Review agents
Review agents
Code review, Security review, Slop review and Intent drift run on your changes, post their findings and a check, and are billed from your credit.
The agents#
YonedaRepo maintains four review agents. Each reviews an open change's diff, read only. The first three report on the lines the change added; Intent drift gives one verdict on the whole change.
| Agent | Looks for | Runs on | Result |
|---|---|---|---|
| Code review | Logic and off-by-one errors, error handling, races, missing awaits, unintended behavior changes, and breaks of the repository's own REVIEW.md, AGENTS.md or CLAUDE.md rules. Skips formatting and anything a linter reports. | Codex, gpt-5.6-luna | Pass or fail |
| Security review | Injection, missing authorization checks, committed or logged secrets, SSRF, XSS, unsafe deserialization, weak cryptography, privilege escalation | Claude Code, claude-sonnet-5-5 | Pass or fail |
| Slop review | Dead or duplicated code, debug leftovers, drift from the repository's conventions, oversized files, filler comments | opencode, mimo-v2.6-flash | A level: low, medium or high (lower is better) |
| Intent drift | Whether the diff does what the change's flight plan set out to do, and nothing it ruled out. A change with no plan starts no model and costs nothing. | opencode, mimo-v2.6-flash | A level: low on course, medium partly, high off course |
The agents never edit, commit, push or use the network, and cannot land, approve or change policy: their token holds only read, review and check, including the right to post their own checks when your landing rules require them. The repository's REVIEW.md, AGENTS.md and CLAUDE.md, as they are at the change's base, are part of each review.
Turn them on#
- Open Agents in the app's sidebar (the page Managed by YonedaRepo).
- On an agent's card choose Enable.
- Pick the repositories under Repositories, adjust the settings if you like, and choose Enable agent.
Only an owner can turn an agent on for a repository (for an organization's repositories, an organization owner). Turning one on adds a repository token named yoneda-managed-agents and a webhook to the repository; turning off the last agent removes both and cancels open runs. Edit and Disable on the card change it later. A repository's Settings → Plugins links to the same page.
Settings#
| Setting | Meaning |
|---|---|
| Harness, Model, Reasoning | Which agent CLI and model run the review (defaults above) |
| Timeout (minutes) | 5 to 30; default 20 (Slop review 15) |
| Instructions | What the agent looks for; Reset to the maintained instructions restores ours |
| When to run | On every push (default) or When marked ready (once, when the change opens) |
| Paths, Branches | Optional filters; branches take prefix/* |
| Block landing at | Slop review and Intent drift: High (default), Medium or higher or Never (report only) |
When they run#
An agent runs when a change is opened and, with On every push, each time its branch moves. Each run reviews one head commit. A newer push replaces a run still going on an older head. There is no manual run button.
What they post#
On the change:
- Line comments, one per finding, at most 20, each on a line the change added.
- A summary comment.
- A check named after the agent (
code-review,security-review,slop-review) on the head commit. Code review and Security review fail it when they post a high or medium finding. Slop review reports low and medium findings; its level is that of its most serious finding, or high with five or more medium findings on lines of the change. It fails the check only at or above Block landing at.
Intent drift posts no line comments. Its summary comment gives the verdict and what drifted, and its check intent-drift carries the level: on course is low, partly on course medium, off course high. So by default an off-course change fails the check. With no flight plan the check passes and its evidence says no_plan.
Findings that cannot be placed on a line of the diff, or that look like a credential, are left out. So are findings past the 2,000 comments a change can hold; they still count toward the check.
A failed check blocks landing when the repository's landing rules list it in required_checks. With Land when cleared on, any failed check sends the Landing around, required or not. Otherwise it is advice. Line comments don't hold a Landing unless Block landing on unresolved agent findings is on.
A run that can't finish leaves no reading. If the check is required it fails, and a comment says so: push again to run the agent. If it isn't, nothing is posted. A pause fails nothing: when the owner resumes automation, agents whose checks are required run on the open Landings still waiting for them, up to 20 runs in a repository. The rest run on their next push.
Go-around#
Nothing wakes an agent when findings arrive. The next session on the branch reads them in its context, with file and line, and so does anyone running yoneda change show <n>. It fixes them and pushes, and the push runs the review again. See Go-around.
Runs#
Agents → View runs lists every run; a run's page shows its findings, its redacted transcript and its cost.
| Status | Meaning |
|---|---|
| Queued, Running | Waiting for or using a container |
| Passed | No blocking findings |
| Found issues | The check failed |
| Paused (no credit) | The credit ran out; see Credit and usage |
| Replaced by a newer push | A newer head is being reviewed instead |
| Stopped, Failed | Cancelled, or the run could not finish |
Request one review through the API#
Workspace owners can POST /api/managed/workspaces/<workspace>/agents/<agent>/run on the website origin, using their signed-in session. Send {} for the usual workspace-wide Run now, or select one current change head:
{ "repo": "<repository id>", "change": 12, "head": "<40-character commit id>" }The agent must be enabled on that repository. An exact request skips automatic trigger filters, but still respects drafts, automation pauses and credit limits. A stale head is refused. Repeating the same request reuses its run; it does not purchase another attempt. The workspace response lists exact_head_review in capabilities when this selection is supported.
Pause or turn off#
- Disable on an agent's card turns it off everywhere and cancels its open runs. To stop it on one repository, Edit it and untick the repository.
- Pausing a repository's automation (owner only; see Break glass and automation pause) stops new runs on that repository, and a run that has not started reviewing stops.
- When the credit runs out, runs pause on their own; see Credit and usage.
Where the review runs#
Each run uses a fresh isolated container. Its model calls go through our model gateway to the model's provider, which receives the code and text the agent reads; see Global network and data location.