rock3r

Babysit a GitHub pull request after creation by continuously polling CI checks/workflow runs, new review comments, and mergeability state until the PR is ready to merge (or merged/closed). Diagnose failures, retry likely flaky failures up to 3 times, auto-fix and push branch-related issues when appropriate, and stop only when user help is required (e.g. CI infrastructure issues, exhausted flaky retries, or ambiguous/blocking review feedback). Use when the user asks to monitor a PR, watch CI, handle review comments, or keep an eye on failures and feedback on an open PR.

Use when the user is writing, running, or debugging end-to-end UI automation tests for a Compose for Desktop (aka Compose Multiplatform on desktop) app — including setting up Spectre, clicking or typing in a running window, finding UI nodes by tag or text, waiting for the UI to settle, capturing screenshots or video, or troubleshooting a Compose Desktop or IntelliJ-hosted UI under test. Also use when the user asks how to automate a live running Compose Desktop app, even if they don't mention Spectre by name.

Use when preparing, validating, publishing, or undrafting a Spectre release.

Use when writing, debugging, or reviewing tests that drive a real, on-screen Compose Desktop window with Spectre — the JVM/Kotlin library for automating live Compose Desktop UIs (and IntelliJ/Jewel-hosted Compose) via the semantics tree plus java.awt.Robot, synthetic AWT input, HTTP, or Java-agent attach. Trigger on mentions of `ComposeAutomator`, `RobotDriver`, `AutomatorNode`, `AgentAttach`, `AttachedAutomator`, `AttachOptions`, `findByTestTag`, `waitForNode`, `waitForIdle`, "automate a Compose window", "live/real-window Compose Desktop UI test", cross-JVM Compose automation, attaching to a running Compose JVM, screenshotting or recording a Compose Desktop window, or any task involving `dev.sebastiano.spectre.*` imports. Also use when the user is writing JUnit 4/5 tests against Compose Desktop **and** the test opens a real top-level window — but NOT when the user wants the off-screen `runComposeUiTest` / `createComposeRule` / `onNodeWithTag` framework, which is a different tool.

Use when working on IntelliJ/Jewel popup hosting, ComposePanel embedding, SwingBridgeTheme, popup rendering modes, or AWT/Compose coordinate conversion.

Use when Detekt reports a violation and you need to fix the underlying problem without papering over it with suppressions or configuration changes.

Use when starting nontrivial feature work that should stay isolated from the main checkout. Creates or reuses a safe git worktree with sensible defaults.

Watch an open GitHub pull request for CI, mergeability, bot reviews, and human feedback until it is ready, closed, or blocked.

Harness-portable autonomous delivery loop for a scoped software task: front-load uncertainty, produce or confirm a plan, implement with TDD and validation, run review passes, prepare git/GitHub artifacts, and babysit the PR when authorized. Use when the user wants a mostly hands-off problem-to-PR workflow.

Use for commit, push, pull request, issue, and GitHub CLI workflows; emphasizes diff-grounded text and safe non-interactive commands.

Harness-portable planning and requirements-capture workflow for turning vague, risky, architectural, multi-workstream, or scope-drift-prone requests into durable artifacts before implementation. Use when the user asks to scope, plan, workshop, write a spec, capture acceptance criteria, prepare an implementation handoff, or when product intent should not be guessed.

Read-only reviewer for checking whether an implementation milestone matches an approved plan or spec.

Guide agents to choose a review pass intentionally and keep findings structured.

Use when starting feature work that should be isolated from the main checkout; creates or verifies an isolated git worktree with safety checks.

Babysit a GitHub pull request after creation by continuously polling CI checks/workflow runs, new review comments, and mergeability state until the PR is ready to merge (or merged/closed). Diagnose failures, retry likely flaky failures up to 3 times, auto-fix and push branch-related issues when appropriate, and stop only when user help is required (e.g. CI infrastructure issues, exhausted flaky retries, or ambiguous/blocking review feedback). Use when the user asks to monitor a PR, watch CI, handle review comments, or keep an eye on failures and feedback on an open PR.

Use when auditing Compose/Jewel UI for SelectionContainer context: interactive composables inside a SelectionContainer without DisableSelection, nested selection containers, or missing exclusion wrappers. Uses indexino (selection-context application) for precomputed facts from a persistent `.indexino/` index — not manual file reading. Bazel monorepos primary; Gradle secondary.

Use when the user asks to audit, assess, review, or find systemic issues in Compose Desktop or Jewel UI architecture: state/effect usage, recomposition behavior, lazy-list keys, composable API hygiene, modifier conventions, and desktop interop seams. For SelectionContainer / DisableSelection context, run compose-selection-audit first. Produces an evidence-first report with severity-ranked findings and top fixes. Desktop/Jewel focus — not Android lifecycle, Navigation, Material mobile patterns, or baseline profiles.

Handle git and GitHub workflows end-to-end: summarize diffs into high-quality commit messages, create commits non-interactively, push branches, prepare PR titles/descriptions, and operate GitHub via gh CLI with file-based inputs to avoid shell escaping bugs. Use whenever a task includes commit, push, PR creation/updates, review replies, merge prep, or gh automation.

Create, revise, evaluate, package, and optimize agent skills. Use when users want to turn a workflow into reusable skill guidance, improve an existing skill, add evals or bundled resources, test skill behavior with subagents, benchmark skill quality, or tune a skill trigger description.

Use when starting feature work that needs isolation from current workspace or before executing implementation plans - creates isolated git worktrees with smart directory selection and safety verification

Safely send or receive one explicit Punaro v3 attachment through the locally provisioned controller. Use for a typed offer whose body has the exact `punaro/attachment-offer/v3:` marker, or when the task owner explicitly authorizes one pre-provisioned local file transfer.

Receive and safely handle messages delivered through a local Punaro-connected agent-mailbox. Use when an agent must await incoming mail, inspect a Punaro typed envelope, acknowledge a completed delivery, or diagnose a local mailbox wake-up without changing relay enrollment or routing.

Reply to a message delivered through Punaro or agent-mailbox using the conversation ID in its typed envelope. Use when an agent receives application/vnd.punaro.message+json, a user asks for a reply through an existing Punaro conversation, or a Telegram-routed agent request needs a durable response.

Monitor a Punaro pull request until checks, reviews, and merge are complete.

Keep nontrivial Punaro work isolated from main.

Babysit a GitHub pull request after creation by continuously polling CI checks/workflow runs, new review comments, and mergeability state until the PR is ready to merge (or merged/closed). Diagnose failures, retry likely flaky failures up to 3 times, auto-fix and push branch-related issues when appropriate, and stop only when user help is required (e.g. CI infrastructure issues, exhausted flaky retries, or ambiguous/blocking review feedback). Use when the user asks to monitor a PR, watch CI, handle review comments, or keep an eye on failures and feedback on an open PR.

Use when auditing Compose/Jewel UI for SelectionContainer context: interactive composables inside a SelectionContainer without DisableSelection, nested selection containers, or missing exclusion wrappers. Uses kotlin-code-index (selection-context application) for precomputed facts from a persistent `.kotlin-index/` index — not manual file reading. Bazel monorepos primary; Gradle secondary.

Use when the user asks to audit, assess, review, or find systemic issues in Compose Desktop or Jewel UI architecture: state/effect usage, recomposition behavior, lazy-list keys, composable API hygiene, modifier conventions, and desktop interop seams. For SelectionContainer / DisableSelection context, run compose-selection-audit first. Produces an evidence-first report with severity-ranked findings and top fixes. Desktop/Jewel focus — not Android lifecycle, Navigation, Material mobile patterns, or baseline profiles.

Use when creating or updating a skill-source.json provenance sidecar for a new, copied, or modified skill, checking whether upstream sources have changed, or folding external guidance into a local skill.

Use proactively before every Gradle invocation, and whenever a Gradle run behaves unexpectedly because of daemon, cache, filesystem, lock, permission, worktree, or sandbox-boundary behaviour. Isolate Gradle daemon registries by a stable execution boundary, retain reuse only within that boundary, and clean up selected homes at the end of their lifecycle.

Upload a screenshot or short demo clip (PNG, JPEG, WebP, GIF, MP4, or WebM) to static.sebastiano.dev and get back a URL to hotlink in a GitHub PR description or comment. Use whenever a PR needs a before/after screenshot, UI diff image, or a short demo recording embedded or linked instead of committed to the repository.

Use when asked to create, search, edit, comment, link, or configure issues in JetBrains YouTrack via the yt command-line tool (package jetbrains-youtrack-cli, YouTrack REST API, issue tracker CLI). Handles global option placement, token precedence, JSON output, and exit code WORKFLOW errors.

Babysit a GitHub pull request after creation by continuously polling CI checks/workflow runs, new review comments, and mergeability state until the PR is ready to merge (or merged/closed). Diagnose failures, retry likely flaky failures up to 3 times, auto-fix and push branch-related issues when appropriate, and stop only when user help is required (e.g. CI infrastructure issues, exhausted flaky retries, or ambiguous/blocking review feedback). Use when the user asks to monitor a PR, watch CI, handle review comments, or keep an eye on failures and feedback on an open PR.