dsh-doublecheck plugin
A delivery quality gate plugin for DeepSeek Harness: requirement questions, red and green tests and a ship or rework verdict.
- 58 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The plugin watches session edits and tests, can stop work in `block` mode, and the critic mode starts an extra session
Install
Manual install
dsh plugin --profile <name> add dsh-doublecheckThe two plugin rows attach to the profile automatically. From Git: `dsh plugin --profile web add github:PerryLink/dsh-doublecheck`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The plugin watches session edits and tests, can stop work in `block` mode, and the critic mode starts an extra session
This is third-party code. Review the repository files before installing.
What it does
dsh-doublecheck turns the phrase "done" into a verdict with evidence. The plugin is two lines for DeepSeek Harness: `doublecheck-grill` interrogates requirements and records a spec, and `doublecheck-guard` stands at the boundaries of edits, tests and handoff. Both read one replayable session log, so after a resume, fork or restart they see the same facts. The `/gate run` command folds what happened into four stages: whether requirements were confirmed item by item, whether the latest test is red or green, whether each change serves a spec dimension and what the review concluded. The result is a ship or rework verdict as a report to paste into a pull request description. The plugin writes no code and does not judge code beauty. Intensity is chosen from `remind`, `warn` and `block`.
Who it is for. DeepSeek Harness users who want delivery discipline: requirements and a test first, then code, with evidence before handoff.
Good fit when
- An agent declared work done but no test ran after the last edit
- A vague task must first become a spec through questions
- Edits outside the spec should show in the report as a finding
- A pull request needs a report with a verdict that can be pasted into its description
Not a fit when
- You do not work in DeepSeek Harness, since the plugin is written for it and does not run outside it
- An assessment of code quality and style is needed, while the plugin checks only process
- Strict blocking from the first minute without setup is needed, for which a separate overlay exists
Example request
Run /gate status, then /gate run and show the verdict: whether requirements are confirmed and whether the last test is red or green.Limitations
The plugin is built for DeepSeek Harness (the `dsh` command) and works only in it. It is an unofficial community plugin with no relation to DeepSeek. Documentation is mostly in Chinese, with English, Spanish, Portuguese and Hindi versions. The default intensity `remind` only reminds, and blocking is enabled in the configuration. Settings are `Config` fields in `cordis.yml`, with 39 keys, and an override by id replaces the whole row. The adversary mode `modules.adversary` is off by default and starts a separate critic session after a green test.
How to disable. Run `dsh plugin --profile web remove dsh-doublecheck` or delete the two rows `doublecheck-grill` and `doublecheck-guard` from the profile patch.
FAQ
Which install channels exist?
Three: `dsh plugin --profile web add github:PerryLink/dsh-doublecheck` from Git, `add dsh-doublecheck` from npm and from a local tarball after `pnpm pack`.
What do remind, warn and block mean?
They are enforcement strengths: `remind` reminds, `warn` warns and `block` stops progress without evidence. Set with the `intensity` key.
How do I turn the check on or off in a session?
With `/doublecheck on` or `/doublecheck off`. Details come from `/doublecheck status` and `/gate status`.
Where is the final report?
In the workspace file `gate-report.md`, while the spec is kept in `doublecheck-spec.md`. Names change with the `gate.reportFile` and `specFile` keys.
What does the critic do?
With `modules.adversary: true` a read-only forked critic starts after a green test and adds findings to the session.
Discussion
No questions or reviews yet.
Related
PluginsA Claude Code plugin that makes an agent check the code around a change and prove the work with a test and independent review
A CLI that rates the maturity of an AI agent harness in a repository from L0 to L4: 36 file checks with no model calls or network.
An adversarial code-audit skill: Hunter finds bugs, Skeptic challenges them, Referee rules, and no source edit happens without your permission
A Claude Code plugin that verifies a change against its plan: boots your stack, runs acceptance criteria, reports evidence.
A skill that reads a repository once and installs checks: code rules become gates that fail a commit or a CI run.
A Claude Code hooks plugin that blocks code without a failing test and implementation beyond the current test
README in short
Evidence comes from the session log, not from the agent’s words. Requirements, the test, edit consistency and review fold into one report.