flaqai Review a DeepSeek Harness Plugin
Produce prioritized, evidence-backed findings for the exact revision under review.
Workflow
- Record the plugin and Harness revisions, source, license, package metadata, lockfile state, install scripts, and build provenance.
- Resolve the full configuration order: Profile, Bundles, local Patch, user Patch, and command-line overrides.
- Trace plugin entry points,
inject, provided Services, Events, Effects, Session writes, remote APIs, and client code. - Review correctness: dependency availability, schema validation, error propagation, cancellation, deterministic ordering, and replay behavior.
- Review lifecycle: listener removal, Service withdrawal, timers, file handles, subprocesses, network connections, temporary files, and remount behavior.
- Review security: credentials, filesystem scope, subprocess arguments, generated code, network destinations, telemetry, data retention, approvals, sandbox enforcement, and supply chain.
- Review tool contracts: strict schemas, authorization, idempotency, output bounds, redaction, canonical results, and model/UI rendering.
- Run existing tests and the narrowest safe reproduction. Use a disposable Profile for installation tests and inspect
--dump-config. - Report findings by severity with exact file and line evidence, impact, triggering conditions, and a bounded remediation.
Severity Guide
- Critical: likely credential disclosure, arbitrary code execution across an asserted boundary, or destructive behavior without meaningful control.
- High: realistic unauthorized side effect, persistent data corruption, replay breakage, or resource leak affecting other plugins.
- Medium: correctness or compatibility defect with a bounded trigger or recovery path.
- Low: maintainability, observability, or documentation gap that materially increases future risk.
Mandatory Questions
- Does every required Service appear in
inject, and what happens when it disappears? - Does every owned Effect dispose on unload, failure, and remount?
- Is every model-visible state transition reconstructable from durable Session events?
- Are dependency visibility, policy approval, and OS sandboxing described separately?
- Can installation or startup execute unreviewed code?
- Are secrets excluded from logs, Events, errors, tool results, and UI payloads?
- Is the package compatible with the exact preview revision being used?
Safety Boundaries
- Keep review actions read-only unless the user explicitly requests fixes.
- Never execute untrusted install scripts or plugins merely to inspect them.
- Never include real credentials or private Session content in findings.
- Do not infer safety from popularity, inclusion in a list, or a clean manifest.
Completion Check
If no actionable issue is found, state that explicitly and list residual risks, untested paths, and revision assumptions. Do not convert uncertainty into a clean bill of health.