@pi-harness/plugin-reviewer-bot

Reviewer Bot — Review tracked Git changes against HEAD for whitespace, likely credentials, and TODO/FIXME markers without executing diff helpers.

Install

npm install --save-exact @pi-harness/plugin-reviewer-bot

Enable

Add the entry to the Cordis profile the harness starts from:

- id: reviewer-bot
  name: "@pi-harness/plugin-reviewer-bot"
  config: {}

The Pi Harness plugin marketplace installs and enables this package for you; the steps above are the manual equivalent.

Scope and limits

review_changes reads staged and unstaged tracked changes against HEAD. It requires an existing commit and excludes untracked files. This is a heuristic whitespace, credential-pattern and TODO/FIXME check, not semantic code review, test execution or a large-file detector. A pass means no configured check matched. Binary content is not inspected. Repository textconv and external diff commands are disabled. Git path prefixes and color are fixed for parsing.

The default diff buffer is 1 MiB (16 KiB–8 MiB configurable), and each Git command has a 15-second timeout (100 ms–60 seconds configurable). Exceeding the buffer fails rather than accepting a partial diff. Reports retain at most 512 files and 100 findings, with full counts and explicit truncation flags. Whitespace reports omit raw source lines. Snapshots are detached, and cancellation or disposal terminates outstanding Git commands and prevents publication; failed calls within the same workspace preserve the previous successful report. Separate Git reads are not an atomic working-tree snapshot: avoid editing files during a review.

Reviews use the current native session workspace and reject results after session replacement. Switching clears the previous report. Before the runtime exists, the launch workspace is used. Inherited Git repository and index redirection variables are removed, optional Git locks are disabled, and filesystem-monitor hooks are disabled.

Timeout remains failure even when Git handles termination by exiting zero; neither diff collection nor whitespace checking can publish a new passing report from that incomplete run. Git commands use the shared bounded runner: stdin EOF, POSIX process-group termination with one-second escalation, and bounded failure settlement even when descendants retain output pipes. Deliberately escaped groups are outside containment. Native Windows tree cleanup remains unverified.

Diff and filename collection run concurrently. If either fails, the other collection command is cancelled as well; the initial failure remains the reported reason. This cancellation is private to that collection and does not cancel subsequent reviews.

The model receives concrete findings and file summaries as JSON in text content, capped at 32 KiB UTF-8 (up to 20 files and 50 findings), alongside full counts and preview truncation flags. Detailed reports retain up to 512 files and 100 findings. Error findings displace retained warnings when capacity is reached and appear first in model previews. Neither format contains raw credential lines.

The panel shows the inspected workspace and finding paths, prioritizing errors among its four visible findings.

The panel distinguishes idle/running/completed/failed/cancelled attempts. A failed same-workspace attempt retains the last successful report but labels it historical rather than showing a current green pass; its bounded diagnostic is displayed separately. While a new attempt runs, the retained report is also historical. Success clears the error, and switching native session/workspace clears both report and attempt state. A diff-output overflow explicitly reports an incomplete review rather than a missing or unreadable repository.