nopus

nopus catches coding-agent answers that disappear into abstract LLM babble. Using deterministic prose checks, it flags responses that cross your chosen complexity threshold and sends them back for one clearer rewrite.

See it in action

Currently supported: Pi, Claude Code, Codex

Use nopus

To use nopus, install Node.js 22 or newer and make sure the node command is available.

Install nopus from npm:

pi install npm:@syzom/nopus

You can install the GitHub repository instead:

pi install git:github.com/Vistyy/nopus

Start Pi normally after installation. The extension checks each completed response. When it requests a rewrite, it hides the rejected response from the terminal transcript by default. The response remains in the Pi session and model history. Because nopus evaluates completed responses, the response remains visible while it streams and disappears only after nopus requests a rewrite. Use these commands while working:

/nopus status
/nopus check
/nopus on
/nopus off
/nopus extra-simple on
/nopus extra-simple off
/nopus hide-original on
/nopus hide-original off

Add the marketplace from GitHub and install the plugin:

/plugin marketplace add Vistyy/nopus
/plugin install nopus@nopus

The plugin uses a Stop hook to request one clearer response. The identifier follows the plugin@marketplace format, so this plugin is nopus@nopus.

Add the marketplace from GitHub and install the plugin:

codex plugin marketplace add Vistyy/nopus
codex plugin add nopus@nopus

Start a new Codex session after installation so the plugin and its skills load. The Codex plugin uses the same bounded Stop hook. The identifier follows the plugin@marketplace format, so this plugin is nopus@nopus.

How it works

nopus examines the prose when a coding agent completes a response. Its measurements are deterministic, so the same prose and sensitivity produce the same decision. It measures these characteristics:

• Uncommon wording - Words that are rare in ordinary conversation.
  • Example: “Commence the task” instead of “Start the task.”
• Very uncommon wording - Words that are rare in both conversation and broad written English.
  • Example: “Ameliorate the problem” instead of “Fix the problem.”
• Abstract vocabulary - Concepts and qualities used without concrete people, objects, or actions.
  • Example: “capability, strategy, governance, and ownership.”
• Abstract sentences - Abstract wording that fills one sentence and continues across several sentences.
  • Example: “The strategy establishes capability ownership through governance transformation.”
• Noun and modifier stacks - Three or more nouns or modifiers compressed into one phrase.
  • Example: “repository state mutation verification strategy.”
• Phrase load - Dense phrases that appear frequently for the amount of prose.
  • Example: A short answer containing both “deployment policy validation process” and “repository state mutation strategy.”
• Formulaic style cues - Literal phrases that add inflated framing or filler.
  • Example: “Here’s where it gets interesting.”

Word rarity comes from packaged conversational and broad-web frequency data. Abstractness comes from human-rated word concreteness, and technical exceptions come from a packaged computing glossary. The style matcher uses only a small allowlist of literal cues rather than trying to classify arbitrary writing habits.

nopus then looks for combinations that make a response materially harder to read. For example, uncommon words become more significant when they appear with sustained abstraction and dense phrase stacks. Several abstract sentences become more significant when loaded phrases recur throughout them.

How nopus avoids unnecessary rewrites

• It removes code blocks, inline code, URLs, file paths, and table rows before measuring prose.
• It excludes obvious identifiers and reduces the weight of established computing terms.
• It uses proportions and phrase density, so a long response is not penalized merely for being long.
• Most decision paths require several measurements to cross their thresholds together.
• A single rare word or dense phrase does not normally trigger a rewrite.
• Style cues require either two distinct cues or one cue supported by enough uncommon wording.
• It requests at most one automatic rewrite, so it cannot create a retry loop.

Necessary technical terms, commands, identifiers, conditions, and qualifications remain part of the requested answer. InteractiveSessionHost stays InteractiveSessionHost; nopus targets the difficult prose around it.

When nopus requests a rewrite, the host displays this message:

nopus requested a clearer rewrite.

What the agent receives

For the response used in the example below, the default rewrite request gives the agent focused guidance and examples from its own response:

Rewrite the response for clarity and directness.
Keep its meaning and necessary detail.
Return only the revised response.

Focus on these areas:
- Replace formulaic framing with plain statements.

Examples from the response:
- "here's where it gets interesting": lecture-hall framing.
- "paradigm shift": tired.

Turning rewrite evidence off removes the Examples from the response block. The agent still receives the rewrite instructions and focus areas.

Extra-simple rewriting is off by default. When enabled, nopus requests a short initial answer that keeps the main conclusion, immediate action, and conditions or warnings that could change that action. It leaves supporting rationale, examples, alternatives, future considerations, and lower-priority implementation details for follow-up.

Choose how sensitive it should be

Sensitivity controls how often nopus intervenes, and the default is medium.

The observed rates come from 5,337 completed Pi responses and provide only a rough comparison because results vary by agent and task.

Configure automatic rewrites

nopus-configure changes the settings used by the automatic integrations. It does not rewrite a response.

The configuration skill writes $XDG_CONFIG_HOME/nopus/config.json, or the platform user-configuration equivalent.

{
  "complexitySensitivity": "medium",
  "includeEvidence": true,
  "extraSimple": false,
  "pi": {
    "hideOriginalResponse": true
  }
}

Rewrite evidence and Pi response hiding are enabled by default. Extra-simple rewriting is disabled by default. NOPUS_COMPLEXITY_SENSITIVITY, NOPUS_INCLUDE_EVIDENCE, NOPUS_EXTRA_SIMPLE, and NOPUS_PI_HIDE_ORIGINAL_RESPONSE override the configuration file for automation.

Simplify a response manually

nopus-simplify is the manual counterpart to the automatic rewrite. Use it when nopus does not trigger but you still find the preceding response too complex. Invoking the skill requests a rewrite immediately instead of applying the selected sensitivity threshold. It uses the same clarity goals as the automatic rewrite and preserves meaning, necessary detail, technical terms, commands, identifiers, conditions, and qualifications. It does not change nopus configuration.

Development

Install dependencies and run all checks:

pnpm install
pnpm check

Update the normalized datasets from their checksum-pinned upstream sources:

pnpm import:data

See evaluation/README.md for the private-corpus calibration and regression workflow. Normal analysis and verification do not download runtime data.

nopus retains normalized JSON rather than the original upstream formats. The SHA-256 values below ide