full |
| < 20 % | ultra |
Override rules:
depth = thoroughtasks clamp to lite regardless of pressure (accuracy matters more than tokens when the stakes are high).phase = inspectingclamps to lite for the gap-analysis output.- Security-sensitive artifacts clamp to lite.
- If the caller sets
caveman_intensityexplicitly in the artifact envelope, use that and skip auto-selection.
See references/budget-thresholds.md for the detailed decision table.
Fallback protocol
If a downstream reader cannot reconstruct intent from a compressed artifact:
- Regenerate the artifact in full prose (no compression).
- Append a one-line entry to
state.md's "Caveman fallbacks" section:- 2026-04-17T12:34Z T-017 verifier fallback: artifact summary re-expanded - If the same artifact category fails 3+ times in one session, clamp that
category to
litefor the rest of the session and log tostate.md.
Envelope
Every compressed artifact carries a tiny header:
<!-- caveman: intensity=full version=1 -->
Readers that see this header know the body is compressed and apply the reverse-expansion grammar (re-add articles, re-expand abbreviations) when rendering for a human.
Examples
| Verbose | Full | Ultra |
|---|---|---|
| "The build completed successfully after three waves." | "Build complete. 3 waves." | "Build ok 3w" |
| "All tests pass, but one test was skipped due to a missing fixture." | "Tests pass. One skipped, missing fixture." | "Tests ok 1 skip fx" |
| "The reviewer flagged two critical issues in the auth middleware." | "Reviewer: 2 P0 in auth mw." | "Rev 2 P0 auth mw" |
Integration with the existing caveman skill
The user-facing caveman skill governs how the assistant speaks to the user
during a session (answers, explanations, replies). This caveman-internal
skill governs how agents write artifacts for each other. Both must be on
or both off for a coherent experience, so caveman_mode = off in config also
disables this skill.
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.