为什么值得关注

它把一个生产级编码代理的内部系统技能架构直接呈现出来，包括自我改进、自我知识、品味等模块，对理解现代 AI 编码工具的设计很有价值。

适合谁使用

构建或扩展 AI 编码代理的开发者
研究代理技能与提示架构的研究者
想了解 Sarvam Code 内置能力的使用者
对 OCR、递归 LM、自我改进模式感兴趣的工程师

典型使用场景

研究编码代理内置技能如何分层组织和渐进式披露
借鉴 SKILL.md、references/ 和 scripts/ 模式到自己的代理项目中
了解文档数字化、PDF 光栅化与递归 LLM 聚合的实践写法
深入阅读 Sarvam Code 的配置项与自我改进流程

项目优势

技能内容逐字提取自 v0.38.0，保留了真实结构和措辞
覆盖技能范围广：OCR、PDF 渲染、递归摘要、自我改进、配置参考、代码品味
采用渐进式披露布局，结构清晰且易于浏览
附有 Harbor 后端说明和完整 config.toml 配置项目录等值得深入阅读的参考文件

使用前须知

这是 v0.38.0 的快照，后续版本可能使它过时
所有技能内容版权归 Sarvam AI，并标注为供学习研究使用，复用受限
只包含内置系统技能，不覆盖其他 CLI 组件或用户自定义技能

README 快速开始

Sarvam Code — bundled system skills

The six system skills that ship inside the Sarvam Code CLI (sarvam-code, v0.38.0, a codex-rs fork), extracted verbatim from ~/.sarvam/skills/.system/. These are the agent's built-in capability + self-knowledge tree — the same progressive-disclosure SKILL.md + references/ + scripts/ layout Claude Code and codex use.

Skill	What it does
document-digitization	OCR an uploaded document to markdown/HTML across 22 Indian languages + English (upload → `digitize_document` w/ language code → poll → fetch).
document-rasterization	Render a local PDF's pages to PNGs so a vision model can read them ("summarize pages 3–10", "what's on page 5").
rlm-mode	"Recursive LM" mode for input-bound work — label/classify/aggregate/dedup/summarize over large sets by holding data in a program (`run_program`) and dispatching recursive `llm()` calls over slices, then combining.
self-improvement	The governor: black-box optimization of the agent harness (config + skills + prompt) around a fixed model, evaluated on the Harbor backend or in-process replicas, scored on a `(reward, cost)` frontier.
self-knowledge	Authoritative reference tree for operating/configuring/extending Sarvam Code itself — `config.toml`, models, features, memory, compaction, config layers, hooks, MCP, sub-agent roles, slash commands.
taste	The code-design/review bar — loaded before choosing a design for changes that carry design decisions (types, refactors, API changes, multi-file edits).

Notable references

self-improvement/references/harbor.md — how Sarvam runs the self-improvement loop on Harbor (harborframework.com): per-dataset config, oracle gate, ATIF trajectories, the two config layers, the document-eval failure taxonomy.
self-knowledge/references/config/*.md — the full config.toml knob catalog.
rlm-mode/scripts/summarize_map_reduce.ts — the map-reduce driver.

Extracted for study. All skill content is © Sarvam AI.

项目描述

The 6 system skills bundled inside the Sarvam Code CLI (v0.38.0), extracted for study

sarvam-code-skills