Skip to content
Back to search

caveman

MarketplaceWorks with Claude CodeGitHub

Ultra-compressed communication mode. Shortens prose while preserving technical terms, code, commands, and exact errors.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026. Its scripts were only partly scanned.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/JuliusBrussee/caveman with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/JuliusBrussee/caveman'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/juliusbrussee-caveman/badge.svg)](/listings/juliusbrussee-caveman)

HTML

<a href="/listings/juliusbrussee-caveman"><img src="/api/listings/juliusbrussee-caveman/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

These steps run a script from the network that findsafeskills does not scan.

npx skills add JuliusBrussee/caveman -g

Works in Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Copilot, and [30+ more](./INSTALL.md). Type `/caveman` if it doesn't start on its own. Say `stop caveman` to go back. One rock. That it.

<details> <summary><strong>Other ways in</strong>: Claude Code plugin, Gemini, every agent at once, Windows, uninstall</summary>

<br>

# Claude Code plugin, auto-starts every session
claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

# Gemini CLI
gemini extensions install https://github.com/JuliusBrussee/caveman

# Every agent on your machine at once, plus the Claude Code statusline badge (Node.js 22.13+)
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v3.1.0/install.sh | bash

Windows, PowerShell 5.1+:

irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v3.1.0/install.ps1 | iex

Changed your mind: `npx -y github:JuliusBrussee/caveman -- --uninstall`

Install broke? Open your agent in this repo and say *"Read CLAUDE.md and INSTALL.md, install caveman for me."* Agent fix own brain.

</details>

About

Found in: claude-skills-marketplace

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 17 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. {
  2. "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
  3. "name": "caveman",
  4. "description": "Ultra-compressed communication mode. Shortens prose while preserving technical terms, code, commands, and exact errors.",
  5. "owner": {
  6. "name": "Julius Brussee",
  7. "url": "https://github.com/JuliusBrussee"
  8. },
  9. "plugins": [
  10. {
  11. "name": "caveman",
  12. "description": "Talk like caveman. Cut filler. Keep technical details explicit.",
  13. "source": "./",
  14. "category": "productivity"
  15. }
  16. ]
  17. }
README.md · README · 382 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. <div align="center">
  2. ## New: Caveman Code. Free.
  3. **21% less than Codex per solved task on Terminal-Bench 4.** Same model on both sides.
  4. <table>
  5. <tr>
  6. <td align="center"><h3>21% cheaper</h3>than Codex<br><sub>Terminal-Bench 4, per solved task, GPT-6.1 Sol: $1.21 vs $1.52, 90 of 192 runs solved both</sub></td>
  7. <td align="center"><h3>Level</h3>with Claude Code within noise<br><sub>Terminal-Bench 4, Opus 5.5: 97 of 180 runs solved vs 83 of 177; $5.19 vs $6.04 per solved task, 95% interval 0.70x to 1.02x</sub></td>
  8. </tr>
  9. </table>
  10. <sub>Measured build, benchmark settings on, not the shipped app: the app ships the same harness with three benchmark-only settings turned off. Same model on both sides: GPT-6.1 Sol against Codex, Opus 5.5 against Claude Code. Our own runs, November 2026, at list prices. <a href="https://caveman.so/code#benchmarks">How we measured</a></sub>
  11. Every coding agent in one window. Free on the subscriptions and keys you already have.
  12. **[Download Caveman Code for Mac](https://caveman.so/download) · [Read the launch post](https://caveman.so/news/introducing-caveman-code)**
  13. ---
  14. <img src="docs/assets/caveman-logo-banner.png" alt="Caveman" width="720">
  15. # why many token when few do trick
  16. **Caveman make your AI agent say less and read less. Code stay exact. Brain still big.**
  17. <a href="https://github.com/JuliusBrussee/caveman/stargazers"><img src="https://img.shields.io/github/stars/JuliusBrussee/caveman?style=flat-square&color=F0A63C&label=stars" alt="GitHub stars"></a>
  18. <a href="https://www.npmjs.com/package/@caveman-ai/cli"><img src="https://img.shields.io/npm/dm/@caveman-ai/cli?style=flat-square&color=F0A63C&label=cli%20downloads" alt="npm downloads"></a>
  19. <a href="./INSTALL.md"><img src="https://img.shields.io/badge/works_with-30%2B_agents-orange?style=flat-square" alt="30+ agents"></a>
  20. <a href="#license"><img src="https://img.shields.io/badge/license-Apache--2.0-green?style=flat-square" alt="License"></a>
  21. <table>
  22. <tr>
  23. <td align="center" width="25%"><h3>33.2% fewer</h3>input tokens, whole sessions<br><sub><a href="#read-less-332-fewer-input-tokens">6 synthetic large-tool-output tasks, 18 pairs</a>, Aug 2026 engine, 18 of 18 answers right</sub></td>
  24. <td align="center" width="25%"><h3>0.46×</h3>the cost of Opus at xhigh, with Auto<br><sub><a href="#auto-046x-the-cost-of-opus-at-xhigh-interim">85 coding tasks</a>, interim, score gap ±7 points, unresolved</sub></td>
  25. <td align="center" width="25%"><h3>397 → 76</h3>tokens for one test run<br><sub><a href="#blocks-code-mode-script-written-once">Blocks</a> vs <code>| tail -20</code></sub></td>
  26. <td align="center" width="25%"><h3>1.4 to 2.4× cheaper</h3>with caveman-style output<br><sub><a href="https://arxiv.org/abs/2606.24083">Adobe Research</a>, eight models</sub></td>
  27. </tr>
  28. </table>
  29. Cited by **[Adobe Research](https://arxiv.org/abs/2606.24083)** · A/B tested by **[JetBrains](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/)** · Remade for Elasticsearch on **[Elasticsearch Labs](https://www.elastic.co/search-labs/blog/elastic-caveman-ai-token-reduction)**<br>
  30. **#1** on [Hacker News](https://news.ycombinator.com/item?id=47647455) · **#1** on GitHub Trending · *"No way this actually works."* [ThePrimeagen](https://www.youtube.com/watch?v=L29q2LRiMRc)
  31. **[Install](#install) · [Six rocks](#six-rocks-one-install) · [Proof](#proof) · [Say less](#say-less-the-caveman-voice) · [Read less](#read-less-332-fewer-input-tokens) · [Auto](#auto-046x-the-cost-of-opus-at-xhigh-interim) · [Blocks](#blocks-code-mode-script-written-once) · [Commands](#commands) · [Caveman Cloud](#caveman-cloud)**
  32. </div>
  33. ---
  34. <table>
  35. <tr>
  36. <th width="50%">Normal agent · 63 tokens</th>
  37. <th width="50%"><img src="docs/assets/dancing-rock.svg" width="18" height="18" alt=""> Caveman agent · 20 tokens</th>
  38. </tr>
  39. <tr>
  40. <td valign="top">
  41. > The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle. When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I'd recommend using useMemo to memoize the object.
  42. </td>
  43. <td valign="top">
  44. > New object ref each render, so React re-renders. Wrap the prop in `useMemo`.
  45. </td>
  46. </tr>
  47. </table>
  48. **Same fix. 63 token become 20. Brain still big.**
  49. Pick your club:
  50. | Skill | Same answer | Tokens |
  51. |---|---|---:|
  52. | `/caveman` | New object ref each render, so React re-renders. Wrap the prop in `useMemo`. | 20 |
  53. | `/ultracave` | Inline object prop, new ref, re-render. `useMemo`. | 14 |
  54. | `/megacave` | 新參照致重繪。`useMemo`。 | 13 |
  55. <sub>Token counts: tiktoken o200k.</sub>
  56. ## Install
  57. ```bash
  58. npx @caveman-ai/cli@latest
  59. ```
  60. One command. Mac, Linux, Windows. Caveman find your agents and your logins, show what it change. Press Enter. Done. Needs Node.js 22.13+.
  61. After that, plain `claude` stays caveman'd. So do Codex, Gemini CLI, OpenCode, and [30+ more](./INSTALL.md). It runs on your machine, with your keys and your Claude Pro/Max login. One rock. That it.
  62. <sub>Had caveman before 4.0? Say <code>@latest</code>, or the old copy on your machine wins. Already set up? <code>caveman setup</code> opens the screen again. Only want shorter answers? <a href="#say-less-the-caveman-voice">Get just the skill</a>.</sub>
  63. ## Six rocks, one install
  64. New in 4.0: caveman is six parts. All on after install. Each has its own switch.
  65. | Rock | What it does | Measured |
  66. |---|---|---|
  67. | **[Say less](#say-less-the-caveman-voice)** | Agent answers short. Code stays exact | 3% shorter than "Answer concisely." (ultracave 35%), n=10 |
  68. | **[Read less](#read-less-332-fewer-input-tokens)** | Logs, JSON, CSV, test output shrink before the model reads them | **33.2%** fewer input tokens per session (6 synthetic large-tool-output tasks, 18 pairs, Aug 2026 engine) |
  69. | **[Auto](#auto-046x-the-cost-of-opus-at-xhigh-interim)** · new | Picks the model and the effort for each ask | 0.46× the cost of Opus at xhigh in an interim run (85 tasks); quality gap unresolved |
  70. | **[Blocks](#blocks-code-mode-script-written-once)** · new | Code mode: agent keeps the scripts it writes and runs them again | **397 → 76** tokens for one test run |
  71. | **[Waste fixes](#smaller-rocks)** | Finds where your tokens go and fixes it. New: keeps the prompt cache warm while you pause | Not measured yet |
  72. | **[Browse](#smaller-rocks)** | Web pages come back compressed | 14.9% smaller full page than Playwright; 121 vs 15,704 tokens with a query, on one 200-row table |
  73. `caveman status` shows what is on. `caveman off browse` turns one off. Only Auto needs an account, and it is free. [4.0.0 release notes](https://github.com/JuliusBrussee/caveman/releases/tag/v4.0.0)
  74. ## Proof
  75. Outside labs first. Ours come with each rock below. Nothing rounded up. Red rows stay red.
  76. | Who | Setup | Result |
  77. |---|---|---|
  78. | **[Adobe Research](https://arxiv.org/abs/2606.24083)**, CAVEWOMAN paper (cites this repo) | Caveman-style output, eight models, five datasets | **Cost cut 1.4 to 2.4× per model, up to 3×** |
  79. | **[Elastic](https://www.elastic.co/search-labs/blog/elastic-caveman-ai-token-reduction)**, Elasticsearch Labs | Caveman mode remade for Elasticsearch, eight live MCP scenarios | **63.6% fewer response tokens.** *"Zero information loss."* |
  80. | **[JetBrains](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/)** | 86 real coding tasks, paired A/B | **No measurable quality loss** (p = 0.82). 8.5% fewer output tokens |
  81. Two findings shaped caveman. Adobe: cavemanning *your* prompt makes answers worse, so caveman never touches your prompt. JetBrains: agent sessions are mostly code and tool calls, which the voice leaves alone. So caveman grew rocks that shrink what the agent *reads* and what it *runs on*.
  82. ## Say less: the caveman voice
  83. **A voice, not broken grammar. Every reply follows the same structure.**
  84. | Rule | What it means |
  85. |---|---|
  86. | **Answer first** | `[thing] [action] [reason]. [next step].` No greeting, no "let me", no recap |
  87. | **One idea per sentence** | Built on [ASD-STE100](https://www.asd-ste100.org/), the controlled English of aircraft maintenance manuals: 20 words max, active voice, one term per thing |
  88. | **Meaning never dropped** | Articles can go. *not*, *never*, *no*, *only* never go. Numbers and units stay exact |
  89. | **Payload verbatim** | Code, commands, paths, and error messages untouched, character for character |
  90. | **No theater** | No "me think". No chatter between tool calls. If caveman phrasing isn't shorter, plain wins |
  91. | **Knows when to stop** | Security warnings, irreversible actions, and confused users get full sentences. Then grunt resumes |
  92. | **Your prompts stay yours** | Never rewritten. [Research say that backfire](#proof) |
  93. Example run, n=10, not a claim. Ten dev questions on claude-opus-5-5:
  94. | Instruction | Output tokens |
  95. |---|---:|
  96. | None | 6,983 |
  97. | `Answer concisely.` | 4,334 |
  98. | `/caveman` | 4,119 |
  99. | `/ultracave` | **2,693** |
  100. "Be concise" is the real baseline; new models already know it. On top of it, `/caveman` cuts 3% more at the median and `/ultracave` 35% more. Measured on the 3.1.0 rules, not yet re-run. The rules cost about 1,160 tokens of input, and if you pay per request (GitHub Copilot), shorter answers save nothing. [Harness](./evals/README.md) · [Every case where caveman loses](./docs/HONEST-NUMBERS.md)
  101. <!-- BENCHMARK-TABLE-START -->
  102. <!-- BENCHMARK-TABLE-END -->
  103. **Only want this rock?** The skill alone, no proxy:
  104. ```bash
  105. npx skills add JuliusBrussee/caveman -g
  106. ```
  107. <a href="https://skills.sh/JuliusBrussee/caveman"><img src="https://skills.sh/b/JuliusBrussee/caveman" alt="skills.sh"></a>
  108. Type `/caveman` (`$caveman` in Codex) if it doesn't start on its own. Say `stop caveman` to go back.
  109. <details>
  110. <summary><strong>Other ways in</strong>: Claude app and phone, Claude Code plugin, Gemini, every agent at once, Windows, uninstall</summary>
  111. <br>
  112. Claude on the web, the phone app, or Cowork? No terminal. Add caveman to your Claude account once, in Customize > Plugins: [steps](./INSTALL.md#claude-apps-claudeai-desktop-mobile-cowork). Not tested end to end yet.
  113. ```bash
  114. # Claude Code plugin, auto-starts every interactive session, subagents too
  115. claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman
  116. # Gemini CLI
  117. gemini extensions install https://github.com/JuliusBrussee/caveman
  118. # Every agent on your machine at once, plus the Claude Code statusline badge (Node.js 22.13+)
  119. curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.1.0/install.sh | bash
  120. ```
  121. Already inside Claude Code? Same rock, slash form. Type these where you talk to Claude:
  122. ```text
  123. /plugin marketplace add JuliusBrussee/caveman
  124. /plugin install caveman@caveman
  125. ```
  126. Windows, PowerShell 5.1+:
  127. ```powershell
  128. irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.1.0/install.ps1 | iex
  129. ```
  130. On npm 12 or newer, new npm block git install. One-liners above handle it themselves from v3.2.0 on. Old one-liner pinned to v3.1.0 or older still fail there: use one above, or this:
  131. `npx --allow-git=root -y github:JuliusBrussee/caveman#v4.1.0`
  132. Changed your mind: `npx -y github:JuliusBrussee/caveman -- --uninstall` (on npm 12+, add `--allow-git=root` too)
  133. Only grabbed the skill? `npx skills remove JuliusBrussee/caveman -g` take it back out.
  134. Install broke? Open your agent in this repo and say *"Read INSTALL.md, install caveman for me."* Agent fix own brain.
  135. </details>
  136. ## Read less: 33.2% fewer input tokens
  137. **The voice shrinks what the agent says. This shrinks what it reads. Same answers: 18 of 18 right.** 6 synthetic large-tool-output tasks, 18 pairs, Aug 2026 engine.
  138. | File type | The file, through caveman | File saved | Whole Claude Code session, 3 runs | Session saved |
  139. |---|---:|---:|---:|---:|
  140. | CSV | 28,041 → **314** | **98.9%** | 165,823 → 74,484 | **55.1%** |
  141. | Logs | 22,810 → **348** | **98.5%** | 148,807 → 74,068 | **50.2%** |
  142. | YAML | 20,447 → **178** | **99.1%** | 132,124 → 71,027 | **46.2%** |
  143. | Test output | 18,806 → **203** | **98.9%** | 150,377 → 108,514 | **27.8%** |
  144. | JSON | 18,837 → **281** | **98.5%** | 147,975 → 108,939 | **26.4%** |
  145. | HTML | 21,670 → 21,670 | none yet | 140,687 → 154,641 | 9.9% worse |
  146. | **All six** | 130,611 → 22,994 | 82.4% | **885,793 → 591,673** | **33.2%** |
  147. *The file* is each test file through today's compressor alone. *The session* is the whole agent run, which also carries the system prompt and tools, so it moves less. Honest part: the session run is from August 2026 on an older engine, HTML has no compressor yet so it lost, and each test file hides one record in 61 to 72 KB of noise. Yours will vary. Headroom on the same suite: 15 of 18 right, 6.7% fewer tokens (interval crosses zero; harness not public). [Method](./docs/WRAP-BENCHMARK.md#per-file-compression)
  148. Every original stays on your disk. The agent can pull it back any time.
  149. ## Auto: 0.46x the cost of Opus at xhigh (interim)
  150. **New model in your agent's picker. Pick Auto and caveman chooses the model and the effort for each ask. Hard ask, big brain. Easy ask, save rocks.**
  151. | Setup | Tasks solved | Cost per try | Median time |
  152. |---|---:|---:|---:|
  153. | Opus 5.5, xhigh effort, every ask | 44.3% | $7.83 | 35.2 min |
  154. | **Auto**, Claude models only | **43.7%** | **$3.59** | **15.9 min** |
  155. | **Auto**, plus OpenAI and open models | 42.9% | $3.47 | 16.7 min |
  156. Auto costs 0.46× what Opus 5.5 at xhigh costs. That is 54% saved, in under half the time. 85 real coding tasks from SWE-Interact and SWE-Together, one try each. Honest part: the run is half done, the 0.5-point score gap is too small to tell apart at this size (±7), and most of the saving is Auto lowering effort on easy asks. Already on Opus at medium? Expect less; not measured yet. [Full table and limits](./docs/AUTO-BENCHMARK.md)
  157. ```bash
  158. caveman login # free account. Auto shows up in the model picker
  159. caveman providers add openai # optional: give Auto more models, on your own key
  160. ```
  161. Works in Claude Code, OpenCode, and Codex on a ChatGPT login, with API keys, Claude Pro/Max, or ChatGPT. Any other model you pick runs as picked. Not signed in: no Auto, nothing sent. [What Auto sends](#privacy)
  162. ## Blocks: code mode, script written once
  163. **skills.sh for code mode.** Code mode is the agent writing a script instead of making tool call after tool call ([Cloudflare](https://blog.cloudflare.com/code-mode/), [Anthropic](https://www.anthropic.com/engineering/code-execution-with-mcp)). Problem: agent write same script every session, throw it away, write it again. Blocks keep it in your repo, check it every pull request, hand it back.
  164. | Input | Raw | What agents type now | Block |
  165. |---|---:|---:|---:|
  166. | `go test -v ./...`, 326 tests | 10,561 | 397 · `\| tail -20` | **76** · `test-summary` |
  167. | 1.4 MB JSON file | 251,893 | 236 · their own heredoc | **227** · `json-peek` |
  168. | 366-line Go file | 3,111 | **250** · `grep -n` | 489 · `grep-defs` |
  169. Tokens the agent reads back. In one week, agents on one machine wrote the same script [394 times](https://github.com/caveman-ai/blocks#one-script-written-394-times-in-a-week). [Purdue](https://arxiv.org/abs/2609.30725) measured the habit at up to 9.57% of task cost. `grep-defs` loses to a bare grep; row stays. [Method](https://github.com/caveman-ai/blocks#blocks-against-what-agents-type-now)
  170. Comes with caveman. Eight blocks ready: `json-peek` · `jsonl-stats` · `test-summary` · `first-error` · `wait-for` · `grep-defs` · `http-json` · `replace-in-file`. When the agent starts typing a script a block already covers, it gets one line: use the block. Never denies or rewrites a command.
  171. ```bash
  172. caveman-blocks scan # what your agents retyped this month. Script text never leaves your machine
  173. caveman-blocks init # rules and block index into this repo's AGENTS.md
  174. ```
  175. Pre-release (v0.1.0-rc.5). Python only, no sandbox yet. Own repo, own stars: **[caveman-ai/blocks](https://github.com/caveman-ai/blocks)**
  176. ## Smaller rocks
  177. | Rock | What it does | Measured |
  178. |---|---|---|
  179. | `caveman learn` | Ranks where your tokens go, from agent history already on disk. `caveman learn implement` applies fixes one diff at a time, only on your yes | Not yet |
  180. | Warm cache · new | Step away and the provider forgets your session, so the next ask pays full price. Caveman keeps Anthropic's prompt cache warm while you pause, only when your own history says you come back in time. Every warm is listed in `caveman stats` | Not yet |
  181. | `caveman browse <url>` | A compressed web page for the agent | **121 tokens instead of 15,704** for a Playwright snapshot of a 200-row table. Tiny forms lose 2.3×. [Bench](./browse/BENCHMARK.md) |
  182. | `/caveman-compress <file>` | Shrinks memory files like `CLAUDE.md`, keeps a backup | **4,138 tokens instead of 6,198** across five files, 22.8% to 49.1% each, every heading, code block, and path intact. [Bench](./skills/caveman-compress/README.md#benchmarks) |
  183. | Middleware | Same shrinking inside the Vercel AI SDK, LangChain, OpenAI, or Anthropic call you already make. [TypeScript](./packages/middleware/typescript/README.md) · [Python](./packages/middleware/python/README.md) · [Every framework](https://docs.caveman.so/caveman/middleware) · [One container for the team](docs/technical/deploy.md) | Not yet |
  184. <p align="center">
  185. <img src="docs/assets/learn-report.png" alt="Caveman Learn report: a summary and savings cards on the left; the biggest places tokens go, with one fix opened, and a chart of how full sessions get on the right" width="900">
  186. </p>
  187. ## Commands
  188. In your terminal:
  189. ```bash
  190. caveman claude # start an agent on caveman: codex · gemini · opencode · aider · kilo · qwen · hermes · openclaw · pi
  191. caveman status # what is on, which agents it reaches, one next step
  192. caveman on browse # switch one rock on or off: output · input · waste-fixes · routing · scripts · browse
  193. caveman login # free account, turns on Auto
  194. caveman learn # where your tokens go
  195. caveman trial -- claude # A/B a real session on your own work
  196. caveman shrink -- pnpm test # compress noisy command output
  197. caveman convert --dry-run # pixel mode: skills the model reads as images
  198. caveman stats # your token history
  199. caveman doctor # one problem per line, with its fix
  200. ```
  201. In your agent:
  202. | Command | What it does |
  203. |---|---|
  204. | `/caveman` · `/ultracave` · `/megacave` | The voice, the grunt, the 文言文. `/caveman status` shows the mode, `/caveman off` stops it |
  205. | `/caveman-commit` | One-line Conventional Commit |
  206. | `/caveman-review` | One finding per line: `L42: 🔴 null deref. Guard it.` |
  207. | `/caveman-compress <file>` | Shrinks memory files and backs up the original |
  208. | `/caveman-stats` | Real token usage for this Claude Code session |
  209. | `/caveman-help` | Every mode and command on one screen |
  210. | `cavecrew` | Subagents that find, edit, and review code, then report back in caveman |
  211. | `investigate-first` · `lean-build` · `surgical-patch` · `safe-refactor` · `migration` · `verify-and-stop` | Work patterns that write less code. Your agent picks them up when a task fits |
  212. | `caveman-setup` · `caveman-discover` · `caveman-learn` · `caveman-manage` · `caveman-optimize` · `caveman-explore` · `caveman-evidence-review` | Drive the proxy from inside your agent |
  213. ## Caveman Cloud
  214. **Everything above is for the coding agent you use. Caveman Cloud is for the agents you build.**
  215. You ship an agent. Every run costs money. Is it good? Is it cheap? Which change made it worse? Cloud answers.
  216. | Cloud does | What you get |
  217. |---|---|
  218. | **Observability** | Every run traced to the work and what it cost. Sets itself up from your repo, your keys, and the logs or OpenTelemetry you already have |
  219. | **Evals** | Grades how good the answers are, on your own tasks, from your real traffic |
  220. | **Autopilot** | Proposes a cheaper change, tests it on your recorded work, ships only when the evals say answers are still as good. You approve the pull request |
  221. | **Proof** | Every number says if it is measured, inferred, or verified. A saving counts only after real traffic shows it |
  222. <p align="center">
  223. <a href="https://app.caveman.so/signup"><b>Sign up free</b></a> &nbsp;·&nbsp; <a href="https://caveman.so/demo">Try the demo</a> &nbsp;·&nbsp; <a href="https://cal.com/caveman/chat">Book a call</a> &nbsp;·&nbsp; <a href="https://docs.caveman.so">Docs</a>
  224. </p>
  225. ## In the wild
  226. <a href="https://www.youtube.com/watch?v=L29q2LRiMRc"><img src="https://img.youtube.com/vi/L29q2LRiMRc/hqdefault.jpg" alt="ThePrimeagen reacts to Caveman: No way this actually works" width="340" align="right"></a>
  227. **ThePrimeagen**: *"No way this actually works."* [Watch](https://www.youtube.com/watch?v=L29q2LRiMRc)
  228. **Adobe Research**: [CAVEWOMAN](https://arxiv.org/abs/2606.24083), arXiv, June 2026
  229. **JetBrains**: *"It is fun, and it costs you nothing measurable in quality."* [Read](https://blog.jetbrains.com/ai/2026/07/speak-to-ai-agents-like-cavemen-tosave-tokens/)
  230. **Elastic**: [elastic-caveman](https://www.elastic.co/search-labs/blog/elastic-caveman-ai-token-reduction) on Elasticsearch Labs, April 2026
  231. **Hacker News**: [#1, 904 points](https://news.ycombinator.com/item?id=47647455)
  232. **GitHub Trending** #1 overall, July 2026 · **[Trendshift](https://trendshift.io/repositories/25391)** #1 repo of the day, April 2026 · **[Product Hunt](https://www.producthunt.com/products/caveman)** #8 of the day
  233. Started as a joke in April 2026. Now past 100,000 stars. Joke got serious. Voice did not.
  234. <br clear="right">
  235. [![Star History Chart](./docs/assets/star-history.png)](https://star-history.com/#JuliusBrussee/caveman&Date)
  236. ## Privacy
  237. - **The skill** runs on your machine and sends nothing. One exception: `/caveman-compress` sends the file you compress to your configured model provider.
  238. - **The CLI** sends usage stats by default: commands run, token counts, caveman-blocks commands and usage counts, a random install ID, OS, and IP. Never your prompts, code, file paths, or scripts you wrote. One person maintains this for free, and those stats show what to build next. Off for good, Blocks included: `caveman telemetry off` or `DO_NOT_TRACK=1`.
  239. - **Auto**, only once you sign in and pick it: each ask sends Caveman Cloud your latest ask (with whatever your agent attaches to it, such as CLAUDE.md, memory or @-mentioned files), the one before it, and the end of the agent's last reply, plus each request's tool names, effort, a few agent headers, and the previous request's token counts. On the Free plan Caveman may keep it to improve routing. `caveman off routing` stops it.
  240. - **Where requests go:** from your machine to your provider, or to another provider you added with `caveman providers add`, on your own key. Only a request routed to a Caveman Cloud model passes through Caveman Cloud, whole (system prompt, tool results, code), on your Caveman project's stored key, billed to that project. `caveman providers cloud off` stops that.
  241. The full list, and how to delete sent telemetry: [SECURITY.md](./SECURITY.md).
  242. ## License
  243. [Apache-2.0](./LICENSE), whole repo. Read it, fork it, ship it, host it. Free like mammoth on open plain. Older releases and third-party notices: [LICENSING.md](./LICENSING.md). "Caveman" and the rock logo are trademarks of Julius Brussee.
  244. ## Cite
  245. ```bibtex
  246. @software{brussee2026caveman,
  247. author = {Brussee, Julius},
  248. title = {Caveman: why many token when few do trick},
  249. year = {2026},
  250. url = {https://github.com/JuliusBrussee/caveman}
  251. }
  252. ```
  253. ---
  254. <div align="center">
  255. 🪨 **Caveman save you token. Star cost zero. Fair trade.**
  256. <sub>
  257. <a href="./docs/README.md">Docs</a> ·
  258. <a href="./INSTALL.md">Install matrix</a> ·
  259. <a href="./docs/HONEST-NUMBERS.md">Honest numbers</a> ·
  260. <a href="./CONTRIBUTING.md">Contributing</a> ·
  261. <a href="https://app.caveman.so/signup">Caveman Cloud</a> ·
  262. <a href="https://github.com/JuliusBrussee/caveman/issues">Issues</a>
  263. </sub>
  264. </div>
install.ps1 · Bundled script · 157 lines · 4 findings marked
  1. # caveman — installer shim (Windows / PowerShell).
  2. #
  3. # Thin wrapper around installer/install.js (the unified Node installer). Every flag
  4. # you'd pass to installer/install.js can be passed here; we just forward them.
  5. #
  6. # One-line install:
  7. # irm https://raw.githubusercontent.com/JuliusBrussee/caveman/v4.1.0/install.ps1 | iex
    • Note: Runs a downloaded PowerShell script (iex / Invoke-Expression)
    • Line 7 on GitHub
  8. #
  9. # Local clone:
  10. # pwsh install.ps1 [flags]
  11. #
  12. # After the installer it hands over to the CLI's first run (`caveman setup`:
  13. # modules, agents, one Continue) in an interactive console, and prints that
  14. # one command otherwise.
  15. #
  16. # Why a Node installer? install.sh + install.ps1 used to be parallel sources of
  17. # truth and constantly drifted (issue #249 was a `node -e "..."` quoting bug
  18. # that silently dropped the JSON merge step on every Windows install). One
  19. # Node script works everywhere without quoting bugs.
  20. #
  21. # Why no top-level param() and everything inside a function? `irm | iex`
  22. # executes this file as a string: script-path variables ($PSCommandPath,
  23. # $MyInvocation.MyCommand.Path) are $null and a top-level param block cannot
  24. # receive arguments through a pipe anyway (issue #565). Wrapping the logic in
  25. # a function and forwarding $args keeps one script working for both the pipe
  26. # path (no args, no script path) and the local-clone path.
  27. function Install-Caveman {
  28. param(
  29. [string[]]$InstallerArgs = @()
  30. )
  31. $ErrorActionPreference = "Stop"
  32. $Repo = "JuliusBrussee/caveman"
  33. $PinnedRef = if ($env:CAVEMAN_REF) { $env:CAVEMAN_REF } else { "v4.1.0" }
  34. # The CLI release the first run comes from when caveman is not installed;
  35. # kept equal to packages/cli/package.json (tests/installer/shim-security).
  36. $CliVersion = "4.1.0"
  37. # Keep Node first; only fall back when it is absent, never when it is old.
  38. $Runtime = "node"
  39. $node = Get-Command node -ErrorAction SilentlyContinue
  40. if (-not $node) {
  41. if (Get-Command bun -ErrorAction SilentlyContinue) {
  42. $Runtime = "bun"
  43. } else {
  44. Write-Error @"
  45. caveman: Node.js (>=18) or Bun required. Install:
  46. - winget install OpenJS.NodeJS.LTS
  47. - or download from https://nodejs.org
  48. - Bun (experimental skills-only fallback): https://bun.sh
  49. "@
  50. $global:LASTEXITCODE = 1; return
  51. }
  52. }
  53. if ($Runtime -eq "node") {
  54. $nodeMajor = [int](& node -p "process.versions.node.split('.')[0]")
  55. if ($nodeMajor -lt 18) {
  56. Write-Error "caveman: Node $nodeMajor too old. Need Node >=18. Upgrade: https://nodejs.org"
  57. $global:LASTEXITCODE = 1; return
  58. }
  59. }
  60. # If we're inside the repo clone, run the local installer directly.
  61. # $PSCommandPath is $null when piped to iex (#565) — the old unguarded
  62. # Split-Path on it was the "Cannot bind argument to parameter 'Path'
  63. # because it is null" crash.
  64. $local = $null
  65. if ($PSCommandPath) {
  66. $here = Split-Path -Parent $PSCommandPath
  67. $local = Join-Path $here "installer/install.js"
  68. if (-not (Test-Path $local)) { $local = $null }
  69. }
  70. if ($local) {
  71. & $Runtime $local @InstallerArgs
  72. } elseif ($Runtime -eq "bun") {
  73. # --bun overrides the package's Node shebang; no separate bunx is needed.
  74. & bun x --bun "github:$Repo#$PinnedRef" @InstallerArgs
  75. } else {
  76. # Curl-pipe path: delegate to npx.
  77. $npx = Get-Command npx -ErrorAction SilentlyContinue
  78. if (-not $npx) {
  79. Write-Error "caveman: npx required (ships with Node >=18). Reinstall Node.js."
  80. $global:LASTEXITCODE = 1; return
  81. }
  82. # Do NOT pass `--` here — npm 7+ npx already forwards trailing args to the
  83. # package, and a literal `--` was tripping installer/install.js's parseArgs as an
  84. # unknown flag.
  85. # npm 12 disables git package fetches by default (EALLOWGIT). Allow only the
  86. # root package requested here; older npm versions do not understand this
  87. # config flag.
  88. # 2>$null mirrors install.sh's `2>/dev/null`: npm writes its "does not support
  89. # Node.js" notice to stderr, and a contaminated value would floor the major to 0.
  90. $npxVersion = [string](& npx --version 2>$null)
  91. $npxMajor = 0
  92. if ($npxVersion -match '^(\d+)') {
  93. $npxMajor = [int]$Matches[1]
  94. }
  95. if ($npxMajor -ge 12) {
  96. & npx --allow-git=root -y "github:$Repo#$PinnedRef" @InstallerArgs
  97. } else {
  98. & npx -y "github:$Repo#$PinnedRef" @InstallerArgs
  99. }
  100. }
  101. if ($LASTEXITCODE -ne 0) { return }
  102. # End in the CLI's first run: modules, agents, one Continue.
  103. $skip = @($InstallerArgs | Where-Object { $_ -in @("-h", "--help", "--list", "-u", "--uninstall", "--dry-run") })
  104. if ($skip.Count -gt 0) { return }
  105. if ($Runtime -eq "bun") {
  106. Write-Host ""
  107. Write-Host "caveman: skills installer finished with Bun. The runtime (smaller inputs, Auto routing) needs Node 22.13+."
  108. Write-Host " Install Node (https://nodejs.org), then run: npx -y @caveman-ai/cli@$CliVersion"
  109. $global:LASTEXITCODE = 0; return
  110. }
  111. # The skills above run on Node 18; the CLI (runtime, routing) needs 22.13.
  112. # A prerelease Node ("25.0.0-nightly…") is not a [version] until its suffix goes.
  113. $nodeVersion = [version](([string](& node -p "process.versions.node")) -replace '[-+].*$', '')
  114. if ($nodeVersion -lt [version]"22.13.0") {
  115. Write-Host ""
  116. Write-Host "caveman: skills installed. The runtime (smaller inputs, Auto routing) needs Node 22.13+; this is v$nodeVersion."
  117. Write-Host " Upgrade Node (winget install OpenJS.NodeJS.LTS), then run: npx -y @caveman-ai/cli@$CliVersion"
  118. return
  119. }
  120. # The caveman on PATH when it is this release or newer: an older CLI has an
  121. # older setup, so npx runs this one and its setup installs it for good, but
  122. # over a newer CLI that would be a downgrade. A prerelease ranks below its
  123. # own release, and two prereleases of one release rank as semver orders
  124. # them; a version that cannot be read counts as older. The probe gives up
  125. # after 10s. Same probe as install.sh, with no double quotes in it:
  126. # Windows PowerShell drops those from a native command's arguments (#249).
  127. $setup = @("npx", "-y", "@caveman-ai/cli@$CliVersion", "setup")
  128. if (Get-Command caveman -ErrorAction SilentlyContinue) {
  129. try {
  130. & node -e 'const r=require(`child_process`).spawnSync(`caveman --version`,{shell:true,encoding:`utf8`,timeout:1e4,killSignal:`SIGKILL`,stdio:[`ignore`,`pipe`,`ignore`],windowsHide:true});const v=s=>(s=/^(\d+)\.(\d+)\.(\d+)(?:-([\w.-]+))?/.exec(s))&&[s[1],s[2],s[3],+!s[4],...s[4]?s[4].split(`.`):[]].map(x=>/^\d+$/.test(x)?+x:x);let h;try{h=v(JSON.parse(r.stdout).version)}catch{}const c=(a,b)=>a===b?0:a===undefined?-1:b===undefined?1:typeof a!=typeof b?(typeof a==`number`?-1:1):a<b?-1:1;const w=v(process.argv[1]),d=h&&w&&[...h,...w].map((_,i)=>c(h[i],w[i])).find(x=>x);process.exitCode=h&&w&&!(d<0)?0:1' $CliVersion 2>$null
  131. if ($LASTEXITCODE -eq 0) { $setup = @("caveman", "setup") }
  132. } catch { }
  133. }
  134. # --non-interactive never prompts: name the first run instead of starting it.
  135. if ($InstallerArgs -notcontains "--non-interactive" -and [Environment]::UserInteractive -and -not [Console]::IsInputRedirected -and -not [Console]::IsOutputRedirected) {
  136. & $setup[0] $setup[1..($setup.Length - 1)]
  137. return
  138. }
  139. Write-Host "Next: $($setup -join ' ')"
  140. $global:LASTEXITCODE = 0
  141. }
  142. # $args is the automatic variable: populated when run as a file
  143. # (`pwsh install.ps1 --force`), empty under `irm | iex`.
  144. Install-Caveman -InstallerArgs $args
  145. # Under `irm | iex` this runs in the user's own session, where `exit` would
  146. # close their window over the output above; only a script file exits.
  147. if ($PSCommandPath) { exit $LASTEXITCODE }
install.sh · Bundled script · 149 lines · 2 findings marked
  1. #!/usr/bin/env bash
  2. # caveman — installer shim.
  3. #
  4. # Thin wrapper around installer/install.js (the unified Node installer). Every flag
  5. # you'd pass to installer/install.js can be passed here; we just forward them.
  6. #
  7. # One-line install:
  8. # curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v1.10.0/install.sh | bash
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 8 on GitHub
  9. # curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/v1.10.0/install.sh | bash -s -- --all
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 9 on GitHub
  10. #
  11. # Local clone:
  12. # bash install.sh [flags]
  13. #
  14. # After the installer it hands over to the CLI's first run (`caveman setup`:
  15. # modules, agents, one Continue) when a terminal is attached, and prints that
  16. # one command otherwise.
  17. #
  18. # Why a Node installer? install.sh + install.ps1 used to be parallel sources
  19. # of truth and constantly drifted (issue #249, etc.). One Node script works
  20. # everywhere without bash/PowerShell quoting bugs.
  21. set -euo pipefail
  22. REPO="JuliusBrussee/caveman"
  23. PINNED_REF="${CAVEMAN_REF:-v4.1.0}"
  24. # The CLI release the first run comes from when caveman is not installed;
  25. # kept equal to packages/cli/package.json (tests/installer/shim-security).
  26. CLI_VERSION="4.1.0"
  27. # Prefer Node, including its existing version check. Bun is a fallback only
  28. # when Node is absent, not a way to hide an old or broken Node installation.
  29. RUNTIME=node
  30. if ! command -v node >/dev/null 2>&1; then
  31. if command -v bun >/dev/null 2>&1; then
  32. RUNTIME=bun
  33. else
  34. echo "caveman: Node.js (≥18) or Bun required. Install:" >&2
  35. echo " macOS: brew install node" >&2
  36. echo " Linux: see https://nodejs.org or use nvm (https://github.com/nvm-sh/nvm)" >&2
  37. echo " Bun (experimental skills-only fallback): https://bun.sh" >&2
  38. exit 1
  39. fi
  40. fi
  41. if [ "$RUNTIME" = node ]; then
  42. NODE_MAJOR=$(node -p "process.versions.node.split('.')[0]")
  43. if [ "$NODE_MAJOR" -lt 18 ]; then
  44. echo "caveman: Node $NODE_MAJOR too old. Need Node ≥18." >&2
  45. echo " Upgrade: https://nodejs.org" >&2
  46. exit 1
  47. fi
  48. fi
  49. # first_run ends the install in the CLI's first run. The terminal comes from
  50. # /dev/tty because under curl | bash stdin is the script itself.
  51. first_run() {
  52. launch=1
  53. for arg in "$@"; do
  54. case "$arg" in
  55. -h|--help|--list|-u|--uninstall|--dry-run) return 0 ;;
  56. # Never prompt: name the first run instead of starting it.
  57. --non-interactive) launch=0 ;;
  58. esac
  59. done
  60. if [ "$RUNTIME" = bun ]; then
  61. echo
  62. echo "caveman: skills installer finished with Bun. The runtime (smaller inputs, Auto routing) needs Node 22.13+."
  63. echo " Install Node (https://nodejs.org), then run: npx -y @caveman-ai/cli@$CLI_VERSION"
  64. return 0
  65. fi
  66. # The skills above run on Node 18; the CLI (runtime, routing) needs 22.13.
  67. node_version=$(node -p "process.versions.node")
  68. node_minor=${node_version#*.}
  69. node_minor=${node_minor%%.*}
  70. case "$node_minor" in ''|*[!0-9]*) node_minor=0 ;; esac
  71. if [ "$NODE_MAJOR" -lt 22 ] || { [ "$NODE_MAJOR" -eq 22 ] && [ "$node_minor" -lt 13 ]; }; then
  72. echo
  73. echo "caveman: skills installed. The runtime (smaller inputs, Auto routing) needs Node 22.13+; this is v$node_version."
  74. echo " Upgrade Node (https://nodejs.org), then run: npx -y @caveman-ai/cli@$CLI_VERSION"
  75. return 0
  76. fi
  77. # The caveman on PATH when it is this release or newer: an older CLI has an
  78. # older setup, so npx runs this one and its setup installs it for good, but
  79. # over a newer CLI that would be a downgrade. A prerelease ranks below its
  80. # own release, and two prereleases of one release rank as semver orders
  81. # them; a version that cannot be read counts as older. The probe gives up
  82. # after 10s (macOS has no `timeout`). install.ps1 carries it too.
  83. if command -v caveman >/dev/null 2>&1 &&
  84. node -e 'const r=require(`child_process`).spawnSync(`caveman --version`,{shell:true,encoding:`utf8`,timeout:1e4,killSignal:`SIGKILL`,stdio:[`ignore`,`pipe`,`ignore`],windowsHide:true});const v=s=>(s=/^(\d+)\.(\d+)\.(\d+)(?:-([\w.-]+))?/.exec(s))&&[s[1],s[2],s[3],+!s[4],...s[4]?s[4].split(`.`):[]].map(x=>/^\d+$/.test(x)?+x:x);let h;try{h=v(JSON.parse(r.stdout).version)}catch{}const c=(a,b)=>a===b?0:a===undefined?-1:b===undefined?1:typeof a!=typeof b?(typeof a==`number`?-1:1):a<b?-1:1;const w=v(process.argv[1]),d=h&&w&&[...h,...w].map((_,i)=>c(h[i],w[i])).find(x=>x);process.exitCode=h&&w&&!(d<0)?0:1' "$CLI_VERSION" </dev/null 2>/dev/null; then
  85. set -- caveman setup
  86. else
  87. set -- npx -y "@caveman-ai/cli@$CLI_VERSION" setup
  88. fi
  89. if [ "$launch" = 1 ] && [ -t 1 ] && { : </dev/tty; } 2>/dev/null; then
  90. echo
  91. "$@" </dev/tty
  92. else
  93. echo "Next: $*"
  94. fi
  95. }
  96. # If we're inside the repo clone, run the local installer directly — saves
  97. # the npx round-trip and keeps offline installs working. BASH_SOURCE is unset
  98. # when bash is invoked from stdin (curl | bash). Do not feed an empty value to
  99. # dirname: dirname "" resolves to "." and would execute an unrelated
  100. # $PWD/installer/install.js from the caller's checkout.
  101. here=""
  102. source_path="${BASH_SOURCE[0]:-}"
  103. if [ -n "$source_path" ]; then
  104. here="$(cd "$(dirname "$source_path")" 2>/dev/null && pwd)" || here=""
  105. fi
  106. if [ -n "$here" ] && [ -f "$here/installer/install.js" ]; then
  107. "$RUNTIME" "$here/installer/install.js" "$@"
  108. first_run "$@"
  109. exit 0
  110. fi
  111. # Force Bun for the package's Node shebang, and retain the same release pin.
  112. # `bun x` is built into bun; a separate bunx executable is not required.
  113. if [ "$RUNTIME" = bun ]; then
  114. bun x --bun "github:$REPO#$PINNED_REF" "$@"
  115. first_run "$@"
  116. exit 0
  117. fi
  118. # Curl-pipe path: delegate to npx. We do NOT pass `--` here — npm 7+ npx
  119. # already forwards trailing args to the package, and a literal `--` tripped
  120. # installer/install.js's parseArgs as an unknown flag.
  121. if ! command -v npx >/dev/null 2>&1; then
  122. echo "caveman: npx required (ships with Node ≥18). Reinstall Node.js." >&2
  123. exit 1
  124. fi
  125. # npm 12 disables git package fetches by default (EALLOWGIT). Allow only the
  126. # root package requested here; keep the old invocation for npm versions that
  127. # predate this config flag.
  128. NPX_VERSION=""
  129. if NPX_VERSION=$(npx --version 2>/dev/null); then :; fi
  130. NPX_MAJOR=${NPX_VERSION%%.*}
  131. case "$NPX_MAJOR" in
  132. ''|*[!0-9]*) NPX_MAJOR=0 ;;
  133. esac
  134. if [ "$NPX_MAJOR" -ge 12 ]; then
  135. npx --allow-git=root -y "github:$REPO#$PINNED_REF" "$@"
  136. else
  137. npx -y "github:$REPO#$PINNED_REF" "$@"
  138. fi
  139. first_run "$@"
packages/cli/src/agents.generated.ts · Bundled script · 778 lines · 1 finding marked
  1. // GENERATED by agents/compile.mjs from agents/profiles/*.json — DO NOT EDIT.
  2. // Run `node scripts/compile-registries.mjs` (wired into the CLI build/test) to regenerate.
  3. export type WireProtocol = "anthropic-messages" | "openai-chat" | "openai-responses" | "gemini-generatecontent";
  4. export type Injection =
  5. | { method: "env"; env: Record<string, string> }
  6. | { method: "config-env-content"; env_var: string; config_content: { local: unknown; managed?: unknown } }
  7. | { method: "config-file"; env_var: string; base_config?: { path: string; env_var?: string; state_dir?: { env_var: string; filename: string }; platform_default?: "qwen-system-settings" }; config_overlay: { local: unknown; managed?: unknown } }
  8. | { method: "native-extension"; host: string; asset: string; loader_flag: string };
  9. export type CommandHook =
  10. | { method: "claude-pretooluse" }
  11. | { method: "codex-pretooluse"; file: string }
  12. | { method: "gemini-beforetool" }
  13. | { method: "opencode-plugin" }
  14. | { method: "hermes-plugin" }
  15. | { method: "openclaw-plugin" }
  16. | { method: "pi-extension" }
  17. | { method: "instruction-note"; file: string };
  18. export type MemoryHook =
  19. | { method: "claude-userpromptsubmit" };
  20. export type SkillsSurface = { format: "skill-md"; user_dirs: string[]; project_dirs?: string[] };
  21. export interface AgentProfile {
  22. schema_version: string;
  23. id: string;
  24. display_name: string;
  25. vendor: string;
  26. homepage: string;
  27. binary_names: string[];
  28. args: string[];
  29. install: string;
  30. wire_protocol: WireProtocol;
  31. injection: Injection;
  32. command_hook?: CommandHook;
  33. memory_hook?: MemoryHook;
  34. skills?: SkillsSurface;
  35. attribution?: { header?: string };
  36. tested_agent_version?: string;
  37. injection_completeness?: "declarative" | "builder-assisted" | "code-only";
  38. last_verified_at?: string;
  39. verified_by?: string;
  40. fallback?: string;
  41. maintainer?: string | null;
  42. }
  43. export const PROFILES: AgentProfile[] = [
  44. {
  45. "schema_version": "1",
  46. "id": "aider",
  47. "display_name": "Aider",
  48. "vendor": "OpenAI / Anthropic",
  49. "homepage": "https://aider.chat/docs/llms/openai-compat.html",
  50. "binary_names": [
  51. "aider"
  52. ],
  53. "args": [],
  54. "install": "python -m pip install aider-chat",
  55. "wire_protocol": "openai-chat",
  56. "injection": {
  57. "method": "env",
  58. "env": {
  59. "OPENAI_API_BASE": "{{cave_base_url}}/openai/v1"
  60. }
  61. },
  62. "attribution": {
  63. "header": "x-cave-agent"
  64. },
  65. "tested_agent_version": "0.86.2",
  66. "injection_completeness": "declarative",
  67. "fallback": "generic-env",
  68. "maintainer": null
  69. },
  70. {
  71. "schema_version": "1",
  72. "id": "claude",
  73. "display_name": "Claude Code",
  74. "vendor": "Anthropic",
  75. "homepage": "https://code.claude.com/docs",
  76. "binary_names": [
  77. "claude"
  78. ],
  79. "args": [],
  80. "install": "npm i -g @anthropic-ai/claude-code",
  81. "wire_protocol": "anthropic-messages",
  82. "injection": {
  83. "method": "env",
  84. "env": {
  85. "ANTHROPIC_BASE_URL": "{{cave_base_url}}",
  86. "ANTHROPIC_AUTH_TOKEN": "{{cave_api_key}}"
  87. }
  88. },
  89. "skills": {
  90. "format": "skill-md",
  91. "user_dirs": [
  92. "~/.claude/skills"
  93. ],
  94. "project_dirs": [
  95. ".claude/skills"
  96. ]
  97. },
  98. "command_hook": {
  99. "method": "claude-pretooluse"
  100. },
  101. "memory_hook": {
  102. "method": "claude-userpromptsubmit"
  103. },
  104. "attribution": {
  105. "header": "x-cave-agent"
  106. },
  107. "tested_agent_version": "2.1.318",
  108. "injection_completeness": "builder-assisted",
  109. "last_verified_at": "2026-11-06",
  110. "verified_by": "CI pinned-binary probe (claude 2.1.318, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); route probe last run on 2.1.308, not re-run on 2.1.318, from a scratch HOME and CLAUDE_CONFIG_DIR with onboarding pre-accepted in ~/.claude.json: ANTHROPIC_BASE_URL redirection observed reaching POST /v1/messages?beta=true under the injected base URL, the injected ANTHROPIC_AUTH_TOKEN (a dummy value) arriving as Authorization: Bearer with no x-api-key, and the run completed (exit 0, stop_reason end_turn). The unauthenticated HEAD /api/hello reachability preflight first recorded on 2.1.306 is still sent to the same injected base URL before the turn, now with a Bun/1.4.4 user-agent rather than the CLI's own -- a mount that answers /v1/messages but rejects HEAD on an unknown path is worth checking against this. The turn itself is a streaming request (stream: true) carrying anthropic-beta: claude-code-20250219 and user-agent claude-cli/2.1.308 (external, sdk-cli)",
  111. "fallback": "generic-env",
  112. "maintainer": null
  113. },
  114. {
  115. "schema_version": "1",
  116. "id": "codex",
  117. "display_name": "OpenAI Codex CLI",
  118. "vendor": "OpenAI",
  119. "homepage": "https://developers.openai.com/codex",
  120. "binary_names": [
  121. "codex"
  122. ],
  123. "args": [],
  124. "install": "npm i -g @openai/codex",
  125. "wire_protocol": "openai-responses",
  126. "injection": {
  127. "method": "env",
  128. "env": {}
  129. },
  130. "skills": {
  131. "format": "skill-md",
  132. "user_dirs": [
  133. "~/.codex/skills"
  134. ],
  135. "project_dirs": []
  136. },
  137. "command_hook": {
  138. "method": "codex-pretooluse",
  139. "file": "~/.codex/AGENTS.md"
  140. },
  141. "attribution": {
  142. "header": "x-cave-agent"
  143. },
  144. "tested_agent_version": "0.166.0",
  145. "injection_completeness": "code-only",
  146. "last_verified_at": "2026-11-02",
  147. "verified_by": "Windows x64 route probes (@openai/codex 0.166.0, Node 24.16.0; --version and --help exit 0): actual caveman wrap codex exec from two scratch homes sent POST /w/codex/v1/responses with dummy bearer auth, model gpt-5.5 from config.toml (no CLI model override), stream true; scripted SSE replies consumed, both exit 0. Attribution uses /w/codex; X-Cave-Agent header absent. API-key path with proxy/MCP/shrink/browse disabled; no subscription, compression, live-provider or task-quality verification",
  148. "fallback": "generic-env",
  149. "maintainer": null
  150. },
  151. {
  152. "schema_version": "1",
  153. "id": "gemini",
  154. "display_name": "Gemini CLI",
  155. "vendor": "Google",
  156. "homepage": "https://github.com/google-gemini/gemini-cli",
  157. "binary_names": [
  158. "gemini"
  159. ],
  160. "args": [],
  161. "install": "npm i -g @google/gemini-cli",
  162. "wire_protocol": "gemini-generatecontent",
  163. "injection": {
  164. "method": "env",
  165. "env": {
  166. "GOOGLE_GEMINI_BASE_URL": "{{cave_base_url}}",
  167. "GOOGLE_VERTEX_BASE_URL": "{{cave_base_url}}/vertex"
  168. }
  169. },
  170. "command_hook": {
  171. "method": "gemini-beforetool"
  172. },
  173. "attribution": {
  174. "header": "x-cave-agent"
  175. },
  176. "tested_agent_version": "0.67.0",
  177. "injection_completeness": "declarative",
  178. "last_verified_at": "2026-11-06",
  179. "verified_by": "CI pinned-binary probe (gemini 0.67.0, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); route probe last run on 0.65.0, not re-run on 0.67.0, from a scratch HOME with GEMINI_CLI_TRUST_WORKSPACE=true and security.auth.selectedType gemini-api-key in its settings: GOOGLE_GEMINI_BASE_URL redirection observed reaching POST /v1beta/models/{model}:generateContent (model-router classifier) and POST /v1beta/models/{model}:streamGenerateContent?alt=sse (the turn) under the injected base URL, carrying the key as x-goog-api-key, and the run completed (exit 0). The classifier's own call was answered by a generic stub rather than a schema-valid classifier response, so it exhausted its retries and fell back to the default model; the route and auth header are what this probe observed, not the classifier's verdict. NOT observed: an env-only launch -- with no selectedType, 0.65.0 (like 0.64.0 and 0.63.0) derives the gateway auth type from GOOGLE_GEMINI_BASE_URL and exits 41 \"Invalid auth method selected.\" before sending any request (re-confirmed on 0.65.0: zero requests reached the recorder)",
  180. "fallback": "generic-env",
  181. "maintainer": null
  182. },
  183. {
  184. "schema_version": "1",
  185. "id": "hermes",
  186. "display_name": "Hermes Agent",
  187. "vendor": "Nous Research",
  188. "homepage": "https://hermes-agent.nousresearch.com/docs/",
  189. "binary_names": [
  190. "hermes"
  191. ],
  192. "args": [
  193. "--provider",
  194. "custom"
  195. ],
  196. "install": "curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash",
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 203 on GitHub
  197. "wire_protocol": "openai-chat",
  198. "injection": {
  199. "method": "env",
  200. "env": {
  201. "CUSTOM_BASE_URL": "{{cave_base_url}}/v1",
  202. "CUSTOM_API_KEY": "{{cave_api_key}}"
  203. }
  204. },
  205. "command_hook": {
  206. "method": "hermes-plugin"
  207. },
  208. "attribution": {
  209. "header": "x-cave-agent"
  210. },
  211. "tested_agent_version": "0.23.0",
  212. "injection_completeness": "builder-assisted",
  213. "last_verified_at": "2026-11-06",
  214. "verified_by": "CI pinned-binary probe (hermes 0.23.0 = release tag v0.23.0 at 2bd378ed, installed with uv sync --frozen from that tag's uv.lock on Python 3.14, agent-conformance pinned-upstream-binary lane on ubuntu-latest, agents/probe-installed.mjs: --version and --help both exit 0). Both install details are required, not incidental: from 0.22 hermes gates every core dependency on python_version >= '3.14', so a 3.12 install carries none of them and --help fails importing ruamel (the 0.22.7 drift report, #1236), and pip cannot resolve it at all (Pillow==13.0.0 against browser-harness's pillow==12.3.0, settled only by the project's [tool.uv] override-dependencies). Launchability only: the hermes-plugin hook, the native config write and a routed model request were not exercised on 0.23.0",
  215. "fallback": "generic-env",
  216. "maintainer": null
  217. },
  218. {
  219. "schema_version": "1",
  220. "id": "kilo",
  221. "display_name": "Kilo Code",
  222. "vendor": "Kilo Code",
  223. "homepage": "https://kilo.ai/docs/code-with-ai/platforms/cli",
  224. "binary_names": [
  225. "kilo",
  226. "kilocode"
  227. ],
  228. "args": [],
  229. "install": "npm install -g @kilocode/cli@7.8.27",
  230. "wire_protocol": "openai-chat",
  231. "injection": {
  232. "method": "config-env-content",
  233. "env_var": "KILO_CONFIG_CONTENT",
  234. "config_content": {
  235. "local": {
  236. "provider": {
  237. "caveman": {
  238. "npm": "@ai-sdk/openai-compatible",
  239. "name": "Caveman Local Proxy",
  240. "options": {
  241. "baseURL": "{{cave_base_url}}/v1",
  242. "apiKey": "{env:OPENAI_API_KEY}",
  243. "headers": {
  244. "X-Cave-Agent": "kilo"
  245. }
  246. },
  247. "models": {
  248. "gpt-5.5": {
  249. "name": "GPT-5.5 (via Caveman)",
  250. "tool_call": true,
  251. "limit": {
  252. "context": 1050000,
  253. "output": 32768
  254. }
  255. },
  256. "gpt-5.4-mini": {
  257. "name": "GPT-5.4 mini (via Caveman)",
  258. "tool_call": true,
  259. "limit": {
  260. "context": 400000,
  261. "output": 16384
  262. }
  263. }
  264. }
  265. }
  266. },
  267. "model": "caveman/gpt-5.5",
  268. "enabled_providers": [
  269. "caveman"
  270. ]
  271. },
  272. "managed": {
  273. "provider": {
  274. "caveman": {
  275. "npm": "@ai-sdk/openai-compatible",
  276. "name": "Caveman Gateway",
  277. "options": {
  278. "baseURL": "{{cave_base_url}}/v1",
  279. "apiKey": "{env:CAVE_API_KEY}",
  280. "headers": {
  281. "x-cave-upstream-key": "{{cave_optional_openai_key_env}}",
  282. "X-Cave-Agent": "kilo"
  283. }
  284. },
  285. "models": {
  286. "gpt-5.5": {
  287. "name": "GPT-5.5 (via Caveman)",
  288. "tool_call": true,
  289. "limit": {
  290. "context": 1050000,
  291. "output": 32768
  292. }
  293. },
  294. "gpt-5.4-mini": {
  295. "name": "GPT-5.4 mini (via Caveman)",
  296. "tool_call": true,
  297. "limit": {
  298. "context": 400000,
  299. "output": 16384
  300. }
  301. }
  302. }
  303. }
  304. },
  305. "model": "caveman/gpt-5.5",
  306. "enabled_providers": [
  307. "caveman"
  308. ]
  309. }
  310. }
  311. },
  312. "attribution": {
  313. "header": "X-Cave-Agent"
  314. },
  315. "tested_agent_version": "7.8.27",
  316. "injection_completeness": "declarative",
  317. "last_verified_at": "2026-11-06",
  318. "verified_by": "CI pinned-binary probe (kilo 7.8.27, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); two route probes last run on 7.8.18 (PR #1416), not re-run on 7.8.27, from separate scratch homes and different loopback ports: this profile's own KILO_CONFIG_CONTENT local overlay made `kilo run --pure --format json` send POST /v1/chat/completions at the injected baseURL carrying X-Cave-Agent: kilo and the dummy OPENAI_API_KEY as Authorization: Bearer (model gpt-5.5, stream: true), receive the scripted SSE reply, and complete (exit 0). The model came from the profile, without a CLI model override. The overlay was rendered from the profile by hand rather than through the CLI's buildWrapEnv, so its base-config merge was not exercised; external plugins were disabled with --pure. No live provider or compression was tested",
  319. "fallback": "generic-env",
  320. "maintainer": null
  321. },
  322. {
  323. "schema_version": "1",
  324. "id": "openclaw",
  325. "display_name": "OpenClaw",
  326. "vendor": "OpenClaw",
  327. "homepage": "https://github.com/openclaw/openclaw#readme",
  328. "binary_names": [
  329. "openclaw"
  330. ],
  331. "args": [
  332. "chat"
  333. ],
  334. "install": "npm i -g openclaw",
  335. "wire_protocol": "openai-chat",
  336. "injection": {
  337. "method": "config-file",
  338. "env_var": "OPENCLAW_CONFIG_PATH",
  339. "base_config": {
  340. "path": "~/.openclaw/openclaw.json",
  341. "env_var": "OPENCLAW_CONFIG_PATH",
  342. "state_dir": {
  343. "env_var": "OPENCLAW_STATE_DIR",
  344. "filename": "openclaw.json"
  345. }
  346. },
  347. "config_overlay": {
  348. "local": {
  349. "mcp": {
  350. "servers": {
  351. "caveman": {
  352. "command": "caveman-mcp",
  353. "args": []
  354. }
  355. }
  356. }
  357. },
  358. "managed": {
  359. "mcp": {
  360. "servers": {
  361. "caveman": {
  362. "command": "caveman-mcp",
  363. "args": []
  364. }
  365. }
  366. }
  367. }
  368. }
  369. },
  370. "command_hook": {
  371. "method": "openclaw-plugin"
  372. },
  373. "attribution": {
  374. "header": "x-cave-agent"
  375. },
  376. "tested_agent_version": "2026.11.1",
  377. "injection_completeness": "builder-assisted",
  378. "last_verified_at": "2026-11-06",
  379. "verified_by": "CI pinned-binary probe (openclaw 2026.11.1, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); config check last run on 2026.10.5, not re-run on 2026.11.1: `openclaw config validate` accepts this profile's injection.config_overlay.local passed as OPENCLAW_CONFIG_PATH (exit 0, \"Config valid\"), and a malformed control file is rejected (exit 1, JSON5 parse error). 2026.10.5 and 2026.11.1 both declare engines.node \">=24.16.0 <25 || >=26.1.0\", so neither will install on Node 22 or 25 at all -- the conformance lane's pinned 24.16.0 sits exactly on that floor. NOT observed: the ~/.openclaw/workspace SOUL.md and workspace-skill bootstrap, or any model request -- both need a live credentialed session",
  380. "fallback": "generic-env",
  381. "maintainer": null
  382. },
  383. {
  384. "schema_version": "1",
  385. "id": "opencode",
  386. "display_name": "opencode",
  387. "vendor": "sst",
  388. "homepage": "https://opencode.ai/docs/config",
  389. "binary_names": [
  390. "opencode",
  391. "opencode2"
  392. ],
  393. "args": [],
  394. "install": "npm i -g opencode-ai",
  395. "wire_protocol": "openai-chat",
  396. "injection": {
  397. "method": "config-env-content",
  398. "env_var": "OPENCODE_CONFIG_CONTENT",
  399. "config_content": {
  400. "local": {
  401. "$schema": "https://opencode.ai/config.json",
  402. "provider": {
  403. "openai": {
  404. "options": {
  405. "baseURL": "{{cave_base_url}}/v1",
  406. "headers": {
  407. "X-Cave-Agent": "opencode"
  408. }
  409. }
  410. },
  411. "anthropic": {
  412. "options": {
  413. "baseURL": "{{cave_base_url}}/v1",
  414. "headers": {
  415. "X-Cave-Agent": "opencode"
  416. }
  417. }
  418. },
  419. "opencode-go": {
  420. "options": {
  421. "baseURL": "{{cave_base_url}}/compat/opencode-go/v1",
  422. "headers": {
  423. "X-Cave-Agent": "opencode"
  424. }
  425. }
  426. }
  427. }
  428. },
  429. "managed": {
  430. "$schema": "https://opencode.ai/config.json",
  431. "provider": {
  432. "caveman": {
  433. "npm": "@ai-sdk/openai-compatible",
  434. "name": "Caveman Gateway",
  435. "options": {
  436. "baseURL": "{{cave_base_url}}/v1",
  437. "apiKey": "{env:CAVE_API_KEY}",
  438. "headers": {
  439. "x-cave-upstream-key": "{env:OPENAI_API_KEY}",
  440. "X-Cave-Agent": "opencode"
  441. }
  442. },
  443. "models": {
  444. "gpt-5.5": {
  445. "name": "GPT-5.5 (via Caveman)",
  446. "limit": {
  447. "context": 1050000,
  448. "output": 32768
  449. }
  450. },
  451. "gpt-5.4-mini": {
  452. "name": "GPT-5.4 mini (via Caveman)",
  453. "limit": {
  454. "context": 400000,
  455. "output": 16384
  456. }
  457. }
  458. }
  459. }
  460. },
  461. "model": "caveman/gpt-5.5"
  462. }
  463. }
  464. },
  465. "command_hook": {
  466. "method": "opencode-plugin"
  467. },
  468. "attribution": {
  469. "header": "X-Cave-Agent"
  470. },
  471. "tested_agent_version": "1.19.5",
  472. "injection_completeness": "declarative",
  473. "last_verified_at": "2026-11-06",
  474. "verified_by": "CI pinned-binary probe (opencode 1.19.5, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); route probe last run on 1.19.4, not re-run on 1.19.5: this profile's own config_content.local overlay loaded and POST /v1/responses reached the injected baseURL twice (title model and turn), each carrying X-Cave-Agent: opencode, run completed (exit 0)",
  475. "fallback": "generic-env",
  476. "maintainer": null
  477. },
  478. {
  479. "schema_version": "1",
  480. "id": "pi",
  481. "display_name": "Pi",
  482. "vendor": "pi.dev",
  483. "homepage": "https://pi.dev/",
  484. "binary_names": [
  485. "pi"
  486. ],
  487. "args": [],
  488. "install": "npm install -g --ignore-scripts @earendil-works/pi-coding-agent",
  489. "wire_protocol": "openai-chat",
  490. "injection": {
  491. "method": "native-extension",
  492. "host": "pi",
  493. "asset": "caveman-pi-extension",
  494. "loader_flag": "--extension"
  495. },
  496. "command_hook": {
  497. "method": "pi-extension"
  498. },
  499. "tested_agent_version": "1.0.4",
  500. "injection_completeness": "builder-assisted",
  501. "last_verified_at": "2026-10-06",
  502. "verified_by": "local pinned-binary probe (pi 1.0.4, agents/probe-installed.mjs); packages/pi-extension suite against pi-coding-agent 1.0.4",
  503. "fallback": "none",
  504. "maintainer": null
  505. },
  506. {
  507. "schema_version": "1",
  508. "id": "qwen",
  509. "display_name": "Qwen Code",
  510. "vendor": "QwenLM",
  511. "homepage": "https://github.com/QwenLM/qwen-code",
  512. "binary_names": [
  513. "qwen"
  514. ],
  515. "args": [
  516. "--extensions=none"
  517. ],
  518. "install": "npm i -g @qwen-code/qwen-code@0.26.8",
  519. "wire_protocol": "openai-chat",
  520. "injection": {
  521. "method": "config-file",
  522. "env_var": "QWEN_CODE_SYSTEM_SETTINGS_PATH",
  523. "base_config": {
  524. "path": "~/.qwen/system-settings.json",
  525. "env_var": "QWEN_CODE_SYSTEM_SETTINGS_PATH",
  526. "platform_default": "qwen-system-settings"
  527. },
  528. "config_overlay": {
  529. "local": {
  530. "providerProtocol": {},
  531. "modelProviders": {
  532. "openai": [
  533. {
  534. "id": "gpt-5.5",
  535. "name": "GPT-5.5 (via Caveman)",
  536. "envKey": "OPENAI_API_KEY",
  537. "baseUrl": "{{cave_base_url}}/v1",
  538. "generationConfig": {
  539. "customHeaders": {
  540. "X-Cave-Agent": "qwen"
  541. }
  542. }
  543. },
  544. {
  545. "id": "gpt-5.4-mini",
  546. "name": "GPT-5.4 mini (via Caveman)",
  547. "envKey": "OPENAI_API_KEY",
  548. "baseUrl": "{{cave_base_url}}/v1",
  549. "generationConfig": {
  550. "customHeaders": {
  551. "X-Cave-Agent": "qwen"
  552. }
  553. }
  554. }
  555. ]
  556. },
  557. "security": {
  558. "auth": {
  559. "selectedType": "openai",
  560. "enforcedType": "openai"
  561. }
  562. },
  563. "model": {
  564. "name": "gpt-5.5",
  565. "baseUrl": "{{cave_base_url}}/v1"
  566. },
  567. "modelFallbacks": "",
  568. "fastModel": "",
  569. "advisorModel": "",
  570. "visionModel": "",
  571. "compactionModel": "",
  572. "imageModel": "",
  573. "voiceModel": "",
  574. "disableAllHooks": true,
  575. "tools": {
  576. "webSearch": {
  577. "enabled": false,
  578. "model": ""
  579. },
  580. "workflowsEnabled": false,
  581. "disabled": [
  582. "agent",
  583. "skill",
  584. "web_search",
  585. "create_sub_session",
  586. "workflow"
  587. ]
  588. },
  589. "agents": {
  590. "builtin": {
  591. "exploreModel": "inherit"
  592. },
  593. "allowedGrades": []
  594. },
  595. "permissions": {
  596. "deny": [
  597. "Agent",
  598. "Skill",
  599. "WebSearch",
  600. "CreateSubSession",
  601. "Workflow"
  602. ]
  603. },
  604. "skills": {
  605. "disabledLevels": [
  606. "project",
  607. "user",
  608. "extension",
  609. "bundled"
  610. ]
  611. },
  612. "experimental": {
  613. "liveVoice": {
  614. "enabled": false,
  615. "apiKey": "",
  616. "model": ""
  617. }
  618. },
  619. "slashCommands": {
  620. "disabled": [
  621. "auth",
  622. "connect",
  623. "login",
  624. "model",
  625. "arena",
  626. "agents",
  627. "workflows",
  628. "settings",
  629. "config",
  630. "permissions",
  631. "skills",
  632. "voice",
  633. "extensions",
  634. "reload-plugins",
  635. "hooks",
  636. "mcp",
  637. "resume",
  638. "continue",
  639. "import-config"
  640. ]
  641. }
  642. },
  643. "managed": {
  644. "providerProtocol": {},
  645. "modelProviders": {
  646. "openai": [
  647. {
  648. "id": "gpt-5.5",
  649. "name": "GPT-5.5 (via Caveman)",
  650. "envKey": "CAVE_API_KEY",
  651. "baseUrl": "{{cave_base_url}}/v1",
  652. "generationConfig": {
  653. "customHeaders": {
  654. "x-cave-upstream-key": "{{cave_optional_openai_key_env}}",
  655. "X-Cave-Agent": "qwen"
  656. }
  657. }
  658. },
  659. {
  660. "id": "gpt-5.4-mini",
  661. "name": "GPT-5.4 mini (via Caveman)",
  662. "envKey": "CAVE_API_KEY",
  663. "baseUrl": "{{cave_base_url}}/v1",
  664. "generationConfig": {
  665. "customHeaders": {
  666. "x-cave-upstream-key": "{{cave_optional_openai_key_env}}",
  667. "X-Cave-Agent": "qwen"
  668. }
  669. }
  670. }
  671. ]
  672. },
  673. "security": {
  674. "auth": {
  675. "selectedType": "openai",
  676. "enforcedType": "openai"
  677. }
  678. },
  679. "model": {
  680. "name": "gpt-5.5",
  681. "baseUrl": "{{cave_base_url}}/v1"
  682. },
  683. "modelFallbacks": "",
  684. "fastModel": "",
  685. "advisorModel": "",
  686. "visionModel": "",
  687. "compactionModel": "",
  688. "imageModel": "",
  689. "voiceModel": "",
  690. "disableAllHooks": true,
  691. "tools": {
  692. "webSearch": {
  693. "enabled": false,
  694. "model": ""
  695. },
  696. "workflowsEnabled": false,
  697. "disabled": [
  698. "agent",
  699. "skill",
  700. "web_search",
  701. "create_sub_session",
  702. "workflow"
  703. ]
  704. },
  705. "agents": {
  706. "builtin": {
  707. "exploreModel": "inherit"
  708. },
  709. "allowedGrades": []
  710. },
  711. "permissions": {
  712. "deny": [
  713. "Agent",
  714. "Skill",
  715. "WebSearch",
  716. "CreateSubSession",
  717. "Workflow"
  718. ]
  719. },
  720. "skills": {
  721. "disabledLevels": [
  722. "project",
  723. "user",
  724. "extension",
  725. "bundled"
  726. ]
  727. },
  728. "experimental": {
  729. "liveVoice": {
  730. "enabled": false,
  731. "apiKey": "",
  732. "model": ""
  733. }
  734. },
  735. "slashCommands": {
  736. "disabled": [
  737. "auth",
  738. "connect",
  739. "login",
  740. "model",
  741. "arena",
  742. "agents",
  743. "workflows",
  744. "settings",
  745. "config",
  746. "permissions",
  747. "skills",
  748. "voice",
  749. "extensions",
  750. "reload-plugins",
  751. "hooks",
  752. "mcp",
  753. "resume",
  754. "continue",
  755. "import-config"
  756. ]
  757. }
  758. }
  759. }
  760. },
  761. "attribution": {
  762. "header": "X-Cave-Agent"
  763. },
  764. "tested_agent_version": "0.26.8",
  765. "injection_completeness": "declarative",
  766. "last_verified_at": "2026-11-06",
  767. "verified_by": "CI pinned-binary probe (@qwen-code/qwen-code 0.26.8, agent-conformance pinned-upstream-binary lane on ubuntu-latest with Node 24.16.0, agents/probe-installed.mjs: --version and --help both exit 0); route check last run on 0.25.3, not re-run on 0.26.8: this profile's own config_overlay.local, rendered into the file named by QWEN_CODE_SYSTEM_SETTINGS_PATH with ENABLE_WEB_SEARCH=0 and QWEN_CODE_DISABLE_WORKFLOWS=1 as buildWrapEnv sets them, made `qwen --extensions=none -p` send POST /w/qwen/v1/chat/completions (X-Cave-Agent: qwen, model gpt-5.5) to a loopback recorder, and the run completed (exit 0). The overlay was rendered from the profile by hand rather than through the CLI's buildWrapEnv, so its base-config merge was not exercised",
  768. "fallback": "generic-env",
  769. "maintainer": null
  770. }
  771. ];

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.