Four ways to review a change. Pick by how much you can leave behind.
Lite, snap, fast and deep differ in speed, in price and in how many bugs they leave in the code. Below: what each is for, its plus and its minus, how each did on our benchmark, and the order we recommend.
The four modes
-
lite $1
For saving: a little more than snap finds, at half the price, and slower.
Plus. Half the price of snap, and it finds a little more.
Minus. Fast and deep find more of the known bugs. A typical lite review takes 85 s: more than twice as long as snap, quicker than fast.
-
snap $2
For the loop: while your agent is still writing the change.
Plus. The answer comes back in under a minute.
Minus. One pass leaves more of the known bugs than fast or deep.
-
fast $4
For the check before the merge. Without a mode it is the review you get from the API and older clients; on a pushed GitHub commit, the plugin and the CLI may run a first step and follow it with fast when that step finds no medium or worse bug (how).
Plus. Fewer known bugs left than snap or lite, in a few minutes.
Minus. It costs more per pass than lite or snap.
-
deep $12
For risky or wide changes: payments, security, data, many folders at once.
Plus. The fewest known bugs left of the four.
Minus. The slowest, and it returns many findings for your agent to check. $24 on a very wide change.
How each mode did on the benchmark
All four modes reviewed the same 85 held-out changes of OMBench, our benchmark, which hold 742 known bugs. Found and left add up to 100% of those known bugs; neither is a share of all bugs in the code, which nobody knows. Deep comes from the four-mode run; lite, snap and fast changed after it, so their rows come from later runs.
| Mode | Price | Found, of the known bugs | Left | Hard cases: found / left | Findings per change: medium+ / all | Server time: typical / 9 of 10 |
|---|---|---|---|---|---|---|
| lite | $1 | 61.5% | 38.5% | 52.5% / 47.5% | 30 / 49 | 85 s / 2.1 min |
| snap | $2 | 57.7% | 42.3% | 47.5% / 52.5% | 27 / 45 | 34 s / 49 s |
| fast | $4 | 76.5% | 23.5% | 77.5% / 22.5% | 35 / 61 | 1.9 min / 5.4 min |
| deep | $12 / $24 | 86.1% | 13.9% | 77.5% / 22.5% | 40 / 80 | 4.6 min / 10.0 min |
Each mode's figures over time, as its setup changes: OMBench results.
How to read it
- Known bugs. Bugs checked against the code before the run, in changes never used to tune any mode. A bug counts as found only when a finding points to the same defect at the same place.
- Hard cases. 40 bugs that earlier reviews missed, in 23 of the changes. A small sample.
- Findings per change. Your agent reads them all, checks each against the code and drops the false ones; you read a short list. The medium-or-worse count is what a fix starts from.
- Server time. From the moment the review starts on our side to the result: the typical change (the median) and 9 changes in 10.
- Price. The price of one review in each mode; deep costs $24 on a very wide change. When a review is charged is in the FAQ · CI billing.
The order we recommend
No single mode fits every moment of a change. You follow this order by naming the mode each time; a named mode never switches on its own (without a mode, see below).
- While the agent works, run snap after each larger step. It is quick enough to sit inside the agent's loop, and it catches more than half of the known bugs early, when they are cheapest to fix.
- Before the merge, run fast on the finished change. Snap's speed comes at a price: it leaves about two in five of the known bugs, and fast leaves about a quarter.
- When a change is risky or wide, run deep as well. It is slower and gives your agent more to check, and it leaves the fewest known bugs of the four.
- When the price matters most, run lite: half the price of snap and it finds a little more, with a typical time of 85 s, between snap and fast. Fast and deep find more.
How to pick a mode
Claude Code or Codex plugin
Name it in the review command:
/bughunter:review snap,lite,fastordeep. Without a name, see without a mode.API and MCP
Send
modeaslite,snaporfastwithsubmit_review, ordeep: truefor deep.get_balancelists the modes your key can use now.CLI and CI
Pass it as a flag:
ohmybug --mode snap,liteorfast(CLI 0.8.0 or later; orOHMYBUG_MODEwhen the flag is absent), orohmybug --deep. Without a mode, see below; a named mode, fast included, runs as asked, with no automatic continuation. See the CLI guide.
Lite and snap read the change from GitHub, so they need a pushed commit in a repository OhMyBug can read; snap takes changes under 1 MB.
Without a mode: a first step, then fast
When you name no mode, the service may start with a first step: snap, or lite on accounts set to lite. What it finds decides what comes next.
| The first step finds | What happens | What you pay |
|---|---|---|
| No medium or worse bug, or the step fails | A fast review of the same commit follows, and it is the answer. | $0 for the first step. The fast review bills as any fast review does: $4. |
| A medium or worse bug, on CI billing | The first step is the answer. No fast review follows. | The first step’s price, charged when it ends: $2 snap, $1 lite. |
| A medium or worse bug, on confirm billing | The service waits for your agent’s verdicts on those findings; they decide whether a fast review follows or the first step is the answer. | The first step bills as any review of its mode does. The fast review, if it follows, bills as above. |
If the fast review cannot be started, or the service refuses it (for example, the key now runs only deep hunts), no fast review follows, and your client says why.
Only a review with no mode named, from a client that follows the route, on a pushed commit and base branch the service reads from GitHub, can be routed. The clients that follow it: the plugin 0.157.0 or later in Claude Code, the plugin 0.158.0 or later in Codex, and the CLI 0.8.0 or later. Older clients, and API callers that do not send the route header, run fast. Uncommitted changes and a diff the client sends run fast too, and so does a change the first step cannot take, such as one over snap’s 1 MB. A named mode, fast included, and a deep hunt are never routed.
More
- OMBench: how the benchmark is built, verified and scored.
- The page for coding agents: how a coding agent should use the findings.
- FAQ: prices and when a review is charged.