J-Space Cognition Suite
Inference-time workspace, verification, and recovery protocols for agents handling deep reasoning and long-horizon tasks.
- Source repo
- Tiger3807861189/J-Space-Cognition-Suite
- Stars
- ★ 3k
- Last updated
- 1mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 73/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer using a Skill-capable coding host who must preserve architecture and naming constraints across several repository files.
- Not a fit if
- Chat or API hosts without a native Skill loader and no custom instruction injection
- Users who need the suite to supply a model, inference service, or task tools
- Teams without Python 3 or filesystem access that still want persistent loop state
- Source review
- 73/100 · Some gaps 1 safety controls not found
What does this agent do, and when should you use it?
J-Space Cognition Suite V3.6 is a model-agnostic inference-time control system for deep reasoning, long-horizon work, tool use, verification, and recovery. It is distributed as a Skill with one registered entry, j-space/SKILL.md, which selectively routes work to nine modules and three supporting references. Its fast, full, and loop passes scale from lightweight single-step handling to delivery checks and durable state for multi-stage tasks. The optional standard-library controller, j-space/scripts/jspace.py, records long-task state under .jspace/ in the active workspace without deciding which solution the model should choose. It does not alter model weights or training; its deployment boundary is either a host Skills directory or a chat/API integration capable of supplying system or developer instructions and retrieving files on demand.
A host loads the sole registered entry, j-space/SKILL.md, which selects fast, full, or loop according to task scope and retrieves only the relevant modules and references. The protocols organize accessible working representations, place shared names, values, constraints, and style anchors in a broadcast hub, and use bridge-before-conclusion reasoning, metacognitive signals, bounded empirical tests, and named verification coverage to advance work. For persistent tasks, j-space/scripts/jspace.py exposes note, seam, ship FILE, and resume: these commands write goals, next actions, hub entries, open questions, and checkpoints to .jspace/, inspect outgoing text, and reload state after a gap. verify_suite.py checks package integrity, while tests/test_jspace.py contains regression tests for the standard-library controller. The underlying model still produces the task result; J-Space supplies workflow routing, continuity, verification, and recovery rather than a standalone model or hosted inference endpoint.
- A developer using a Skill-capable coding host who must preserve architecture and naming constraints across several repository files.
- An engineering team conducting a multi-stage repository audit and wanting each finding tied to a verifier and stated coverage.
- An agent user running work across tools, turns, or long interruptions who needs a persistent ledger for reliable resumption.
- A researcher evaluating inference-time control without retraining a model or changing its weights.
- A chat or API integrator who can expose instruction files on demand and wants selective loading instead of placing the entire protocol in every context.
How do you install or deploy this agent?
You need Python 3 and the user-level Skills directory used by the host. Download or clone the repository, then copy the complete j-space/ directory into that location so the installed entry is <skills-directory>/j-space/SKILL.md; preserve modules/, references/, and scripts/. Verify it with:
<python-command> <skills-directory>/j-space/scripts/verify_suite.pyReplace <python-command> with an available python, python3, or py -3 command, and reload the host if it discovers Skills only at startup. If the host has no native Skill loader, provide j-space/SKILL.md as a system- or developer-level instruction and expose modules/ and references/ through file or retrieval tools; this is an integration path rather than a native installation. Include the repository's LICENSE and THIRD_PARTY_NOTICES.md when redistributing j-space/ as a standalone package. No credentials are documented as required.
How do you use this agent?
Invoke the installed Skill through the host's supported mechanism, such as a Skill picker, /j-space, $j-space, or a direct request: Use j-space for this task. Audit this repository, preserve its architecture, verify every finding, and keep the work consistent across all affected files. The entry gate selects the lightest suitable pass automatically. For durable state, keep the task workspace as the current directory and run the controller through its resolved Skill path: <python-command> <skill-root>/scripts/jspace.py note --goal "what done means" --next "first action". Continue with note --next to replace the next action, note --check "..." --by "..." to append a checkpoint, seam to reread recent state, ship OUTPUT_FILE to inspect outgoing text, or resume after a long gap. The controller records and reports state; the model remains responsible for choosing the solution.
What are this agent's strengths and limitations?
- Operates at inference time without retraining or changing weights, and the project reports reproduction across DeepSeek, Qwen, GLM, GPT, and Claude model families.
- Uses one entry and selective loading, so the nine modules and three references do not all occupy every task context.
- Its optional controller uses only the Python standard library and confines working state to
.jspace/inside the active task workspace. - Includes an integrity verifier, controller regression tests, and three-platform CI for checking the distributed package.
- Defines distinct
fast,full, andlooppasses that add delivery checks, ledgers, seams, and recovery only as task scope grows.
- Native use depends on a compatible Skill loader; ordinary chat or API deployments require custom instruction injection and on-demand file retrieval.
- Installation must preserve the full relative directory structure, because
SKILL.mdroutes to files undermodules/,references/, andscripts/. - Persistent loop state requires Python 3 and filesystem access, and creates a
.jspace/directory in the task workspace. - The published benchmark table combines project results with comparator records from their respective evaluation contexts, so it is not evidence from one fully standardized environment.
- The controller does not choose solutions and the suite does not supply a model, inference service, or task tools; outcomes remain dependent on the base model and host environment.
How does this agent compare with similar options?
J-Space is not a replacement for models such as DeepSeek, GLM, Kimi, or Claude; it is a control protocol layered onto an inference workflow. The project directly compares a DeepSeek V4-Flash-0731 baseline with the same model assisted by J-Space V3.6 and also lists published records for GLM-5.3, Kimi-K3, Opus-4.8, and Fable 5. Because comparator values retain their providers' evaluation contexts and configurations, the table is more useful as directional project evidence than as a strictly controlled model ranking.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| J-Space Cognition Suite This agent | 73 · Some gaps | Agent plugin / skillFree + model costs | ★ 3k | 1mo ago | Python | — |
| Harness AI Kit | 69 · Some gaps | CLIFree | ★ 25 | 1mo ago | Python | Codex · Claude Code |
| Pydantic AI Skills | 86 · Good | Library / SDKFree + model costs | ★ 379 | 2d ago | Python | OpenAI API |
| Kungfu | 81 · Good | CLIFree | ★ 4.5k | 1mo ago | C++ | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The controller is described as standard-library-only and confined to writing runtime state under the task-local .jspace/ directory, while the installation prompt requires asking before replacing an existing installation. This provides reasonably limited permissions and visible external effects. However, the supplied material defines no general confirmation gate for tool calls, sensitive tasks, or data processing; it also provides no policy for identifying, redacting, retaining, or deleting secrets and no complete rollback procedure. Local data flows are explained, but the trust boundary created by loading modules as system/developer instructions is not fully addressed. GitHub Actions dependencies use major-version tags rather than immutable commit pins, although the dependency surface is small. Research attribution, licensing, and third-party notice boundaries are thorough, justifying full source-attribution credit.
The README, structure description, controller lifecycle, and tests are mutually consistent. Integrity checks cover entry metadata, exact routing, and drift of important anchors, while regression tests cover corrupt ledgers, decoding failures, invalid input, and recovery behavior. Runtime availability is strong because only a common Python 3 standard library is required and three-platform CI is configured. Failure cases use nonzero exit codes and specific messages without traceback leakage, with direct test assertions. These criteria therefore have strong static evidence; tests were not executed, and lack of execution was not used to reduce these scores.
The source clearly separates fast, full, and loop modes for short tasks, multi-step work, tools, verification, and recovery, and describes both native Skill-loader and generic chat/API integration. Automatic selection of the lightest pass and on-demand module loading provide precise routing. The controller is explicitly bounded to recording state rather than choosing solutions, and environment-dependent effect sizes are acknowledged. Full boundary credit is withheld because there is no systematic account of unsuitable tasks, high-risk domains, or conditions under which activation should stop or be refused.
The bilingual entry points, project tree, single registered entry, module roles, installation procedure, command examples, and maintainer verification commands are well organized. Names such as j-space and fast/full/loop are stable, with integrity checks protecting key entry contracts. Examples are extensive but there is no structured FAQ. Limitations cover environmental variation, selective loading, and controller responsibility, but not a comprehensive set of known defects or operational risks. The complete Apache-2.0 license and redistribution guidance are present. The release sequence is documented, but per-version changes, migrations, and breaking changes are not. Unknown publisher identity is not itself penalized; maintenance responsibility remains thin because no maintainership roster, support policy, response commitment, or security-reporting path is supplied.
Mode guidance, controller commands, and copyable installation and invocation examples make the product practically usable, while selective loading, checkpoints, and recovery offer plausible marginal value beyond an ordinary single prompt. The README reports substantial benchmark gains, but the supplied evidence lacks the project's raw run records, per-task data, and reproducible evaluation materials, so marginal-value credit is limited. Efficiency figures represent one run per condition and use undisclosed scaling coefficients; without absolute elapsed time, token counts, variance, or the suite's context and cognitive overhead, cost-benefit support is thin.
The README identifies comparator sources, evaluation parameters, and environmental qualifications, helping distinguish conditional observations from universal claims. Nevertheless, the supplied files do not provide revision-linked raw records, per-item results, calculation scripts, or independent review for the J-Space benchmark numbers and cross-model reproduction claim. The tests verify controller and package behavior rather than effectiveness claims. External references can corroborate comparator context and terminology but not the project's claimed improvements, so traceability and cross-source corroboration receive only thin credit. Cautious language separates some observations from inference, although the move from project-level records to broad compatibility remains insufficiently supported.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- Do not treat the README's benchmark gains or cross-model reproduction statement as independently verified; obtain raw per-task records, evaluation scripts, exact run configurations, and repeated statistical results before adoption.
- Before installing this as system- or developer-level instruction material, manually review SKILL.md and every selectively loaded module and reference; their privileged instruction contents are not fully present in the supplied evidence.
- Do not use it directly in workspaces containing credentials, personal information, or regulated data unless separate redaction, retention, access-control, and deletion rules are established.
- The controller creates task-local .jspace/ state. In sensitive repositories, inspect that content, configure suitable ignore rules, and define backup or cleanup procedures.
- CI actions are referenced by major-version tags; high-assurance deployments should audit and pin them to trusted commit digests.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
sv1 J-Space Cognition Suite SV1
Release notes excerpt
新增 - 持久控制器 scripts/control.py — 门禁、任务地图、子代理、证据管理 - 宿主适配器 scripts/host_bridge.py — 把宿主事件转为上下文与放行决策，可移植适配，非厂商专属注册 - 四个新模块 — repository（仓库工程）、orchestration（子代理协作）、epistemics（认识论边界）、cyber（授权安全分析） - 三份新参考 — controller（控制器契约）、engineering-evidence（一手来源）、host-integration（宿主集成） - 五个新测试文件 — 含 CI 完整性校验与安全检查 变更 - SKILL.md 重构：272 → 177 行，入口更精简，路由更直接 - 全部 26 个原有文件均有修订 兼容性 - 无破坏性变更：14 个新增，26 个修订，0 个删除 - Python 3.10+（CI 覆盖 3.10 / 3.13） - 仅标准库，无网络访问 - 运行时只写 .jspace/，不修改 skill 本身 说明 - 概念 DOI 10.5281/zenodo.21971181 覆盖全部版本 …View on GitHub ↗ -
v3.7.4 J-Space Cognition Suite V3.7
Release notes excerpt
V3.7.4 hardens the controller and settles the audit: decline messages now echo the received --by (issue #52, reported by @raelldottin), and the hand-restatement rule is now stated wherever restating is taught (issue #17, reported by @menoxz). The benchmark tables move to a GLM-5.3-Flash baseline with vendor-official figures, third-party Terminal Bench 2.1 measurements, and clearly marked estimates. The controller gains six edge-case fixes with 32 regression tests, and CI moves to node-24 actions.View on GitHub ↗
FAQ
Does J-Space train or modify my model?
Does it require network access or API credentials?
.jspace/ storage.Are all nine modules loaded for every request?
SKILL.md selects the lightest suitable pass and retrieves relevant material on demand; fast loads nothing extra.What if my host does not support Skills?
j-space/SKILL.md as a system or developer instruction and expose modules and references through file or retrieval tools. That path requires host-side integration.How can I verify the package?
<python-command> <skills-directory>/j-space/scripts/verify_suite.py. Maintainers can also run <python-command> -m unittest discover -s tests -v from the repository root.