v0.4 Beta Final Engineering Closure — 5 gates passed, 555 tests, 0 regressions.
A file-based knowledge workspace for Claude Code and compatible Agents. OKS helps Agents turn sources into reviewed, traceable and recallable knowledge.
Open Knowledge Studio is a lightweight knowledge engineering workspace for AI Agents.
It gives Agents a stable file-based memory layer: project knowledge, source evidence, failure lessons and human decisions survive across sessions instead of being re-explained every time.
The only recommended path (v0.4.0+):
Agent reads Source
→ selects Capability / Provider
→ produces EvidenceManifest
→ oks raw-commit
→ Raw Bundle
→ AgentObservation → Candidate
→ Human Review (mandatory gate)
→ Wiki
→ Search / Recall → Agent Output
Agents may write Candidates. Humans approve Wiki. Core never calls AI APIs. Provider never writes Raw directly. Recipe never hardcodes a specific Provider.
Requirements: Python ≥ 3.12, Git, pipx.
# Install the latest canonical main branch
pipx install open-knowledge-studio
oks --version
# Create a workspace (this also installs the Agent skills into it)
oks init ./my-knowledge-base
export OKS_ROOT=./my-knowledge-baseIngesting a source: in a supported Agent Host (Claude Code, Codex), just say:
收录这个 PDF
The Agent reads the /ingest skill, checks capabilities, selects Providers,
collects evidence, runs oks raw-commit, and writes a Candidate to drafts/.
Pure terminal (no Agent): oks ingest <source> creates a Run Workspace and
outputs instructions for Agent continuation. It does not invoke an Agent on its
own.
After human review:
oks drafts promote <slug>
oks recall "how should agent memory be managed?"
oks lintoks init Create knowledge workspace
oks raw-commit Commit evidence bundle to Raw (Fragment ↔ Manifest consistency enforced)
oks capability List / doctor / status available Providers (incl. user_impact metadata)
oks config Manage global config — including strategy (lightweight/quality/privacy/ask_each_time)
oks drafts Manage Candidate drafts
oks wiki Manage reviewed knowledge pages
oks recall Two-path episodic + knowledge recall (--knowledge-only, --type)
oks lint Quality scan
oks status Overview dashboard
| Feature | Description |
|---|---|
| Fragment ↔ Manifest consistency | 4 core fields validated (artifact_id, kind, method, agent_judgment); fail-closed on mismatch |
| ASR transcript semantics | kind=transcript + method=asr_transcription now accepted by schema |
| Guided Decision UX | Strategy config (oks config set strategy); 11 providers carry user_impact metadata; Strategy-Aware Ingestion in skill templates |
| Feishu 集成 | 参考实现见 examples/feishu-loop/（已从核心迁出） |
| 17 Providers | 25 capability actions, 7 modality recipes; provider.yaml + SKILL.md per provider |
555 tests passed (1 pre-existing env failure, 0 regressions).
OKS keeps the core lightweight. Heavy capabilities are installed on demand.
| Capability | Purpose |
|---|---|
document |
Office (docx/pptx/xlsx), HTML, CSV |
pdf-lite |
Lightweight text-layer PDF extraction (pymupdf4llm) |
pdf |
Full PDF extraction (MinerU, ~300 MB) |
watch |
Video, audio, subtitle and OCR extraction |
formula |
PaddleOCR formula candidates |
OKS is Claude Code-first, but not Claude Code-only. It works with any Agent that can read files, run commands and follow project rules.
Do not build a new platform when existing Agent capabilities already do the job.
- Core Architecture
- Capability Boundaries
- Agent-Native Ingest Walkthrough
- Kimi K3 Deep Analysis
- Security — Remote Governance
In v0.4.0 the legacy extractor pipeline was removed:
--legacyflag — removedOKS_ENABLE_LEGACY_PROVIDERSenv var — removedraw_bundle_adapter.py,source_router.py, old extractors — permanently deletedoks ingest <source> --mode quick— replaced by Agent-native/ingestskill
Legacy code is preserved in Git tag v0.4.0-legacy-final.
MIT
v0.4 Beta 最终工程收口 — 5 个验证 Gate 全部通过，555 项测试，0 回归。
Open Knowledge Studio 是一个面向 Claude Code 和兼容 Agent 的文件式知识工作区。
它让 Agent 把外部资料、项目经验、失败教训和人工判断沉淀成可追溯、可审核、可召回的长期知识， 而不是每次新会话都重新解释上下文。OKS 不试图替代 Obsidian、Notion、Roam 或用户已有的编辑器——它负责的是：把用户已有的文件、网页、媒体、平台内容，经 Agent 提取、人工审核后，沉淀成可召回的文件系统知识。
Agent 读取 Source
→ 选择 Capability / Provider
→ 生成 EvidenceManifest
→ oks raw-commit
→ Raw Bundle
→ AgentObservation → Candidate
→ Human Review（强制人工门禁）
→ Wiki
→ Search / Recall → Agent Output
Agent 可以写 Candidate。人工审核后进入 Wiki。Core 不调用 AI API。 Provider 不直接写 Raw。Recipe 不写死具体 Provider。
pipx install open-knowledge-studio
oks --version
# oks init 会同时把 Agent 技能装进该目录
oks init ./my-knowledge-base
export OKS_ROOT=./my-knowledge-base收录来源：在支持的 Agent Host（Claude Code / Codex）中说：
收录这个 PDF
Agent 读取 /ingest Skill、检查能力、选择 Provider、采集证据、执行
oks raw-commit，最后将 Candidate 写入 drafts/。
纯终端（无 Agent）：oks ingest <source> 创建 Run Workspace 并输出 Agent
接管提示，不自行调用 Agent。
人工批准后：
oks drafts promote <slug>
oks recall "how should agent memory be managed?"
oks lintv0.4.0 已移除旧 extractor 链路：
--legacy标志 — 已移除OKS_ENABLE_LEGACY_PROVIDERS环境变量 — 已移除raw_bundle_adapter.py、source_router.py、旧 extractor — 已永久删除oks ingest <source> --mode quick— 由 Agent-native/ingestSkill 替代
旧代码保留在 Git tag v0.4.0-legacy-final 中。