dsh-policy-drift-proof
Curated pickMaintenance: Activedongsheng123132/dsh-policy-drift-proof
Content-addressed, value-redacted policy drift evidence for DeepSeek Harness
Install
dsh has no central install command — add this plugin’s entry (documented in its README below) to your profile or patch config, then restart.
How installs work3
stars
0
forks
JavaScript
Language
MIT
License
2026-08-14
Created
2026-08-20
Last push
README
dsh-policy-drift-proof
Read-only, content-addressed policy/configuration drift evidence for DeepSeek Harness.
Version 0.2 adds host-neutral DSH ToolDefinitions, a proof-only inline Codex MCP surface, real ToolRuntime calls and a stock Web Loader regression. The package exposes namespace exports only and does not bundle a second DSH runtime.
This plugin does not enforce tool calls, approve actions, scan repositories, or repair configuration. dsh-tool-policy already provides pre-execution policy routing, while SecurStack provides security scans and policy gates. This plugin covers the missing evidence question: did the policy snapshot actually observed by an operator differ from the pinned baseline, and was the difference weakening, tightening, exact, or unclassified?
Evidence model
The explicit manifest pins a baseline and observed policy-snapshot/v1 by SHA-256 and revision. Rules cover declared JSON Pointer roots:
ordered-not-weaker: enum order is restrictive to permissive; moving right fails.set-no-additions: additions fail, removals are recorded as tightening.exact: any change fails.- any changed covered leaf without a rule fails closed as
UNCLASSIFIED_DRIFT.
Reports include paths, classifications and SHA-256 digests only. They never include policy values. Secret-shaped fields, raw output fields, absolute/escaping paths, symlinks, oversized inputs and excessive structure are rejected. The verifier performs no network calls or child processes and writes only one content-addressed JSON report beneath the explicit artifactDir, then reads it back and verifies its digest.
verified means both snapshot hashes and revisions matched and no weakening, exact or unclassified violation was found. Tightening remains observable through driftStatus: "changed". This is evidence about supplied snapshots, not a security certification.
CLI
node bin/dsh-policy-drift-proof.mjs inspect \
--workspace examples/basic \
--manifest policy-drift.manifest.json
node bin/dsh-policy-drift-proof.mjs verify \
--workspace examples/basic \
--manifest policy-drift.manifest.json \
--artifactDir artifacts
The CLI prints one JSON object to stdout. A drift violation exits 2; invalid usage exits 1.
DeepSeek Harness bundle
dsh plugin --profile policy-proof add github:dongsheng123132/dsh-policy-drift-proof#<commit>
dsh profile --profile policy-proof compose
The bundle registers:
dsh_policy_drift_inspectdsh_policy_drift_verify
The companion stdio MCP server exposes policy_drift_inspect and policy_drift_verify through .mcp.json. MCP accepts an inline manifest and pinned inline snapshots, performs no filesystem/network/subprocess/write operations, rejects secret- and raw-output-shaped fields, and never returns policy values. DSH and CLI remain the surfaces for workspace-bound files and content-addressed report publication.
Verification
npm test
npm run check
npm run smoke:plugin
npm run smoke:mcp
DSH_CHECKOUT=/path/to/built/deepseek-harness npm run smoke:dsh
DSH_CHECKOUT=/path/to/built/deepseek-harness DSH_HOME=/path/to/isolated-home npm run smoke:web-loader
python C:/Users/ZhuanZ/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py .
CI runs tests, structural checks and MCP smoke on Ubuntu and Windows. Node.js 22 or newer is required.
License
MIT
More in Security & Auth
tencentmeeting-cli
by tencentcloud
Tencent Meeting command-line tool with OAuth2 authorization for meeting, recording and attendance-report management.
dsh-auto-review
by perrylink
Second-model AI auto-review for DeepSeek Harness approval requests: a read-only reviewer subagent returns structured allow/deny verdicts with reasons, fail-closed by default, fully auditable from the session log (approval/asked -> autoReview/verdict -> approval/decided).
dsh-permission-rules
by perrylink
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
dsh-clawrouter
by blockrunai
A safety gate for DeepSeek Harness: a stronger model reviews dangerous tool calls before they run. Plus vision and 67 models from one wallet, paid per request over x402.
