J-Space-Cognition-Suite-V3.6

Four Ways Long Tasks Quietly Go Wrong

Long agent tasks tend to fail at seams. A plan can drift away from the original requirement. A repository map can become stale after edits. Delegated work can arrive without evidence. A final answer can sound complete even when an acceptance condition is still open.

J-Space addresses those failure modes with a shared control record, source refreshes, evidence reports, and delivery gates. It does not change model weights or guarantee better reasoning. Its value is procedural: it makes important state inspectable and asks the host to stop when required evidence is missing.

Try prompts such as:

Use j-space at medium level to update this API example and verify every parameter against the implementation.

Use j-space at high level to repair this repository issue. Preserve public contracts and finish with evidence for each acceptance condition.

Working-Set Overload and Representation Drift, Explained

A long task contains more than chat history. It contains the goal, the files that define current behavior, unresolved questions, decisions, tests, and outputs. If those pieces live only in transient context, later steps can operate on an outdated representation.

The suite keeps durable state outside the prompt. Its controller can record source receipts, a semantic repository map, agent ownership, reports, and completion evidence. The host still has to read the real files and run real checks. The stored map is a navigation aid, not proof that the code behaves as described.

For adjacent workflows, compare Vibe-Skills for Skill routing and mem9 for cross-session memory.

How Verification Discipline Catches Premature Completion

Use a short gated loop:

1. Define the acceptance conditions before editing.
2. Refresh the relevant source files and update the task map.
3. Execute the change and record observed checks with their evidence.
4. Run the delivery gate and resolve every blocking result before shipping.

Example input with an explicit output shape:

Audit the authentication change. Return: requirement, source evidence, observed test, unresolved risk, and final PASS or FAIL.

This structure makes a missing test or stale source visible. It does not convert an untested assumption into evidence.

Installing J-Space in a Model-Agnostic Workflow

Use the J-Space quick start and keep the complete directory structure intact. Python 3.10 or later is required for executable controllers and validation.

1. Download or clone the repository, then copy its complete j-space/ directory to the DSH Skills directory. For the default user-level layout, the entry is ~/.dsh/skills/j-space/SKILL.md.
2. Keep modules/, references/, and scripts/ beside that SKILL.md; avoid a nested j-space/j-space/ path.
3. Verify the installed copy with python3 ~/.dsh/skills/j-space/scripts/verify_suite.py.
4. Reload DSH if it discovers Skills only at startup, select j-space, and confirm that a routed module and controller --help output can be read.

If your DSH configuration uses another Skills root, substitute that documented path in both commands.

Host-status check on 2026-09-28: DeepSeek Harness is still a developer preview, and upstream says compatibility-breaking changes will occur. Reconfirm Skill loading after each host upgrade.

Carry Verified Work Into EvoMap

EvoMap is separate infrastructure for organizing verified agent assets; J-Space does not publish to it by itself. After a controlled task is validated, a team can explicitly package its reusable strategy or outcome and review the EvoMap Wiki. Compare brooks-lint when the asset includes engineering-review evidence.

FAQ

Is J-Space a model fine-tune?

No. It is an instruction, state, and verification suite used at inference time. The project explicitly says it does not modify model weights.

Does it guarantee that a long task is correct?

No. Gates can expose missing evidence and stale state, but correctness still depends on the sources, tests, and human decisions supplied to the workflow.

Is V3.6 the upstream project name?

The project is named J-Space Cognition Suite and labels its architecture SV1. V3.6 remains the catalog route label for this listing.

Can I use it without Python?

The documented prose fallback can cover lighter use, but the strict controllers and validation require Python 3.10 or later. Pair long tasks with Vibe-Skills when Skill routing is also needed.

Related Pages

• Route tasks with Vibe-Skills
• Review code with brooks-lint
• Share memory through mem9

Reuse a Verified J-Space Workflow

Turn a verified J-Space workflow into reusable agent experience with clear evidence and acceptance conditions.

Browse reusable assets on EvoMap Compare Vibe-Skills routing