Find and implement the next change worth testing
Start from your own brief or let Outrider find a candidate. It checks the change against your repo, implements it at a real call site, and validates the result before review.
You decide what goes any further.
$ remyxai outrider init --repo owner/name --auto-interest
Two Outrider-drafted contributions have merged into Hugging Face PEFT.
// SHIPPED_UPSTREAM
Add Riemannian-preconditioned LoRA optimizer
"this PR is already quite mature"
huggingface/peft maintainer
Add Super-Tuning as a new PEFT method
"The PR looks good from my perspective"
huggingface/peft maintainer · final approval
Start with research or your own brief
The source can change. The implementation and validation path stays the same.
Scheduled discovery against your Research Interest.
# no command needed
Give Outrider a method family. It finds the top matching paper.
--search-method \ "riemannian preconditioning LoRA"
Name an exact arXiv id for a repeatable run.
--pin-arxiv 2410.20305v2
Give Outrider an issue, design doc, Colab, or inline spec. No paper required.
--brief @design.md
Discover → implement → validate → decide
Start with the change
Use alerts, search, a pinned paper, or your own brief. Outrider checks the change against your repo, its history, and a real call site before it writes code.
→Build it on a branch
The coding agent runs in a fresh runner, implements the change in your repo, writes tests, and keeps the work scoped to the integration.
→Check the draft
Outrider checks fidelity, repo conventions, lint, and targeted tests before the work is ready for review.
→Review the result
Outrider opens a PR, files an Issue, keeps the work on a branch, or skips. Your team decides what goes any further.
Every run has an explicit outcome
Outrider would rather stop than open noise. A run can produce a PR, an Issue, a branch for your team to inspect, or a recorded skip.
The change fits the repo and clears the checks. Outrider opens a PR with the implementation, tests, references, license context, and review notes.
The idea is relevant but needs a better integration point or maintainer judgment. Outrider files an Issue and preserves the attempted work for discussion.
Push the implementation without opening a PR or Issue. Use branch mode to inspect several candidates before choosing one to put in front of maintainers.
Nothing clears the fit, dedup, or safety gates. No maintainer-facing artifact is created, and the run records why it stopped.
Use evidence from earlier runs
With repo state enabled, Outrider records where prior changes fit, which shapes were rejected, and why. Later runs can use that history alongside merged PRs and existing Issues.
Call sites and integration shapes that prior runs reached cleanly.
Rejected mechanism shapes, the reason they failed, and the conditions that could change that result.
Merged PRs, coordination Issues, and prior Outrider artifacts help prevent repeated work.
As new findings arrive, Outrider can evaluate them against a repo history that is getting more specific.
Before review, every PR is checked three ways
The default chain runs after the draft is created and stays in the same workflow.
Compares the draft to the reference implementation when one exists, or to the paper when it does not. The run records what is covered, deferred, or different.
Shapes the PR around the repo's recent merged work, contribution docs, and existing templates.
Runs lint and targeted tests on the changed files. Passing work moves to ready-for-review; failures stay visible for a human to take over.
Use the model you want
Anthropic is the default. Route a run through z.ai GLM, Moonshot Kimi, Bedrock, Vertex, or your own Anthropic-compatible endpoint without changing the Outrider harness.
# the same spec, validators, and refinement chain run downstream.
Switch it on in minutes
Hand Outrider one change now, or set up scheduled discovery. Either way the run lands a PR, Issue, branch, or recorded skip.
# your own brief $ remyxai outrider trigger --repo owner/name --brief @design.md # method-family search $ remyxai outrider trigger --repo owner/name --search-method "riemannian preconditioning LoRA" # exact paper $ remyxai outrider trigger --repo owner/name --pin-arxiv 2410.20305v2
$ pip install remyxai
$ export REMYXAI_API_KEY=...
$ remyxai outrider init \
--repo owner/name \
--auto-interest
Installs Outrider on the repo and creates a Research Interest from it in one step.
# no terminal required create a Research Interest from your repo, then choose "Set it up for me"
Connect GitHub once, create the interest from your repo, and let Remyx provision the workflow.
$ remyxai outrider setup-local \
--repo owner/name \
--auto-interest \
--anthropic-key sk-ant-...
Provisions with your own authenticated gh CLI. PRs then come from your account instead of remyx-ai[bot].
Runs in GitHub Actions. Nothing merges on its own
The coding agent runs in a fresh runner with a stripped-down environment. GitHub and Remyx credentials stay out of the agent subprocess, outbound PR and Issue content is checked for secrets, workflow edits are blocked, and every merge still requires a human.
- same-repo installs use a repo-scoped bot token for GitHub writes
- model provider keys stay in your repo's Actions secrets
- .github/workflows/** is always blocked from agent edits
- dependency, CI, hook, container, and shell changes are flagged for review
- a credential match in outbound content aborts the run
- teams can add their own path blocklist
- nothing auto-merges
Your first Outrider run is one command away
Start from your own brief or let Outrider find a change. Free to install and runs in GitHub Actions.