Outrider — GitHub Action
https://github.com/user-attachments/assets/4b22a207-d878-4b4d-a1f1-02e886a8e994
Keep the systems you maintain improving — discover, implement, and validate your next great idea, delivered as a review-ready PR.
Outrider runs as a GitHub Action. It finds the next high-fit change, wires it into a real call site in your repo, and returns a draft PR whose body already carries the evidence a maintainer needs — references cited, license flagged, tests written, scope kept honest, and the implementation anchored on your repo's own merged history and conventions. Discovery is the tool, drafting is the agent, the judgment is yours: every run produces a branch you review, run, and decide on. Run after run, the loop compounds — discover → implement → validate → decide — and your repo accumulates a record of what actually fits it.
- uses: remyxai/outrider@v1
with:
interest-id: ${{ vars.REMYX_INTEREST_ID }}
Each dispatch runs the coding agent in a fresh, ephemeral runner — candidates don't share state, testing variance stays low, and you can dispatch dozens per week without context pollution. Backends are pluggable: Anthropic Opus for the shipping commit, z.ai's GLM-5.2 at ~20× lower cost for scouting and branch-mode exploration.
Trigger patterns
Four ways to specify what Outrider should implement — same harness downstream:
| Trigger | Where the intent comes from | How to fire |
|---|
| Alerts | System-sourced. The ranker picks from arXiv against your ResearchInterest on a weekly cron. | Scheduled — no command needed |
| Search | Semi-sourced. You supply a method-family query; the ranker returns the top arXiv hit. | --search-method "riemannian preconditioning LoRA" |
| Pin | Reproducible. You name an exact arXiv id. | --pin-arxiv 2410.20305v2 |
| Brief | Operator-sourced. You write the design brief directly — an issue body, a Colab notebook, a hand-authored spec. No arXiv anchor required. | --brief @design.md or --brief "add exp backoff to the HTTP client" |
Every pattern produces the same output shape: a draft PR with implementation + tests + license section + convention-aligned body + honest scope citations. What differs is only where the "why" comes from.
Bring your own context — even when it's underspecified. The scaffolding fills in what the brief doesn't (references cited, license flagged, tests, convention alignment) so a two-line issue body still yields a review-ready PR.
What you get
- Draft PRs wired to an existing call site, with a self-review noting what was implemented vs. left out
- Issues when preflight, validators, or self-review route the intent to discussion instead
- Branch-only mode (
publish: branch) — pushes to the fork without opening a PR or Issue; explore N candidates before committing to any one
- No duplicate work — a paper isn't re-recommended once Outrider or a maintainer Issue references it
- A selection narrative in the step summary — why this candidate, or why nothing this run
Agent and model backends
Two independent axes. agent picks the coding-agent CLI that does the work; provider picks the model behind it.
agent | CLI |
|---|
claude (default) | Claude Code |
codex | OpenAI Codex |
backboard | Backboard R-CLI |
| Backend | Cost / full run | Best for |
|---|
| Anthropic Opus | ~$2–3 | Finalize a draft PR |
| z.ai GLM-5.3 | ~$0.05–0.10 | Draft PR |
| Moonshot Kimi-K3 | ~$1 | Finalize a draft PR |
Route per-dispatch via the agent and provider inputs — see docs/backends.md for the generated pair table, the auth-header matrix and the switching workflow template. Rule of thumb: GLM for the exploration ladder, Opus for the candidate you commit to ship.
Leave agent unset and nothing changes: every existing workflow keeps the Claude Code path it has today. Not every agent/provider pair is valid — an agent speaks one API family and a provider serves one or more — so the action rejects an impossible pair before spending anything and names the agent that does serve your provider. mode: smoke verifies a new agent/provider/model combination against the vendor in seconds without cloning or opening anything.
Quickstart
pip install remyxai
remyxai outrider init --repo owner/name --auto-interest
Installs the action, writes the workflow, sets the secrets (REMYX_API_KEY, ANTHROPIC_API_KEY). Scheduled cron handles the weekly cadence from there.
Add --github-secrets-only to keep your provider key in one place: it is encrypted on your machine, written to the repo's Actions secrets, and no copy is stored anywhere else.
Trigger an ad-hoc run:
# Paper-anchored — exact arXiv id or a method-family search
remyxai outrider trigger --repo owner/name --pin-arxiv 2410.20305v2
remyxai outrider trigger --repo owner/name --search-method "riemannian preconditioning LoRA optimizer"
# Brief-anchored — a design brief you supply directly, inline or from disk
remyxai outrider trigger --repo owner/name --brief "add exponential backoff to the HTTP client"
remyxai outrider trigger --repo owner/name --brief @design.md
--pin-arxiv implements the exact paper; --search-method searches for the top hit; --brief runs the paper-less flow where the design brief you supply IS the spec. See remyxai-cli for bulk-install and per-dispatch routing.
Setting up by hand instead of via the CLI? See docs/manual-install.md.
Examples
Case study: three recent contributions to huggingface/peft
Three parameter-efficient fine-tuning methods surfaced from arXiv, drafted on the smellslikeml/peft fork, and shepherded upstream to huggingface/peft:
Full case study — per-method deep dives, PR-shape cohort comparison, coordination-issue-first workflow — at docs/case-studies/peft.md.
Case study: two training-free high-resolution methods for huggingface/diffusers
Two complementary routes to 4K generation from off-the-shelf FLUX checkpoints — no fine-tuning, no new weights — surfaced from arXiv and shaped to diffusers' own conventions:
- HRDiT — huggingface/diffusers#14480 (in review). NTK RoPE scaling + structure guidance for up to 4096², landed as a self-contained community pipeline with no core-library changes. Built minimal-first: a position-alignment port, then NTK scaling and structure guidance layered only onto the stages that needed them.
- DyPE (+ optional spectral attention) — proposed via huggingface/diffusers#14520. A training-free positional-embedding hook (
apply_dype) that reproduces the reference bit-for-bit (Δ=0) and adds an optional spectral-attention mode which cuts residual 4K speckle ~6× while keeping fine detail. Offload-robust, a no-op at/below the trained resolution.
The pair puts the same problem — training-free high-res on a RoPE DiT — into two different diffusers shapes. Choosing which shape is itself the contribution: HRDiT self-contains as a community pipeline, while DyPE sits closer to the core hooks (apply_faster_cache, apply_pyramid_attention_broadcast). Because precedent points both ways for DyPE, that placement is raised with maintainers in #14520 rather than assumed.
Full case study — per-method detail, the placement question, validation figures — at docs/case-studies/diffusers.md.
More examples
Each PR below shows the match (paper → repo) and the shape (how the wiring landed):
- huggingface/pytorch-image-models #2767 — EfficientViM backbone family (arXiv:2411.15241), merged. Match: timm's sub-25M efficient-backbone lane. Shape: the HSM-SSD ("Mamba duality") core reduces to pure PyTorch (no
mamba_ssm / custom CUDA); resolved the family's long-standing Google-Drive-only weights gap (mlvlab/EfficientViM#1) by converting the 12 checkpoints to safetensors on the HF Hub.
- OLMo-core #13 — Preemptive training instability monitor (arXiv:2606.28116). Match:
train/callbacks/ has the reactive StabilityMonitorCallback; the preemptive variant registers alongside. Shape: MechanismMonitorCallback with QK spectral entropy + MoE routing entropy, gated by a parameter-free rolling one-sided z-score; 12 tests.
- OpenRLHF #14 — MRPO step-level reward penalty (arXiv:2606.31825v1). Match: PPO advantages already carry per-step weighting. Shape: MRPO's decay factor slots into
RemoteExperienceMaker.compute_advantages_and_returns as a second multiplier, opt-in flag, default-off byte-identical.
- ag2 #9 — Adaptive Context Elasticizer (arXiv:2606.31564v1). Match:
MiddlewareFactory already extends the LLM-call pipeline. Shape: new elastic middleware alongside HistoryLimiter / TokenLimiter, per-instance abstraction cache for reversibility.
- — Dense Embodied Chain-of-Thought supervision (). the annotator has staged language modules (plan / vqa). new wired in as phase 4.5.
Documentation
License
Apache 2.0. See LICENSE.
