Install
Inside DeepSeek Harness, with dsh-market
dsh plugin --profile web add dshmarket
Or from the command line
dsh plugin --profile web add dsh-flakefinder
Installing runs third-party code with your own permissions — it can read your files, use your credentials and reach the network. Review the source first, and pin a commit (github:owner/repo#sha) when you can.
README
Teach agents the difference between broken code and a flaky test.
A DeepSeek Harness plugin for test stability: run tests repeatedly, classify flaky cases, keep history, maintain a quarantine manifest, and gate writes with approvals. Zero runtime dependencies.
Tools
| Tool | Purpose | Write |
|---|---|---|
flaky_detect |
Repeat tests N times and classify stable-pass / stable-fail / flaky | history only |
flaky_history |
Query detection history, filter by target | no |
flaky_report |
Combine quarantine manifest and history into a report | no |
flaky_quarantine |
Write flaky cases to .flakefinder.json |
yes (approval) |
flaky_clear |
Remove recovered cases from the manifest | yes (approval) |
Frameworks: vitest / jest / pytest (setup.cfg / pyproject / pytest.ini / tox.ini) / node:test, auto-detected in that order. JUnit XML entities and TAP SKIP directives are parsed.
Compatibility
Validation host: Harness 0.2.0-rc.1 built from official sources (commit 407e65c8) with Node 24.16.0 on 2026-09-28. All 44 plugin tests pass in an isolated environment; all 18 plugins mount together in one host registering 6 tools, with tool schemas and health-check contracts passing. No live ports or external services were exercised in this round.
2026-09-13 fix: retain the service receiver when calling subprocess.spawn, preventing failures caused by passing the method as an unbound callback. Verified against a real isolated host subprocess service. flaky_detect runs three real Node test processes through the host subprocess service and returns stable-pass.
Install
dsh plugin --profile web add dsh-flakefinder
Uninstall
dsh plugin --profile web remove dsh-flakefinder
Then restart the web service. To clean up fully, also remove the plugin entry from your profile cordis.patch.yml if you overrode it.
Example
User: src/checkout.test.ts keeps failing recently, is it real?
Agent:
flaky_detect(target="src/checkout.test.ts", runs=5)
-> verdict: flaky (3/5 passed; failures cluster on fake timers)
-> flaky_quarantine(tests=["src/checkout.test.ts > timer restore"], reason="timer race")
Config
See cordis.patch.yml; defaults: defaultRuns=5, maxRuns=20, timeoutMs=120000, writeApproval=true, pythonPath=python (python3 on Linux/macOS; used for pytest runs).
Development
pnpm test
pnpm typecheck
MIT
Comments
Comments live in GitHub Discussions. Sign in with GitHub to post or react.