AgentScan Text Action

Runs the interlinked detector on a pull request's description and classifies it the way agentscan classifies accounts: organic, mixed, automation, or insufficient-data when there's too little text to judge. Mixed and automation PRs get a label. In full and description modes the action also publishes the analysis, as a PR comment by default or appended to the description inside an <!-- interlinked:start --> … <!-- interlinked:end --> block. Later runs edit that comment or block in place, and the block is stripped before analysis so it never affects the score.

name: AgentScan Text

on:
  pull_request_target:
    types: [opened, edited, reopened]

permissions:
  contents: read
  pull-requests: write

jobs:
  analyze:
    runs-on: ubuntu-latest
    steps:
      - uses: MatteoGabriele/agentscan-text-action@v0

Use pull_request_target so the token can write to PRs from forks. The action never checks out or runs PR code, so that's safe here. Edits made with the default GITHUB_TOKEN don't trigger new workflow runs, so updating the description or comment doesn't loop.

Inputs

Input	Default	Description
github-token	${{ github.token }}	Reads the PR template, updates the description, manages labels.
allowed-users		Usernames to skip, comma-separated or as a JSON array.
trusted-author-associations	member,owner	Author associations to skip (collaborator, contributor, first_timer, first_time_contributor, member, owner), comma-separated or as a JSON array. Set to "" to analyze everyone.
mode	labels	full (description and labels), labels (labels only), description (description only) or silent (outputs and job summary only).
analysis-location	comment	Where full and description modes publish the analysis: comment (a separate PR comment, edited in place on reruns) or description (appended to the PR description). See the warning below.
comment-on-organic	false	Publish the analysis for organic and insufficient-data descriptions too. Either way, an earlier run's analysis is brought up to date, so an edit that clears a flag doesn't leave the old verdict behind.
label-mixed	agentscan:mixed-text	Label to add when the classification is mixed.
label-automation	agentscan:automated-text	Label to add when the classification is automation.
auto-close	false	Close the PR when its classification is in auto-close-classifications.
auto-close-classifications	automation	Classifications that trigger auto-close (mixed, automation), comma-separated or as a JSON array.
fail-on-classifications		Classifications that fail the step (mixed, automation), comma-separated or as a JSON array. Empty, so the step never fails.
message-organic / message-mixed / message-automation		Custom message per classification, shown in place of the default one.

Labels are never removed automatically, so editing the description can't clear them; a maintainer removes them.

These follow the same conventions as agentscan-action. The default labels share its agentscan: prefix but differ from its account labels, so both actions can run on the same PR.

Warning

With analysis-location: description, the analysis lives in text the PR author controls. They can edit or delete it, or rewrite it to say the opposite. The label and the default comment can only be changed by maintainers, so rely on those for anything you enforce.

Outputs

Output	Description
classification	organic, mixed, automation, or insufficient-data.
score	How human the description reads, 0 to 100 with 100 human, on the same scale as agentscan's account score. -1 with insufficient data. Below 50 is automation, 70 and up is organic.
confidence	Confidence in the classification, 0.5 to 1.
signals	The signals that fired, as JSON.

The analysis is also written to the job summary.