J-Space Harness
简体中文
@spy2006/dsh-jspace-harness packages the J-Space Cognition Suite V3.6 as a
DeepSeek Harness agent preset. It adapts the upstream protocol to the
Harness runtime the right way: Harness-native scoped prompt sections,
per-agent state isolation, request-cache-stable turn checkpoints, and durable
V3.6/Harness verification — instead of guessing provider payloads or mutating
the conversation transcript the way the pi adapter does.
What this preset gives you
| Capability | Harness-native behavior |
|---|
| Cognitive protocol injection | Registers one scoped systemPrompt.section() evaluated per assembly with the calling agent's context (dsh-plan-mode/dsh-persona pattern). No transcript pollution, no provider-payload mutation. |
Gate (fast / full / loop) | Classified once per turn, persisted to a per-session ledger file, and surfaced in the prompt section. |
| Per-agent state isolation | Ledger keyed by session.id under $DSH_HOME/j-space/<session-id>.json with a per-agent WeakMap cache; sessions never share state. |
| Cache stability | The system-prompt section stays byte-identical across a turn's steps: pass is classified once per turn and checkpoint advances only at turn boundaries. The stable prefix [system + history] can be reused by the request cache; only the new tail is re-encoded. |
| Prompt construction cost | compact and reminder protocol text is built once at load; full reads SKILL.md lazily and caches it (no re-read per assembly). |
| State write resilience | Ledger writes are non-fatal, use a unique temporary file (<pid>.<ts>.tmp), and never interrupt an agent turn on disk failure. |
| Verification | Upstream V3.6 anchors (PREMISE, INVARIANTS), strict frontmatter/routing checks, and Chinese claim/coverage vocabulary — enforced by two durable verifiers. |
| Runtime control | /jspace command: status, on, off, compact, full, reminder. |
Install
Via DSH Plugin Market (dshmarket v1) — recommended
The package is published on npm: @spy2006/dsh-jspace-harness@0.1.1.
- Make the catalog entry visible:
- DSH 1024Store (auto-collected): this repository carries the
dsh-plugin
topic and its default branch already satisfies the 1024Store static check
(package.json with dsh.bundle.patch + cordis.patch.yml in the same
tree), so it is picked up by the pipeline's incremental scan. Search
jspace in the 1024 Store tab of Plugin Market.
- Custom standard source: deploy
market/worker.mjs
(npx wrangler deploy), then register
https://<worker>/catalog-source.json in Plugin Market → Sources.
- Open the entry in DSH Plugin Market, Preview (the host re-validates the
exact npm identity, repository backlink, lifecycle scripts, engine, and DSH
bundle evidence), confirm, and restart DSH Desktop.
- On first launch
index.js copies preset/ into
$DSH_HOME/.agent-presets/jspace-harness. An existing user-owned preset is
deliberately not overwritten, so local customizations survive.
- Start a new session and pick the J-Space Harness preset.
Manual install
git clone git@github.com:2006spy/jspace-harness.git
# copy the preset directory into the DSH user preset root:
robocopy preset %USERPROFILE%\.dsh\.agent-presets\jspace-harness /E # Windows
cp -R preset ~/.dsh/.agent-presets/jspace-harness # macOS/Linux
Restart DSH and select J-Space Harness in the new-session preset picker.
Usage
In any session on the preset, /jspace controls the runtime:
| Command | Effect |
|---|
/jspace or /jspace status | Show enabled state and mode |
/jspace on / /jspace off | Enable / disable the cognitive layer |
/jspace compact | Default: ~1.5k-char protocol block; modules are read on demand |
/jspace full | Inject the whole SKILL.md (truncated at 18 000 chars); use sparingly |
/jspace reminder | Lightest: one gate line only, best for long sessions |
State is stored per session in $DSH_HOME/j-space/<session-id>.json
(override with JSPACE_STATE_DIR). The bundled skill (skills/j-space)
provides the optional jspace.py controller (note / seam / resume /
ship) and its regression tests.
Benchmarks (measured on this machine)
Method: Python 3.12.10, cold subprocess timings, fresh temporary workspace per
controller lifecycle, median of N rounds. Raw data lives in
benchmarks/.
Controller lifecycle (upstream V3.6 vs this preset)
Full ledger lifecycle (note goal/core/open/close + seam + ship with a
Chinese covered verification claim), 20 rounds:
| Condition | Median | Mean | Min | Max |
|---|
Upstream V3.6 jspace.py | 800.299 ms | 814.206 ms | 774.202 ms | 914.101 ms |
This preset's jspace.py | 781.112 ms | 780.362 ms | 749.431 ms | 836.777 ms |
This preset is ~2.4% faster on the median lifecycle despite carrying the V3.6
controller logic, strict verifier, and Chinese claim/coverage vocabulary.
Adapter cold parse (pi adapter vs this preset)
node --check of each adapter entry, 20 rounds:
| Condition | Median | Mean | Bytes |
|---|
pi adapter (tonyxu721/pi-j-space) | 206.535 ms | 200.709 ms | 12 175 |
This preset's preset/j-space.mjs | 205.755 ms | 212.285 ms | 10 852 (-10.9%) |
Same-model protocol smoke comparison (10 tasks, blind-scored)
Same model, no tools, one fixed rubric (0–2 correctness + 0–1 verification per
task, max 30). Upstream generic protocol vs this preset's forced gate/ledger:
| Condition | Correctness | Verification | Total |
|---|
| Upstream generic protocol | 20 | 10 | 30/30 |
| Harness forced protocol | 20 | 10 | 30/30 |
Verdict: tie — Harness forcing does not regress correctness or verification
coverage on short deterministic tasks. The task set is intentionally too short
to claim a capability separation.
Verification suite
- Upstream V3.6 regression tests: 18/18 pass
verify_suite.py (upstream anchors + strict frontmatter/routing): cleanverify_harness.py (Harness adapter contract): clean
Reproduce
python benchmarks/run_jspace_bench.py # engineering timings -> engineering-results.json
python preset/skills/j-space/scripts/verify_suite.py
python preset/skills/j-space/scripts/verify_harness.py
Scope of these numbers
Engineering measurements describe local controller/adapter behavior, not an
LLM capability score. The same-model comparison is a 10-task protocol smoke
test. Provider-level prompt-cache hit rate and long-horizon durability need a
fixed provider/model, temperature/seed, task corpus, and token telemetry.
Discovery in DSH Plugin Market
- DSH 1024Store (auto-collected): the pipeline scans GitHub repositories
carrying the
dsh-plugin topic and validates package.json +
dsh.bundle.patch + patch file on the default branch. This repository
satisfies all checks, and the published npm package with a matching
repository backlink makes the entry installable through the managed path.
See https://github.com/imsai-sh/awesome-deepseek-harness-plugins.
- Custom standard source: deploy
market/worker.mjs and register
https://<worker>/catalog-source.json. See market/README.md.
- dshfind is a third-party read-only index; listing there is controlled by
dshfind, not by this repository.
Development checks
python preset/skills/j-space/scripts/verify_suite.py
python preset/skills/j-space/scripts/verify_harness.py
node --check index.js
node --check preset/j-space.mjs
npm pack --dry-run
Project layout
jspace-harness/
├── index.js # plugin entry: copies preset/ into .agent-presets on first launch
├── cordis.patch.yml # DSH bundle patch (market installer mounts the plugin row)
├── package.json # npm package: dsh.bundle.patch, no lifecycle scripts
├── preset/ # the agent preset (agent.cordis.yml + j-space.mjs + skills/j-space)
│ └── skills/j-space/ # SKILL.md, 9 modules, references, jspace.py, dual verifiers
├── market/ # deployable Cloudflare Worker standard-source catalog
├── benchmarks/ # reproducible engineering + protocol comparison data
└── .github/workflows/ # CI: verifiers + syntax + npm pack
Attribution
The bundled J-Space skill derives from
J-Space Cognition Suite V3.6.
See THIRD_PARTY_NOTICES.md and LICENSE.
