yylo — Claude Code plugin: what it does, how to install
yylo is a Claude Code and Codex plugin that isolates agent tasks in separate git worktrees and coordinates Pi and Codex subagents, verifying changes before merge.
MIT ★ 1,596 updated 2026-10-29
/plugin marketplace add hashgraph-online/awesome-codex-plugins
/plugin install yylo@awesome-codex-pluginsWhat yylo does
The plugin orchestrates coding agents straight from the command line. For each incoming task, it spins up an isolated git branch and worktree. This keeps your working tree clean while an agent modifies files.
Execution is handed off to Pi and Codex subagents. The tool enforces boundaries across task types, validation checks, and release readiness. Every change inside the repository is backed by receipts to confirm what was altered.
Your environment must support git worktree checkouts for this workflow to function. When branch operations fail on an untracked local conflict, the orchestrator halts execution.
What it helps with
- isolate agent tasks into separate git worktrees
- delegate coding jobs to Pi and Codex subagents
- enforce release-readiness boundaries before merging
- track repository modifications with structured receipts
Example prompts for the agent
- Run the refactoring task inside an isolated worktree using the Codex subagent
- Verify branch changes against the release boundary and validate receipts
- Delegate bug reproduction to the Pi subagent in a dedicated worktree
How to install yylo
Plugin marketplace
/plugin marketplace add hashgraph-online/awesome-codex-plugins/plugin install yylo@awesome-codex-pluginsStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about yylo
How to install yylo in Claude Code?
First add the catalog with /plugin marketplace add hashgraph-online/awesome-codex-plugins. Then run /plugin install yylo@awesome-codex-plugins.
What does yylo do?
It manages coding workflows by setting up a dedicated git worktree per task. It directs work to Pi and Codex subagents and verifies typed boundaries before merges.
Which subagents does yylo work with?
The tool delegates tasks to Pi and Codex subagents. Each task runs within strict boundaries backed by change receipts.
Source text
yylo · MIT
# YYLO CLI [](https://github.com/kailiu42/awesome-coding-agents) [](https://github.com/Piebald-AI/awesome-gemini-cli) [](https://github.com/hanishrao/collective-ai-tools) [](https://github.com/andyrewlee/awesome-agent-orchestrators) [](https://github.com/currenjin/site-for-developers) [](https://github.com/acvnace/awesome-vibe-coding-resources) [](https://github.com/agentrust-io/awesome-ai-governance) [](https://github.com/alantriesagain/awesome-local-first) [](https://github.com/alvinreal/awesome-opensource-ai) [](https://github.com/ARUNAGIRINATHAN-K/awesome-ai-agents-2026) [](https://github.com/Aurelius77/AI-TOOLS-DIRECTORY) [](https://github.com/awesome-gptX/awesome-gpt) [](https://github.com/brandonhimpfen/awesome-ai-agents) [](https://github.com/charlesteh/awesome-ai-tools) [](https://github.com/chgaowei/ai-agent-infra-list) [](https://github.com/CiferaTeam/awesome-agents-team) [](https://github.com/DhanushNehru/Ultimate-NodeJs-Resources) [](https://github.com/dr-mushtaq/Awesome-AI-Tools-Resources) [](https://github.com/ejboy/awesome-efficient-devtools) [](https://github.com/Engineering4AI/awesome-spec-driven-development) [](https://github.com/frangelbarrera/awesome-ai-agents) [](https://github.com/HA2345567/awesome-autonomus-ai-agents) [](https://github.com/ichangyou/awesome-codex) [](https://github.com/ishandutta2007/Awesome-Agent-Orchestration-Platforms) [](https://github.com/ishandutta2007/awesome-ai-tools) [](https://github.com/ishandutta2007/Awesome-Multi-Agent-AI-Harnesses) [](https://github.com/kiliczsh/ailandscape.org) [](https://github.com/lgaggini/awesome-cli-tui-software) [](https://github.com/LinuxSuRen/awesome-ai) [](https://github.com/M1n9X/llm_agents_devtools) [](https://github.com/michielhdoteth/awesome-ai-agent-tools) [](https://github.com/milisp/awesome-chatgpt-claude-agents) [](https://github.com/milisp/awesome-codex-cli) [](https://github.com/planetabhi/riseofmachine) [](https://github.com/rudy2steiner/awesome-agent-loops) [](https://github.com/RyanAlberts/best-of-Agent-Harnesses) [](https://github.com/sajadh76/awesome-AI-tools) [](https://github.com/shaftoe/awesome-pi-coding-agent) [](https://github.com/shospodarets/awesome-platform-engineering) [](https://github.com/systempromptio/awesome-ai-agent-governance) [](https://github.com/tejas-singh-0212/tool-directory) [](https://github.com/tysoncung/awesome-vibe-coding) [](https://github.com/yadavanujkumar/awesome-agentic-ai) [](https://github.com/yeaight7/awesome-ai-devtools) [](https://github.com/youxufkhan/awesome-ai-coding) [](https://github.com/sraodev/awesome-macOS) [](https://github.com/Jenqyang/Awesome-AI-Agents) [](https://github.com/robertsmrek/AI-Tools-Catalogue) [](https://github.com/frangelbarrera/Artificial-Intelligence-Universe) [](https://github.com/NipunaRanasinghe/awesome-ai-agents) [](https://github.com/Piebald-AI/awesome-gemini-cli-extensions) [](https://github.com/dalisoft/awesome-ai-coding) [](https://github.com/thechampagne/awesome-windows) [](https://github.com/jimy-r/awesome-agent-workspaces) [](https://github.com/neomatrix369/awesome-ai-ml-dl) [](https://github.com/hadez8877/awesome-opensource) [](https://github.com/ishandutta2007/Awesome-Antigravity-Skills) [](https://github.com/ArshdeepGrover/ai-tools-manager) [](https://github.com/CodeSigils/awesome-agent-trust) [](https://github.com/abordage/awesome-mac) [](https://github.com/yerdaulet-damir/awesome-solo-ai) [](https://github.com/benjamincanac/whichcodingtools) [](https://github.com/Awakehsh/awesome-agent-tools) [](https://github.com/fabricioctelles/jump-skills) [](https://github.com/last9/awesome-sre-agents) [](https://github.com/Scottcjn/awesome-agents) [](https://github.com/hashgraph-online/awesome-ai-plugins) YYLO is a command-line orchestrator for coding agents, repeatable workflows, and receipt-backed repository changes. It is for developers who want a quick agent loop and for project operators who need typed task, validation, merge, and release-readiness boundaries. - npm package: [`@yylo/cli`](https://www.npmjs.com/package/%40yylo%2Fcli) - Commands: `yylo` and `yy` (equivalent); `ypl` is `yy pi --live` - Source: [yylo-dev/yylo](https://github.com/yylo-dev/yylo) YYLO orchestrates work. [YYLO Ledger](https://github.com/yylo-dev/yylo-ledger) is the independent Git-native Record/task store, and [YYLO Benchmark](https://github.com/yylo-dev/yylo-benchmark) is the independent evaluation/evidence package. `yy ledger` and `yy benchmark` delegate to those separately installed CLIs; they are not bundled alternate implementations. ## Terminal **Live agent session** — `ypl` (`yy pi --live`) runs an interactive terminal agent; tool calls, streamed output, and the status bar render in the terminal:  **Execution summary** — after each run the orchestrator prints iterations, tool calls, cost, and session ids:  ## Quick start: install to first successful command **Prerequisites:** Node.js 20.10 or newer, npm, and Git. ```bash npm install --global '@yylo/cli@latest' yy --version mkdir yylo-demo cd yylo-demo git init yy init --task "Document the onboarding path" --subagent pi pwd # execute explicitly; watch is an optional read-only observer ``` A successful run prints the installed YYLO version, initializes `.juno_task/`, then emits a watch receipt with `"state":"COMPLETED"`, `"exit_code":0`, and nonzero `log_bytes`. This canary does not contact a model provider. In an empty, unborn Git repository, `yy init` creates the initial workspace commit, keeps the original branch as the product target, and creates a detached protected integration-owner worktree under the user state directory. Existing or dirty repositories are never committed or rearranged by this bootstrap. Inspect the initialized workspace and exact command surface: ```bash yy info --json yy doctor workspace yy --help ``` `doctor workspace` is intentionally nonzero when it finds an actionable topology problem; it never fetches or changes the workspace. ### Stable and prerelease channels The stable npm channel is `@latest` (`0.2.2`). Prereleases remain on `@next`: ```bash # Stable npm install --global '@yylo/cli@latest' # Explicit prerelease npm install --global '@yylo/cli@next' # Exact version matching this checkout (once published) npm install -g @yylo/[email protected] npm view '@yylo/cli' version dist-tags --json yy --version ``` This unreleased source targets CLI **0.2.12**, requiring **Ledger 0.4.0 exactly** for universal Record retrieval and storage-kind-prefixed IDs. Older Ledger 0.3.3, prereleases, and later versions are rejected before delegation. Install the matching Ledger explicitly once published; ordinary delegation does not auto-install a substitute. A source bump neither publishes these versions nor activates installed runtimes. Pin an exact version in CI. Installing `@next` is an intentional prerelease choice. The first guarded release-helper checkpoint is exact `--set v0.1.0-rc.1`; later releases must use their separately authorized exact SemVer. Next: [run an agent](#beginner-agent-workflow), [manage a typed task](#typed-task-and-merge-flow), or [build a managed workflow](#managed-workflows-and-evidence). ### Install agent skills YYLO skill content is versioned independently in the public [`yylo-dev/yylo-skills`](https://github.com/yylo-dev/yylo-skills) repository and is not bundled in `@yylo/cli`. This CLI source requires stable skills `^2.1.1`, declared in `package.json` as `yyloSkills.version`. Install the latest compatible stable release, or pin an exact compatible version once it is published. Source version 2.1.1 is not evidence that this skills release is available: ```bash yy skills install yy skills install --version 2.1.1 yy skills update yy skills status ``` Fresh Simple `init` installs skills by default; use `yy init --mode simple --no-skills` to opt out. Guided Simple setup accepts the same flag. Previews remain offline and saved fresh plans bind the choice. Otherwise use `skills install` or `skills update` explicitly. These skill acquisition paths access the network. Acquisition is staged through `npx skills add` first and falls back to a shallow exact-tag Git clone. The seven user-intent-first skills (`artifact-yylo`, `ledger-tasks-yylo`, `plan-ledger-tasks-yylo`, `ralph-loop-yylo`, `understand-project-yylo`, `wiki-yylo`, and `workflow-yylo`) are copied to `.agents/skills`, `.claude/skills`, and `.pi/skills`. During install/update, unchanged receipt-owned copies upgrade automatically without `--force`. Customized or unrecorded differing canonical directories refuse before mutation; review before explicitly forcing replacement. An explicit install/update retires a legacy YYLO skill only when its local install record and current digest prove it unchanged; customized/unrecorded legacy and unrelated skills remain preserved. Status flags these legacy copies because they can still inject obsolete instructions; review them separately, never infer cleanup authority from successful installation. `skills list` and `skills status` inspect local bytes and receipts offline. Status reports the required range and outdated releases even when their hashes match. No compatible published release means installation fails without changing skills; it never falls back to an old release or an unsupported future major. Installing upgrading the CLI alone does not install/update skills, and `scripts update` does not own skills. Maintainers must publish the reviewed immutable skill release before shipping a CLI that requires it; publication remains separately authorized. ## What YYLO owns | Need | Public surface | Boundary | | --- | --- | --- | | Agent run | `yy pi`, `yy start`, and the agent aliases listed by `yy --help` | Provider credentials and model availability remain external. | | Session continuity | `continue`, `clone`, `branches`, `switch`, `continuity` | Scope state is isolated and explicit; cleanup is planned and reversible. | | Optional observation | `watch status|await|follow` | Read existing execution evidence; never launch, retry, cancel, or complete tasks. | | Validation evidence | `evidence run|status|await` | Content-addressed task evidence tied to exact inputs. | | Repository topology | `info`, `where`, `doctor workspace`, `integration` | Read-only discovery is separate from guarded sync/repair/push. | | Feature lifecycle | `task start|run|status|checkpoint|preflight|finish` | Implementation belongs in the returned exact-base task worktree. | | Protected delivery | `merge status [TASK_ID]`, `merge land TASK_ID`, `merge project TASK_ID` | One-task native Git composition plus expected-old ref update; Ledger projection is separate and retryable. | | Maintainer release | repository `scripts/release-cli.sh` | Prepare is read-only; publish needs separate authority. | | Records/evaluation | `ledger`, `benchmark` | Transparent delegation to independently installed canonical packages. | Run `yy --help` for the complete top-level inventory of your installed version; each listed command prints its own usage when invoked with `-h`. The old `lifecycle` command is removed; use typed `task` and `merge` commands. ### Stable machine output Use `--format json|ndjson --raw` on task, merge, and integration commands, or Ledger's `-f json|ndjson --raw`, for a strict stdout-only data channel. The `capabilities` command publishes command-specific projections in either machine format. See [Machine output contract](docs/machine-output.md) for the versioned envelope and migration policy. ## Beginner agent workflow Install the coding agent you intend to use and configure its provider credentials separately. Pi is optional and supports multiple providers: ```bash npm install --global '@mariozechner/pi-coding-agent' yy pi --help yy pi --no-session 'Summarize this repository and make no changes' ``` The final command may contact the configured model provider. `--no-session` prevents Pi session persistence; it does not disable provider usage. For a reusable prompt or shell-sensitive text, prefer a file: ```bash printf '%s\n' 'Explain the test layout. Do not edit files.' > prompt.md yy pi --prompt-file prompt.md --no-session ``` Interactive Pi uses the `ypl` shortcut: ```bash ypl 'Inspect the current task' ``` `ypl` expands to `yy pi --live`. Live mode requires an interactive terminal and enabled Pi extensions. ### Controlled iterations ```bash yy -s pi -m :gpt -i 3 -p 'Implement the next small verified increment' yy loop -n 2 \ --step 'yy pi "Implement the next increment"' \ --step 'npm test' ``` Quote prompts so the shell does not expand backticks or `$()` before YYLO receives them. `-i` bounds iterations inside an agent invocation; `yy loop -n` bounds the outer command workflow. The loop's `-n/--iterations` bounds the outer command workflow; agent `-i/--max-iterations` still bounds iterations inside an agent invocation. Inline steps are repeatable: ```bash yylo loop -n 5 \ --step 'yy pi "Implement the next increment"' \ --step 'yy cc "Inspect and improve your work"' \ --step 'npm test' ``` For a reusable workflow, save the same contract as `flow.yaml`: ```yaml iterations: 5 continuity: iteration on_error: continue steps: - run: yy pi "Implement the next increment" - run: yy cc "Inspect and improve your work" - run: npm test ``` ```bash yylo loop --workflow flow.yaml ``` Every step receives one-based loop metadata through `YYLO_LOOP_ID`, `YYLO_ITERATION`, `YYLO_ITERATION_COUNT`, `YYLO_STEP`, and `YYLO_STEP_COUNT`. ### Ledger references in prompts Use `##ABC123`, `##readable-slug`, or `##{readable-slug}` to include a Ledger Record by exact ID, slug, or retained alias. Existing `## ABC123` spacing works. Tasks, Wiki/PDR/workflow documents, and artifacts share native Record lookup; ambiguous identities are never resolved by guessing. Brace references next to punctuation that could be part of a slug (for example `##{design.v2}.`). Missing references remain unchanged. Lookup errors produce a manual-resolution warning; references do not execute workflows or recursively expand Record text. Hydration is bounded to 32 distinct identities, a shared 30-second lookup budget, and 64 KiB of inserted context (16 KiB per Record). Oversized Records carry a retrieval command instead. Artifacts include metadata; only verified inline UTF-8 text up to 4 KiB is included. Binary, local, and external artifact bytes are not automatically read or downloaded. Use `yy ledger record get ID_OR_SLUG` for explicit retrieval. Immutable IDs remain authoritative for task lifecycle and Record relations. ## Models and project shortcuts `yy pi --help` is the source of truth for shipped aliases. Current Pi shortcuts include: | Shortcut | Resolved model | | --- | --- | | `:luna` | `openai-codex/gpt-6-luna` | | `:sol` | `openai-codex/gpt-6.1-sol` | | `:gpt` | `openai-codex/gpt-6-astra` (Pi default) | | `:astra` | `openai-codex/gpt-6-astra` | | `:mini` | `openai-codex/gpt-5.6-terra` | | `:sonnet` | `anthropic/claude-sonnet-4-6` | | `:opus` | `anthropic/claude-opus-4-6` | Aliases are subagent-specific; for example, Pi and Codex do not share the same `:mini` mapping. Set a per-project default: ```bash yy pi set-default-model :sol ``` Add project shortcuts in `.juno_task/config.json`: ```json { "modelShortcuts": { "pi": { ":team-default": ":sol" } } } ``` Project shortcuts are scoped to the selected subagent and can reference shipped or project shortcuts. Unknown targets, malformed data, and cycles fail with an actionable error. Managed Workflow Runner model authorization is separate: explicit selectors must be exact members of `workflowModels`; an unflagged `yy pi` continues to inherit the configured default. ## Observable local commands `watch` optionally observes existing run evidence without owning execution: ```bash npm test # execute explicitly in the authorized workspace # If an existing producer has published a watch-compatible run: yy watch status RUN_ID yy watch follow RUN_ID yy watch await RUN_ID ``` `RUN_ID` is an existing run identity, not one created by `npm test`. `watch exec` is retired. All watch operations are read-only, and interruption stops only the observer. Missing or malformed terminal evidence is not success. Process exit, semantic outcome, cleanup, and task completion are distinct; `task finish` verifies delivery. Historical run directories are preserved. ## Managed workflows and evidence Fresh `yy init` installs managed scripts, prompts, and wiki guidance under `.juno_task/`. Update checksum-managed assets with: ```bash yy scripts update ``` Locally customized files are preserved and a package candidate is written to a managed-conflict path. `--force` is an explicit replacement operation and creates backups; inspect conflicts before using it. ### Workflow Runner Use a reviewed YAML workflow for ordered, repeatable steps: ```bash ./.juno_task/scripts/workflow_runner.sh \ --init-example agent-chain .juno_task/workflows/agent-chain.yaml ./.juno_task/scripts/workflow_runner.sh lint \ --workflow .juno_task/workflows/agent-chain.yaml ./.juno_task/scripts/workflow_runner.sh \ --workflow .juno_task/workflows/agent-chain.yaml --dry-run \ --print-output none --no-print-step-stdout ``` A workflow run retains rendered command identity, stdout/stderr, responses, session IDs, declared receipt hashes, attempts, and terminal manifests. `--tmux` creates an observer session; it does not detach the producer. If a producer was interrupted, diagnose before mutation: ```bash ./.juno_task/scripts/workflow_runner.sh recover-attempt RUN_DIRECTORY --dry-run ./.juno_task/scripts/workflow_runner.sh doctor RUN_DIRECTORY ``` `RUN_DIRECTORY` is the printed durable run path. Recovery resumes only the first invalid step after verifying unchanged successful evidence. Never edit historical manifests to make them reusable. ### Task validation evidence At a clean coherent task commit: ```bash yy task checkpoint TASK_ID yy evidence run TASK_ID yy evidence status TASK_ID yy evidence await TASK_ID ``` These commands plan and retain exact-input validation evidence. They do not finish or merge the task. ## Typed task and merge flow This section is for repositories initialized with the current controller/task policy. Run lifecycle commands from the registered metadata controller. Never edit the integration-owner checkout. ### Deterministic delivery, external implementation ```bash yy task start TASK_ID # External agent edits, tests and commits in the returned worktree. # Retain the returned token privately for gated commands: yy task finish TASK_ID --lease-token <current-token> yy merge land TASK_ID # Only if automatic Ledger projection needs recovery: yy merge project TASK_ID ``` `TASK_ID` is a Ledger task ID. Start prepares an isolated workspace; finish verifies the clean committed result and queues it. Merge uses native Git and automatically records verified integration in Ledger. Tests and semantic review remain explicit project checks. Task run/resume and automatic implementation budget recovery are retired. No implementation model, retry or repair engine is hidden behind watch. Process exit is not completion. For interrupted work, preserve the existing workspace and inspect status/lease-status before explicit continuation; do not automatically reset or replay historical attempts. ### Choose Simple or Advanced For a new Simple project, run `git init`, then `yy init --mode simple` to initialize immediately. Add `--directory PATH` to select an existing Git root. For guided setup, `yy init` asks a final mode question: - **Simple (recommended to start):** code, notebooks, notes and Ledger in one checkout. Agents share files; `yy task local` is bookkeeping, not delivery. No implicit commits, worktrees or package installation. - **Advanced:** a separate controller, isolated task worktrees and managed merging. Use `yy init --interactive --mode simple` to skip the mode question. Inline initialization retains Advanced behavior; `yy init "Build an API" --mode advanced` makes that explicit. **Plain `yy init --mode simple` now writes files**, without confirmation or an external plan. Existing preview automation must add `--dry-run` (zero writes) or use optional `--plan-file` then `--apply-plan` for reviewed plans. Initialization installs no dependencies and does not certify agent readiness. Normal initialization does not convert existing workspaces. For a settled registered metadata-only controller, preview then apply a fresh Advanced → Simple copy: ```sh yy init --mode simple --from-advanced /controller --directory /new-project --plan-file /external/conversion.json yy init --mode simple --apply-plan /external/conversion.json ``` Source workspaces remain unchanged. Old instructions and settings are retained inactive for manual review. Simple → Advanced is not supported, including with `--force`. Read [Simple and Advanced workspaces](docs/simple-workspaces.md) for exact commands, readiness recovery, persistence, and shared-checkout limits. ### Manual implementation path (managed mode) ```bash yy task start TASK_ID # Change directory to the worktree printed by start. # Read its AGENTS.md/CLAUDE.md, implement, run focused tests, and commit. yy task finish TASK_ID yy merge status TASK_ID yy merge land TASK_ID yy merge project TASK_ID ``` The ordinary path is clean commit -> `yy task finish TASK_ID`. Optional read-only `yy task preflight TASK_ID` diagnoses admission before finish; it is not a prerequisite. Delivery remains a separate one-task operation. Safety invariants: 1. `task start` freezes the protected target SHA, creates a dedicated branch/worktree, and completes configured dependency hydration before reporting `WORKING`. 2. Product edits and focused tests occur only in that task worktree. Controller metadata and integration-owner product bytes are separate authorities. 3. `preflight` is optional and read-only. `finish` independently enforces scope, dirty-byte, runtime, hydration, requirements, ownership, and selected-validation checks, requires a clean committed tip, and queues it; it does not merge. 4. Tests and semantic reviews are explicit project checks outside merge. Merge launches no models, chooses no reviewers, schedules no suites, and maintains no validation cache. 5. `merge land` selects one immutable task source, composes in a private detached candidate, and uses Git expected-old ref protection. A moved target requires recomposition and renewed candidate checks. 6. A conflict remains private to its task and cannot block an unrelated task. Preserve conflicts and unrelated dirty bytes; do not reset, stash, force, rebase, squash, or clean to bypass them. 7. Git success and Ledger projection are separate. A projection retry never repeats Git integration and never blocks another independent land. Observation commands are safe to repeat: ```bash yy task status TASK_ID yy task doctor TASK_ID yy merge status yy merge status TASK_ID ``` Merge status reports only independently landable active tasks and always reports `model_calls: 0`. Historical queue receipts remain immutable data, not an executable compatibility surface. ## Maintainer npm release Normal tasks and `yy merge` own integration. After an ordinary version bump is integrated, maintainers use `scripts/release-cli.sh prepare CLI_VERSION BENCHMARK_VERSION`, obtain explicit publication approval, and then run the separate `publish` command. The CLI has no release command. ## Controller generation upgrades Recognized controller generations migrate automatically on first task/agent use; controller-local scripts, policy and executable identity move together. Diagnose without changing state using `yy scripts generation doctor`. Unknown/customized bytes are preserved, and fallback requires a trusted retained executable plus scripts. See [upgrade, recovery and validation guidance](docs/controller-generation-upgrades.md). Maintainer release preparation gates the exact npm tarball and binds its immutable acceptance report; this does not authorize publication. ## Receipt-bound workspace relocation When an entire controller is moved between machines, do not rewrite lifecycle JSON or historical receipts by hand. From a clean physical controller checkout, create and review an external plan, apply it once, then verify its immutable receipt: ```bash node juno-code/scripts/workspace-relocation.mjs plan --controller "$PWD" \ --map /old/controller=/new/controller --map /old/worktrees=/new/worktrees \ --output /secure/relocation-plan.json node...
© YYLO contributors · MIT · hashgraph-online/awesome-codex-plugins