# ActiveLoom > Open-source (Apache 2.0) engineering skills, adversarial PR review protocols, commit-pinned review ledgers, bounded automation, and herd orchestration for Claude Code, OpenAI Codex, and Gemini / Antigravity (Agy). From Loomantix. Skills are distributed per harness: Claude Code `.claude/skills//SKILL.md`, Codex `.codex/skills//SKILL.md`, Gemini / Agy `.agents/skills//SKILL.md`. Links below point at the Claude Code copy on `main` unless noted; swap the root for your harness. In a consumer repository, prefer the installed copy: check `.activeloom-config.yml` (selected harnesses, sync) and `ls .claude/skills .codex/skills .agents/skills`, or run `npx activeloom detect`, which writes nothing. [llms-full.txt](https://activeloom.dev/llms-full.txt) bundles this index with the full agent entry guide (task selector, install scopes, source ownership). Skill files are linked, not inlined. ## Start here - [Agent Entry Guide](https://activeloom.dev/agent-guide.md): Copyable session prompt, task-to-skill selector, prerequisites, and source ownership map. - [README](https://github.com/loomantix/activeloom/blob/main/README.md): Scope, non-goals, and a "choose a starting point" table. - [Getting Started](https://github.com/loomantix/activeloom/blob/main/docs/getting-started.md): Adoption tiers 0–3, requirements, and onboarding. - [Everyday Workflows](https://github.com/loomantix/activeloom/blob/main/docs/workflows.md): How the common skills are started and hand off to each other. - [Set Up for Autonomous Development](https://github.com/loomantix/activeloom/blob/main/docs/autonomous-development.md): Stage-by-stage setup from install to a first allowlisted `agent-loop` run. - [Herd Orchestration](https://github.com/loomantix/activeloom/blob/main/docs/herd-orchestration.md): Driving review chains, grills, builds, and promotions in herdr panes. ## Pick by task - Design unclear before code: `grill`. Whether or what to build: `product-grill`. - Reproducible bug or regression: `diagnosing-bugs`. - One issue to a draft PR: `issues start `. Ready queue: `issues ready`. - Make the backlog agent-ready: `backlog-refinement`. Then an allowlisted queue: `agent-loop`. - Review an existing PR: the harness's `REVIEW_WORKFLOW.md`, then `critique` (Lean) or `deepcritique` (Deep). The depth comes from the workflow's triggers, not diff size. - Find how a reviewed PR fails once live: `red-team`. - Set up a repository from zero: the Set Up for Autonomous Development guide. - The same work in visible herdr panes: `hreview`, `hgrill`, `hloop`, `horch`, `hred-team`, `hpromote`. ## Core Engineering Skills - [grill](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/grill/SKILL.md): Relentless pre-code interview that stress-tests a plan or design until nothing is silently assumed. - [product-grill](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/product-grill/SKILL.md): Product interview that stress-tests a plan from the user's point of view; can conclude don't build. - [diagnosing-bugs](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/diagnosing-bugs/SKILL.md): Disciplined diagnosis loop for hard bugs and regressions: isolate, repro, RCA, minimal verified fix. - [feature-dev](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/feature-dev/SKILL.md): Guided feature development with codebase understanding and architecture focus. - [issues](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/issues/SKILL.md): Work an issue end to end (scope check, claim, worktree, implement, draft PR); ready queue with dependency resolution. - [backlog-refinement](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/backlog-refinement/SKILL.md): Prepare a GitHub backlog for autonomous work: assess, prioritize, and rewrite issues into agent-ready form. - [agent-loop](https://raw.githubusercontent.com/loomantix/activeloom/main/.codex/skills/agent-loop/SKILL.md): Autonomous issue implementation loop with strict allowlisting, one linked worktree and draft PR per issue, and bounded local review. Linked as the Codex copy; it needs both the Claude and Codex CLIs. - [handoff](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/handoff/SKILL.md): Write or resume a session handoff comment on a GitHub issue. - [onboard](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/onboard/SKILL.md): After `init`, draft the project-specific `TODO(activeloom)` values from repository evidence for confirmation. ## Review Skills - [critique](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/critique/SKILL.md): PR-first adversarial review. Posts verified findings inline before fixing. Lean mode (highest-signal lenses) vs Deep mode (full relevant review matrix). - [deepcritique](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/deepcritique/SKILL.md): High-fidelity review chain for changes whose review depth resolves to Deep. - [red-team](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/red-team/SKILL.md): Pre-merge red team of a reviewed PR: how the change fails once live, sorted into fix-before-merge, follow-up issue, or accepted risk. - [review-setup](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/review-setup/SKILL.md): Set reviewer and worker model/effort pairs and local review engine order. ## Herd Orchestration (herdr) One agent session drives other agent sessions in visible [herdr](https://herdr.dev) panes. Requires herdr 0.9.0+ with the driving session running inside it (`HERDR_ENV=1`), ActiveLoom installed in the repository including the Codex harness (the review run controller lives at `.codex/skills/critique/scripts/local-review-handoff.py`), the review profile (`review-setup`), an authenticated `gh`, plus `git`, `jq`, and Python 3. For `horch` to merge, the `ship-staging` skill (Codex and Gemini / Agy roots only; not pulled in by `add horch`) and a `ship` section in `.activeloom-herd.json` are also required. See [Before you start](https://github.com/loomantix/activeloom/blob/main/docs/herd-orchestration.md#before-you-start) and [Running unattended safely](https://github.com/loomantix/activeloom/blob/main/docs/herd-orchestration.md#running-unattended-safely). - [horch](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/horch/SKILL.md): Workspace orchestrator: starts review chains, builds, grills, and promotions in tabs; merges converged work only into the integration branch. - [hreview](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/hreview/SKILL.md): Runs the ordinary review chain on a PR, one engine per pane, until it converges or runs out of rounds. Never merges. - [hgrill](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/hgrill/SKILL.md): Grills an issue on each configured engine in turn, then posts product calls and assumptions for a human check. - [hloop](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/hloop/SKILL.md): Batch loop: a build tab per agent-ready issue, then a fresh review chain, then a ship through horch's gate. - [hred-team](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/hred-team/SKILL.md): Red-teams a reviewed PR in a read-only pane; blocker and high findings are fixed on the PR, the rest become issues. - [hpromote](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/hpromote/SKILL.md): Runs the repository's promotion recipe across visible panes; stops at the human approval and merge. - [auto-glance](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/auto-glance/SKILL.md): One non-author model stands in for the human glance on small or docs-only PRs. Opt-in per repository; never merges. ## Review Protocols & Ledgers - Review workflow, per harness: [Claude Code](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/REVIEW_WORKFLOW.md), [Codex](https://raw.githubusercontent.com/loomantix/activeloom/main/.codex/REVIEW_WORKFLOW.md), [Gemini / Agy](https://raw.githubusercontent.com/loomantix/activeloom/main/.agents/REVIEW_WORKFLOW.md). PR-first review lifecycle, review depth (Lean vs Deep), and multi-engine convergence. - [Review Ledger Protocol](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/references/local-review-ledger.md): Engine-neutral contract. The draft PR is the ledger: roster and attestation markers carry sha256 digests and name the exact head commit they cover. Evidence, not a merge gate. - [Review Chain Runner](https://raw.githubusercontent.com/loomantix/activeloom/main/.codex/references/review-chain-runner.md): Deterministic automatic review relay (Codex control surface). ## CLI & Installation - Inspect first (writes nothing): - `npx activeloom detect`: What this repository looks like (harnesses, existing config). - `npx activeloom add`: List the skills available. `npx activeloom tiers`: Explain the four adoption tiers. - Add `--dry-run` to `add` or `init` to report what would change without writing. - Install: - `npx activeloom add ... --harness `: Tier 0, personal skill directories; pulls each skill's required skills. - `npx activeloom init --harness `: Tier 1, write the selected harness trees and `.activeloom-config.yml` into this repository (`--harness` is repeatable: claude, codex, gemini; omit it to use detection). Then run the `onboard` skill. - `npx activeloom init --sync`: Tier 2 (recommended), scheduled GitHub Actions sync PRs. `--sync --app`: Tier 3, GitHub App signed commits. - `npx activeloom review-config `: Show or set the review profile. - Requirements: Node 18.17+; `init` also needs Python 3.9+ with PyYAML (`add` needs no Python). The CLI fetches content from the `sync-v2` ref by default (`--ref` overrides); links here track `main`. ## Optional - [refactorpass](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/refactorpass/SKILL.md): PR-first refactor pass: read-only cleanup proposals, verified findings posted before edits. At most once per PR per engine. - [promote](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/promote/SKILL.md): Integration-to-production promotion from the repository's `.activeloom-promote.json` recipe; stops at the human approval and merge. - [ship-staging](https://raw.githubusercontent.com/loomantix/activeloom/main/.codex/skills/ship-staging/SKILL.md): Merge converged PRs to a staging branch; required for `horch` to ship (Codex and Gemini / Agy roots only). - [dep-analysis](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/dep-analysis/SKILL.md): What a dependency update could break, the tests that exercise it, and a verdict at the exact head. - [simplify-comments](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/simplify-comments/SKILL.md): Cut comment bloat without changing code. - [copilot-review](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/copilot-review/SKILL.md) and [reviewit](https://raw.githubusercontent.com/loomantix/activeloom/main/.claude/skills/reviewit/SKILL.md): Hosted review integrations, where repository policy permits.