Give your agents discipline.
Evidence for every PR.
Coding agents write code fast, but without guardrails they regress, wave through their own reviews, and loop. ActiveLoom is an open-source set of skills and review protocols for Claude Code, Codex, and Gemini: independent models review each draft PR, findings are posted before fixes, and the evidence is tied to the exact commit. Try one skill, run whole review chains in panes you can watch, or sync the set across every repository.
Installs one skill for you (Tier 0). Nothing is written to a repository. More ways to adopt ↓
Engineered for teams that treat agent code like production code.
ActiveLoom replaces fragile ad-hoc system prompts with structured engineering discipline, rigorous review ledgers, and deterministic fleet sync.
- 1 · Skills
The base. Install any one on its own: interviews, debugging, issue work, review. - 2 · Review protocol
Review skills post findings on a draft PR before fixing, and record them in a ledger pinned to the commit. - 3 · Automation
agent-loopworks an allowlisted issue queue headlessly. The optional herd skills run the same reviews, interviews, and builds in terminal panes you watch. - 4 · Sync
Keeps all of it current across your repositories through pull requests, without touching your local rules.
Not chat prompts. Operational skills.
ActiveLoom skills encode strict operational procedures with validation checklists and deterministic outputs. From product and design interviews (product-grill, grill) and deep troubleshooting (diagnosing-bugs) to agent-ready backlogs (backlog-refinement) and pre-merge red teams (red-team).
Claude, Codex, and Gemini in sync.
Don't maintain three divergent prompt trees. Shared skills are authored once in prompts/ and rendered into .claude/, .codex/, and .agents/ (Gemini / Antigravity). Review prompts are tuned per engine on purpose.
Adversarial review. Proof before edit.
No rubber-stamping. Reviewers inspect the draft PR diff and post verified findings inline on GitHub before touching any code. Each PR's review depth is resolved first: Lean runs the highest-signal lenses; Deep runs the full matrix for risky surfaces such as auth or crypto.
Review evidence pinned to the commit.
Every review pass is recorded on the draft PR as ledger markers carrying sha256 digests and the exact head commit. Push a new commit and attestations for the old head stop counting.
Autonomous queues with guardrails.
Let agents work your GitHub backlog safely with agent-loop. Strict issue allowlists, one linked Git worktree and draft PR per issue (agent-loop/issue-<N>-<run>), bounded local review rounds, fresh-base validation, and automatic bail-outs when blocked. Setup guide ↗
Sync every repo without losing local rules.
A scheduled GitHub Actions workflow proposes skill and review updates as pull requests from the upstream's sync-v2 release ref, while protecting repo-owned rules such as .review/addendum.local.md. How sync works ↗
One session drives the herd. You watch every pane.
The herd skills let one agent session drive other agent sessions in visible herdr panes (herdr is a terminal workspace manager for coding agents). Every review pass, interview, and build runs where you can see it, and you can type into any pane to step in. Any of the three harnesses can drive, and any of them can be driven.
horch ORCHESTRATOR Orchestrates a whole herdr workspace: starts review chains, builds, grills and promotions in tabs, and merges finished work into your integration branch (for example staging, never production).
hreview REVIEW CHAIN Drives the same critique passes you would run by hand, one engine per pane, until the reviewers converge or the rounds run out.
hgrill INTERVIEWS Interviews (“grills”) an issue on each configured engine in turn, then posts the biggest product calls and assumptions to the issue.
hloop BATCH BUILDS The visible counterpart to agent-loop: a build tab per agent-ready issue, then a fresh review chain, then a merge through horch’s checks.
hred-team RED TEAM Red-teams a reviewed PR in a read-only pane. The author engine fixes blocker and high findings; the rest become issues.
hpromote PROMOTION Promotes the integration branch to production by running your repository’s recipe across visible panes, with a read-only risk review and fix lanes.
auto-glance OPT-IN GLANCE Small or docs-only PRs normally wait for a quick human look (a “glance”). This lets one model that didn’t write the PR do that read and mark it ready. Opt-in per repository; never merges.
- Run the driving session inside herdr 0.9.0+
- ActiveLoom in the repo, including the Codex harness
- A review profile (
review-setup) - An authenticated
gh, plusjqand Python 3 horchmerges only withship-stagingand ashipsection in.activeloom-herd.json
What a herd session looks like.
A simulated example. One driving session (horch) runs reviews, interviews, and builds in herdr panes.
Every pass is an ordinary ActiveLoom review, recorded in the PR ledger.
Click a tab or pane to see what it is doing ↗
src/auth/refresh.ts#L142 logger.info(). Breaks the repo's no-secrets-in-logs invariant.
4d2e1b).redactSecrets() from the shared package; thread replied and resolved.What ActiveLoom deliberately does NOT do.
Software engineering tools earn trust through explicit non-goals. We don't lock you into a cloud, hide the prompts, or replace human review.
- ✔ Standardized engineering skills: Planning, debugging, and review workflows that any team agent can invoke.
- ✔ PR-first adversarial review: Verified findings posted inline to GitHub before code modifications occur.
- ✔ Commit-pinned review ledgers: Review evidence recorded on the PR and bound to the exact head commit it covers.
- ✔ Visible herd orchestration: One session drives review chains, grills, builds, and promotions in herdr panes you can watch and type into.
- ✔ Upstream fleet sync: Keep every repository updated via proposed PRs without clobbering repo-local rules.
- ✖ No proprietary agent runtime: Bring your own Claude Code, OpenAI Codex, or Google Antigravity (Agy). No vendor lock-in.
- ✖ No monthly SaaS seat tax: No hosted workspace fees, no credit tokens. Apache 2.0 open source. The npm installer has no runtime dependencies; you bring your own agent CLIs and model access.
- ✖ No unattended production merges: agent-loop opens draft PRs and the review runner never marks ready or merges.
horchships converged work only into your integration branch, and promotion to production stops at your approval and merge. - ✖ No clobbering of local rules: Upstream prompt improvements never overwrite your repository's custom compliance invariants.
From personal trial to fleet sync.
Adopt gradually. Start with a single skill in your personal setup, then scale up to scheduled multi-repo updates. (Adoption tiers are separate from Lean/Deep review depth.)
Personal Evaluation
Try a single skill in your home directory without touching any shared repository files or configurations.
Repository Setup
Scaffold harness directories (.claude/, .codex/, .agents/), then run onboard to fill project fields and shared project instructions for your team.
Scheduled Sync
Automated GitHub Actions workflow proposing upstream skill improvements via pull requests using built-in GITHUB_TOKEN.
App-Backed Fleet
Scheduled sync through a GitHub App, so propagation PRs land with GitHub-signed commits.
Give your agents real discipline today.
Start with one skill, then set up a repository when you're ready. Your workflow stays yours. Your agents' reviews now leave evidence pinned to the commit they read.
Installs one skill into your personal skills directory. Nothing is written to a repository.
init: Python 3.9+ with PyYAML · Add --dry-run to preview any installgh; agent-loop needs both the Claude and Codex CLIs