OMP Best Of plugin
An Oh My Pi plugin: several agents solve a task in isolated copies, a verifier picks the best solution and applies the patch.
- 79 ★ on the repository
- Last commit
Extended access
Why this level
- Several parallel agents edit repository copies and spend tokens
- The `--apply` option writes the winning patch into the main working tree
Install
Manual install
omp plugin marketplace add wolfiesch/omp-best-ofThe first of two Oh My Pi commands.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Several parallel agents edit repository copies and spend tokens
- The `--apply` option writes the winning patch into the main working tree
This is third-party code. Review the repository files before installing.
What it does
The command `/best-of --n 5 --apply Fix the failing authentication test` launches 2 to 8 Oh My Pi candidates on one task, each in its own isolated copy of the repository (copy-on-write, a git worktree when needed). A verifier then ranks the full trajectories: `logprob` on the LLM-as-a-Verifier package using the Kwok et al. tournament algorithm, or `sampled` with pairwise comparison on an OMP subscription model. The winning patch reaches the project only with `--apply` and only if `HEAD` has not changed. The plugin refuses a dirty tree, drops candidates with a non-zero exit code and always keeps artifacts. A standalone `omp-best-of` command serves scripts and CI.
Who it is for. Oh My Pi users who want several attempts at a hard change and the best one picked instead of trusting the first.
Good fit when
- A failing authentication test is fixed by five candidates in parallel.
- A parser refactor is to be compared across variants without touching the working tree.
- Trajectories, patches and token usage of every attempt must be kept for review.
- Best-solution selection is wired into a script or CI through `omp-best-of`.
Not a fit when
- The working tree has uncommitted edits: the plugin will refuse to start.
- A different agent than Oh My Pi is used: the plugin is written for its extension API.
- The token budget is tight: running five candidates and a verifier costs far more than one attempt.
Example request
Run `/best-of --n 3` and rewrite the parser so its public API does not change.Limitations
Oh My Pi 17 or newer, Bun 1.3 or newer and git are needed, and the `logprob` mode also needs `uv` and a route to a model that returns log probability scores. The working tree must be clean. Candidates inherit the model and thinking level of the current session, so costs scale with the number of candidates. The `sampled` mode is described as an ordinary pairwise judge, not the algorithm from the paper. The default verifier model is set by `--verifier-model`. The license is MIT and the plugin ships with a task set for measurements.
How to disable. Run `omp plugin uninstall best-of` or unlink the local copy and restart Oh My Pi. Delete saved run artifacts by hand.
FAQ
How do I choose the candidate count?
With `--n` from 2 to 8, default 3. More candidates raise the chance of a correct one and the token cost.
Does the plugin change my code by itself?
No. Without `--apply` it only ranks solutions and keeps artifacts. The patch applies only if `HEAD` and tree status are unchanged.
How do `logprob` and `sampled` differ?
`logprob` uses the LLM-as-a-Verifier tournament and a model with scores, `sampled` runs pairwise comparison on an OMP subscription model.
What about time limits?
The `--max-time` flag limits each candidate, 20 minutes by default.
Can it run outside an Oh My Pi session?
Yes, with `omp-best-of --n 3 --apply -- "task"`. Progress goes to stderr and the JSON summary to stdout.
Discussion
No questions or reviews yet.
Related
Pluginspstack skill catalog ported to OMP: a task router, TDD, code review and subagent playbooks, each invoked by its own command.
An Antigravity CLI plugin for engineering work: a separate design turn, documentation agents and edits only within a sealed plan
OpenCode plugin that turns a session into a team of agents led by an architect and runs the work through plan phases
A coding agent for small, cheap LLMs: splits a goal into a task tree with contracts and verifies each leaf with builds and tests.
A Pi agent package: code search, LSP, web search, desktop control and confirmations for risky commands.
A Windows plugin for OpenCode Desktop that sets edit boundaries, screens commands and ties test results to the current code.
README in short
The README explains the idea of picking among several solutions, the step-by-step flow, requirements, installation, example commands and a flag table.