⚠️ This repository has been renamed: XiaoSHang05/dsh-rice-guard — click to visit the new location. Everything below is preserved as originally indexed.

@XiaoSHang05/dsh-loop-guard avatar

@XiaoSHang05/dsh-loop-guard

A passive watchdog for DeepSeek Harness: watches a turn's thinking and text streams and cancels it when the model starts repeating itself.

PluginAI ModelsAgent orchestration
Verification
L4 · Install tested
stale
Risk
Out of scope
Health
Active
Trust
Silver

Install verification: sandbox install ok, web smoke failed (L5) · tested on dsh 0.2.1-alpha.2 · last checked Install verification details →

What it doesAI

Watches a DeepSeek Harness turn's thinking and text streams and cancels the turn when the model starts repeating itself.

  • Passively monitors a turn's thinking and text streams
  • Detects model self-repetition and cancels the runaway turn
  • Ships as a dsh Cordis watchdog plugin
aiAIagent-orchestrationAI

Capability tags are AI-inferred or rule-extracted from the README (see badge) — vocabulary-limited, not install-verified. AI = inferred; README = literal keyword from the readme.

Installation

dsh plugin --profile web add github:XiaoSHang05/dsh-loop-guard
verified

Install method: GitHub · sandbox install ok, web smoke failed (L5) · verified on dsh 0.2.1-alpha.2

Verification & Compatibility

Grouped by plugin version: each card lists the dsh versions actually tested against that artifact version and their install verdicts (L5 wins). Untested combinations are omitted — absence means untested, never assumed compatible. L1–L3 are a plugin-level static archive (independent of the dsh version — the same verdict on every dsh row): they sit in each row's level grid next to L4/L5, marked "static", while their evidence is stored once per plugin inside that row's "checks & evidence" instead of being duplicated per version.

External credentials (plugin-level): none detected

Declared DSH compatibilityfrom package.json · peerDependencies
author declared no dsh version requirement

○ The author declares no dsh version requirement — install is not blocked by version.

Footnote: the author also wrote engines.dsh = ">=0.2.0-0 <0.3.0" — DSH does not read that field (not part of the verdict).

Install compatibility matrix (rows = plugin versions〔latest 5〕, columns = dsh versions〔latest 5〕)
At-a-glance → evidence in the cards below
plugin \ dshdsh 0.2.1-alpha.2
v0.2.0·
Install compatibility (by plugin version, dsh versions within)
Legend:✓ L5 runtime verified○ ecosystem plugin◼ ecosystem app✗ failed
v0.2.0github2026-10-14

https://github.com/XiaoSHang05/dsh-loop-guard

No runtime install test for this version (static scan only) — compatibility unknown, never assumed.

Only the latest 3 published versions are shown; source-snapshot (commit hash) groups, unstated groups and older versions are omitted — the full archive stays in the data layer (snapshot/unstated groups appear only when a plugin has no published version).

Author badges

Embed the official badges in your README to showcase security & install-test status:

dsh.so risk
`[![dsh.so risk](https://www.dsh.so/badge/dsh-loop-guard-5.svg)](https://www.dsh.so/artifact/dsh-loop-guard-5/)`

Security Report

Two automated layers, neither a manual review: a static scan (per-plugin rule checks) and a model-driven dynamic security audit that covers only the subset static-flagged at warn or above and never executes plugin code.

dsh-loop-guard-5Automated review · daily

Static low, dynamic low — stricter wins → overall low.

static low→dynamic low→overall low
0 critical·0 high
lowoverall · low
1static scan · vet primaryvet AST static analysis (npm artifact or git source)
plugin versionv0.2.0dsh manifest
scanned versionv0.2.0git2026-10-14
current versionv0.2.0same version
vet verdictclean · git

Vet is an AST static scan (npm artifact or git source).

show static findings (5)
  • criticaleval · code-exec_probe/browser/verify.mjs:19

    const box=await page.$eval('.lg-root',node=>{const r=node.getBoundingClientRect();return {width:r.width,height:r.height}})

  • criticalnewFunction · code-exec_probe/client/load.mjs:6

    new Function('window',readFileSync(new URL('../../lib/client.js',import.meta.url),'utf8'))(window)

  • warningchildProcess · code-exec_probe/mutation.mjs:47

    if(file.includes('client')) assert.equal(spawnSync(process.execPath,['scripts/build-client.mjs'],{cwd:root,stdio:'ignore'}).status,0)

  • warningchildProcess · code-exec_probe/mutation.mjs:48

    const run=spawnSync(process.execPath,[test],{cwd:root,stdio:'ignore',timeout:15000})

  • warningchildProcess · code-exec_probe/mutation.mjs:53

    if(file.includes('client')) assert.equal(spawnSync(process.execPath,['scripts/build-client.mjs'],{cwd:root,stdio:'ignore'}).status,0)

2model reads redacted source context · never executes plugin code
highest gradelow
merged intooverall low
audited at2026-10-16
findings0 critical·0 high0 medium / 3 low hidden
show 3 audit findings
  • lowDev probe reads the local DSH browser-session secret and mints a cookie_probe/env.mjs:65

    The probe helper reads ~/.dsh/.credentials.yaml, extracts the browser-session HMAC secret and signs a valid local web session cookie. This is developer tooling invoked manually (node _probe/*.mjs); nothing in the plugin runtime imports it, the cookie is authority-bound and is only sent to the local host (DSH_WEB_URL, default 127.0.0.1:19387) — no exfiltration channel and no attacker-controlled input.

  • lowTest loader evaluates the built client bundle with new Function_probe/client/load.mjs:6

    Runs the project's own build artifact (lib/client.js) inside the probe process. The input is a local file produced by scripts/build-client.mjs, not remote or attacker-supplied data, and the probe only stubs a __ModuleLoader__ shim; no runtime path of the installed plugin uses this loader.

  • lowMutation harness rewrites plugin sources and spawns node, restoring them in finally_probe/mutation.mjs:45

    The script patches src/index.js, src/detect.js and src/client.js, re-runs scripts/build-client.mjs and the probe tests via spawnSync, then restores the original bytes in a finally block. Targets, patterns and commands are hard-coded local constants; it is dev-only tooling a developer must invoke deliberately, and it corrupts nothing in the pinned plugin runtime.

Static-scan criticals sit in dev-only probe tooling (local credential read, dynamic eval); no plugin runtime entry point loads them, and no runtime-reachable risk was found in the harvested evidence.

Rule: the dynamic audit can only raise risk, never lower it (raise-only, never an exemption); a wrong raise is withdrawn by human review and the tier falls back to the static verdict. Overall verdict = the stricter of the two layers.

Disclaimer: the static scan and the dynamic security audit are both automated analysis, not a security guarantee. Always review what you install.

Activity

Last commit 2026-10-14 · activity: Active

• Repo created: 2026-10-13

• Stars: ★ 0 · Forks: 0

• Health: Active — committed within last 30 days

Source

GitHub: github.com/XiaoSHang05/dsh-loop-guard

Was this page helpful?