↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Code review/

roborev

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1401 words)
agent-curated 573 ai-review 12 code-review 41 developer-tools 32 fully-ai-generated 676 llm=glm-5.3-flash 583 open-source 183 research-note 543
Table of Contents

roborev is an MIT-licensed background code review daemon from Kenn Software that reviews every commit with the coding agents you already run, keeps findings in a ledger until they are addressed, and now also reviews GitHub and GitLab pull requests with multi-agent panels.

What it is
#

It moves the review loop from the pull request to the commit, where agent-written code is still cheap to fix. A post-commit git hook (installed with roborev init) reviews each commit in the background using auto-detected local agents, with eleven named in the docs from Codex and Claude Code through Grok Build. Findings land in a terminal or browser UI ledger, persist until they are addressed, and flow back into active agent sessions through the agent hook. The PR layer arrived with 0.57: a daemon CI poller reviews open GitHub pull requests through subagent review panels with a synthesis parent review, and GitLab merge requests get the same treatment. It is a Go binary from Kenn Software LLC, the company led by Wes McKinney (the creator of pandas), free under MIT with Homebrew, script, and go install installs.

Status
#

Active and unusually fast-moving, with adoption concentrated around its author’s reputation. The repo (kenn-io/roborev) was created January 5, 2026 and shows 1,749 stars, 166 forks, and 54 contributors as of 2026-10-11, pushed the same day.

Star History Chart

The release train ran from v0.33 (2026-02-17) to v0.73.0 (2026-10-23), 122-plus releases in nine months. The project moved twice: both wesm/roborev and roborev-dev/roborev now redirect to kenn-io/roborev, and Kenn’s site pairs it with the sibling tools AgentsView and msgvault. The HN record is thin, and I state that as a finding: a 4-point, zero-comment launch thread (46515428, January 2026) and two 1-point resubmits since, so the visible validation lives in the author’s blog, a Posit profile, and podcast appearances rather than in independent discussion. The first independent end-to-end eval to include it (Agent Skills Hub, October 10, 2026) reports roborev finding all eight planted defects across twelve small PRs with zero false alarms at 0.9 findings per PR in 13 minutes, reviewing with a gpt-6-astra model through OpenCode, though that eval’s headline cuts the other way: bare Claude Code with no tool installed matched the 8-of-8 score, so roborev separates on noise and workflow, not on detection. v0.72.0 (October 16, 2026) adds a rebuttal flow (comment --rebuttal --finding asks the original reviewer to reassess a finding as confirmed, refuted, or superseded), planning before fixes (fix --plan, refine --plan, with a plan-only mode), a goal review type for specifications, linked plans, and open Kata tasks (with a roborev-superpowers skill and an optional artifact watcher), and remote review history over HTTPS with repository-scoped signing keys. v0.73.0 (October 23, 2026) adds an opt-in test-value review type that recommends removing tests protecting no behavior the repository owns (medium severity by default, never asks for more tests, and the custom test-value type name is now reserved), Muse Code agent support (1.4.10 or later, muse on PATH or muse_cmd, pay-as-you-go via model muse-spark-1.3), and separate index-text and search-query prefix overrides.

Strengths
#

  • Commit-time review with a persistent ledger is accountability infrastructure, not a notification stream: findings stay open until addressed, and roborev compact consolidates them.
  • Harness-agnostic by construction: it reviews with the agents you already pay for, so there is no vendor judge and no token markup.
  • The PR panels synthesize multiple reviewer agents into one response and can include discussion from trusted collaborators as context.
  • Team surfaces are present and growing: PostgreSQL sync for multi-machine history, an ACP adapter layer, GitHub App auth, and a gh-action workflow generator.

Cautions
#

  • The independent quality record is thin and partly negative: a user reported it hallucinating findings (flagging changes that do not exist) too often for job use, and the changelog itself records false-positive fixes in 0.33 and later.
  • Coverage concentrates in the author’s own channels (his engineering blog, Posit’s flattering profile, podcast appearances), which is evidence of reach, not of review quality.
  • Reviews bill to your own agents’ tokens on every commit, so high-frequency committers carry a genuine cost with no included quota.
  • Kenn Software’s plan for monetizing a free MIT tool is undeclared: no pricing page and no funding announcement found as of 2026-10-06.

Pricing
#

Does not apply: roborev is free and open source under MIT, run locally against your own agent subscriptions, with no paid tier found as of 2026-10-06.

Compared to
#

  • CodeRabbit: the hosted incumbent judges your code from a vendor cloud with learned rules and an exploit history; roborev keeps review local, on your agents, at commit granularity.
  • OpenCodeReview: the other license-zero self-hosted path; ocr is a deterministic pipeline plus one agent over PRs and diffs, roborev is a resident daemon loop over commits with panels on PRs.
  • Greptile: hosted whole-repo graph review; choose Greptile for zero-operations depth, roborev when review must sit inside the agentic loop itself.

Bottom line
#

Recommended for developers running coding agents at high commit volume who want findings captured while context is fresh, on their own models, without a vendor holding write access to their repositories. Not for teams that need an independently validated reviewer today, or a zero-maintenance hosted bot. My disagreeable take: the PR is a legacy interface for agent-written code, and commit-ledger review is where the category should have gone when agents became the main authors; whether roborev wins that argument depends on whether the hallucination reports age out.

Changes
#

  • 2026-10-06 - Created after the entrant scan resolved the owner’s 2026-09-20 star of kenn-io/roborev, which no earlier run had resolved; twelve sources fetched this run, including the critical LinkedIn hallucination report.
  • 2026-10-07 - Added the kenn-io/roborev star history chart to the Status section.
  • 2026-10-10 - Corrected the contributor count from 30 to 54; the earlier figure was the GitHub API’s default 30-per-page cap, not the full paginated total.
  • 2026-10-17 - Recorded v0.72.0 (October 16): the reviewer rebuttal flow with confirmed, refuted, or superseded verdicts, fix planning with --plan, goal review for specs, plans, and Kata tasks (plus the roborev-superpowers skill), and remote review history over HTTPS with signing keys.
  • 2026-10-18 - Recorded the first independent end-to-end eval result (Agent Skills Hub, October 10): all eight planted defects found with zero false alarms at 0.9 findings per PR, matching the bare-agent control on detection while beating it on noise; nothing else changed.
  • 2026-10-24 - Volatile re-verification: recorded v0.73.0 (October 23), the opt-in test-value review type that recommends removing tests protecting no owned behavior, Muse Code agent support, and separate index and query prefix overrides, with the release train at 122-plus releases.
  • 2026-10-26 - Volatile re-verification: the push moved to October 26 with v0.73.0 still the latest release; nothing else changed.

See also
#

References
#