AI Agent Hub
Back to plugins
🤖

falsify-dsh

Model Inference Updated 2026.08.15

Run the following command in DeepSeek Harness:

dsh plugin install shi275773124/falsify-dsh

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install shi275773124/falsify-dsh inside DeepSeek Harness to install this plugin; the full source is at https://github.com/shi275773124/falsify-dsh

About this plugin

In a fast-moving DeepSeek Harness workflow, an agent saying "looks fine" is not a receipt, and a green lint is not a ship. Between the model"s self-assessment and the release gate there has been no independent adjudicator. falsify-dsh fills that gap: it is a thin, purpose-built adapter that plugs the public MIT Falsify CLI into DSH and makes the verdict come from the CLI, not from the model.

The adapter exposes three tools—falsify_lint for static tag and blocker checks, falsify_review for epistemic document review, and falsify_gate for a production gate receipt. The CLI is the sole adjudicator; the adapter only supplies file paths, invokes python -m falsify, and returns the verbatim PASS / PASS_WITH_DEBT / BLOCK verdict. It never invents a green: if the CLI is missing the result is CLI_ERROR. capital_authority is always NONE, the adapter opens no network sockets of its own, and it uploads nothing.

It is best for engineers and teams working inside DSH who need a verifiable, auditable delivery receipt before anything ships—replacing "the agent says it is fine" with a concrete, independently issued gate signal.

Use Cases

  • Requiring an objective go/no-go signal that is independent of agent self-assessment before merging a PR
  • Static lint is fully green but a document-level epistemic review is still needed before release
  • Replacing the verbal assurance of "the agent says it is fine" with an auditable, CLI-issued receipt

Best For

  • Engineers who develop daily inside DeepSeek Harness
  • Tech leads who need a traceable, auditable delivery pipeline
  • QA or SRE roles who insist on an independent adjudicator for any model-produced artifact