agy-worker
Let Codex use Google Antigravity CLI (agy) for repository exploration, bounded feature work, and project-scale implementation. Use when a Codex task benefits from delegated repository work while Codex retains diff review, verification, repair, and delivery assurance.
git clone --depth 1 https://github.com/aiskillstore/marketplace /tmp/agy-worker && cp -r /tmp/agy-worker/pending/cagdasyurekli/agy-worker ~/.claude/skills/agy-workerSKILL.md
# Delegate repository work and verify the result Use this skill for useful repository exploration, bounded implementation, or broad project work through `agy`. The worker discovers ordinary structure and proposes or makes changes; Codex remains responsible for scope, diff review, driver-owned checks, repair decisions, and the final assurance label. Resolve the installed package instead of guessing a checkout path: ```bash PIPELINE="$(bash "$SKILL_ROOT/scripts/resolve-pipeline.sh")" || exit $? ``` Optionally check offline prerequisites before spending provider quota: ```bash "$PIPELINE/doctor.sh" --repo /absolute/path/to/target ``` `ready` covers offline prerequisites only. It does not prove authentication, provider availability, task quality, or future job success. For installation and package orientation, read [Package README](README.md). ## Before every provider launch Obtain explicit approval for the exact transmission mode, repository content, and task being sent unless that exact provider transmission was already approved. Without `--provider-scope`, treat the entire disposable worktree passed as `--workdir` as worker-readable and potentially transmissible to Google/Gemini, regardless of requested edit paths; `--add-dir`, prompt denylist instructions, `qa-gate --only`, and `--allow` do not narrow that default read boundary. Optional `--provider-scope FILE --approve-transmission-sha SHA256` instead binds exact reviewed read entries, their selected-content digest, and a write subset, then stages only selected entries in a fresh owner-private mode-`0700` Gitless provider cwd. The primary `workflow.sh run` facade has no implicit transmission mode: launch requires either `--approve-whole-worktree MANIFEST_SHA256` or the scoped pair above. The deprecated `--approve-preview-sha` spelling cannot launch by itself and remains temporarily available only with `--legacy-preview-approval`. The controller still locally enumerates and validates worktree paths and the scope policy before staging; scoped mode is not a filesystem, network, `PATH`, `HOME`, or same-UID sandbox and retains the documented local-owner and mutation-race residuals. Provider-scope approval grants neither provider execution, Git action, driver acceptance, nor publication. Before each launch, ensure secrets, credentials, private keys, user-denied paths, and unrelated private files are absent from every entry approved for provider transmission; telling the worker not to read an approved entry is not a control. Keep raw worker logs and local controller state outside the worktree and out of prompts. Installation does not authorize provider transmission, Git actions, publication, or acceptance. Before every provider-launch attempt (initial start/run, resume, continue, and restart), tell the user in one or two concise user-facing sentences what task is being sent to AGY. Include a short public-safe task label, caller-selected model information, caller-selected effort when separately selectable, and the exact resolved model slug. For default selection where no model is selected or the default tier is used, state truthfully that the provider default model is used and that model or effort is unresolved, without inventing a resolved slug or thinking level. For fixed/compound/literal models where effort is not separately selectable, state that accurately without inferring backend reasoning or inventing a thinking level. The notice must precede every dispatch attempt and remain accurate afterward. If preflight fails before provider launch, explicitly state that the task was not sent to AGY. If provider reach is genuinely uncertain, state that it is unverified rather than claiming success. Direct model and effort selection remain caller-owned; recommendations are advisory. For the complete transmission, environment, verifier, and compatibility boundaries, read [Security and compatibility](references/SECURITY_AND_COMPATIBILITY.md). ## Route the request | User intent | Workflow | Default cycle budget | Driver responsibility | |---|---|---:|---| | Explore, understand, review, or plan | `explore` | 2 | Spot-check material claims and state coverage limits. | | Implement a feature, refactor, tests, or a bounded repair | `task` | 2 | Inspect the diff and run relevant project checks. | | Build a project or perform broad audit-and-fix work | `project` | 5 | Review repo-wide changes and run build/test/lint as applicable. | `explore` and `task` accept `1..2` cycles; `project` accepts `1..5`. Personas are optional prompt specializations, not capability, approval, or quality gates. Do not route ordinary work through compatibility evidence or data-only profile commands. For material UX, lifecycle, trust-boundary, security, data-semantics, or other domain plans: A coordinator and suitable domain expert must co-plan. Freeze user journeys, acceptance tests, and authority/privacy constraints before implementation. The final acceptor must be a different agent or fresh context; no planner or implementer may self-accept. Purely mechanical changes are exempt. Verification v2 and the controller bind candidate evidence, not agent identity or governance. The final human-readable handoff must report the planner/reviewer separation. Explicit delegation-first requires running the `delegation-policy.sh` evaluator before substantive repository work. The controller records are local: the runtime cannot infer prior work or approval and must never silently authorize direct-Codex fallback after a missing approval, hard stop, preflight failure, provider failure, or exhausted budget. Direct-Codex and second-eye work remain explicit policy choices. ## Use the primary lifecycle Prefer the portable `workflow.sh` facade for `run --preview`, approved `run`, read-only `status`, and `verify-finalize`. For ordinary use, supply only an absolute repository and job ID; optional `--base` overrides the first-call `HEAD` binding. The facade derives owner-private state
Use when replying. Refresh current date and time.
Create polished, validated architecture, workflow, sequence, data-flow, and lifecycle/state diagrams as explorable standalone HTML with inline SVG, dark/light themes, optional trace motion, and PNG/JPEG/WebP/SVG/WebM export. Accept plain-language requirements or pasted Mermaid flowchart, sequenceDiagram, and stateDiagram input; inspect repository evidence when the diagram must reflect real code. Use when the user asks to visualize system architecture, infrastructure, cloud/security/network topology, technical workflows, API call sequences, request lifecycles, data pipelines, ETL/ELT, data lineage, state machines, or to convert/beautify Mermaid.
Design systems, plan implementations, review architecture decisions - Use when you need to plan a complex feature, design system architecture, or make high-level technical decisions.
Use when automating iOS Simulator interactions, capturing screenshots/video, or inspecting accessibility via AXe CLI.
Use when working with Canton Network participants, DAML smart contracts, Splice applications, or debugging LF version and package ID issues.
>
HSK4級レベルから流暢さを目指す学習者向け。中国語表現の使用場面・自然さを分析し、作文を「ネイティブらしい流暢な表現」に改善。bilibili等のコンテンツ理解とネイティブとの会話をサポート。実際の用例をWeb検索で提示
Troubleshoot Claude Code extensions and behavior. Triggers on: debug, troubleshoot, not working, skill not loading, hook not running, agent not found.