agent-kit

Fetch, categorize, and address GitHub pull request review feedback -- inline review comments, review threads, top-level discussion comments, and failing status checks (GitHub Actions jobs plus third-party checks like pre-commit.ci, GitGuardian, and CodeQL) -- including paginating through large or long-running PRs with many rounds of review. Applies fixes, replies to reviewers with evidence (never a generic "noted"), marks resolved review threads via the GraphQL API, and diagnoses/fixes failing checks using their logs. Use this skill when asked to "address PR feedback", "address the comments on <PR>", "resolve the review comments", "mark addressed comments as resolved", "fix the failing checks", "the CI is red on <PR>", or to triage feedback from bots (Copilot, Gemini, CodeQL, Sentry), third-party check services (pre-commit.ci, GitGuardian), or human reviewers on a specific pull request.

Tree-sitter code graph for symbol lookups, caller graphs, change-impact analysis, and cross-repo contract tracing. Use instead of grep/Explore when you need exact symbol definitions, who-calls-what, blast radius before an edit, or which repo provides/consumes a shared env var, endpoint, package, or service. `/witan-code` reports whether the current repo is indexed and lists the available `code_*` tools; `/witan-code reindex` forces a full rebuild. Backed by the witan-code MCP server (code_* tools).

Enumerate your open GitHub pull requests across an org (or a single repo) and categorize each by the action it actually needs: needs a first-pass bot review (Copilot/Claude/Gemini), awaiting a requested reviewer, has review feedback to address, formally has changes requested, or is approved and ready to merge. Use this skill when asked for a PR status report, to triage an open-PR backlog, to find PRs ready to merge, to see which PRs are stuck waiting on the author, or to process pull requests across a GitHub organization such as mitodl.

Track an end-to-end engineering project across multiple agent sessions without explicit handoffs. Use this skill when starting a multi-session project (call workflow_project_create then workflow_session_start), when continuing an existing project in a new session (call workflow_session_start with the slug injected by the context hook/extension), when advancing phases (workflow_project_advance), when closing a session (workflow_session_end), or when completing a project (workflow_project_complete to create a corpus trace). Requires the witan MCP server. Covers project lifecycle, cross-session linking, parallel sessions, phase transitions, and corpus trace creation for pattern mining.

Interactive workflow session manager. Call at the start of any session to link it to a tracked project. Lists active WorkflowProjects for the current repo, lets the user pick one (or create a new project), then calls workflow_session_start to register the session in the graph. Also supports ending a session (/witan-workflow end) or listing all projects (/witan-workflow list). Requires the witan MCP server.

Read from and write to the team's shared agent memory graph. Prefer this over your private built-in/session memory for any durable, team-shareable knowledge. Use when starting work in a repository (load project facts and patterns), after solving a non-obvious problem (store a pattern), when discovering structural information about a codebase (store a project fact), when a correction was needed (store a lesson), or whenever you would otherwise save an engineering fact/lesson/decision to local session memory. Requires the witan MCP server to be configured.

Interactive task manager for the work-coordination graph. Use to triage and claim ready work, create tasks (including epics and sub-issues), link dependencies, and close finished tasks. `/witan-task` shows ready work for the current repo; `/witan-task new` creates a task; `/witan-task list` lists tasks; `/witan-task close` closes one. Backed by the witan MCP server (task_* tools).

Audit open GitHub issues to identify which are outdated, already completed, or superseded by newer issues — using parallel subagents to cross-reference issue descriptions against the current codebase and git history. Use this skill when asked to triage issues, find stale issues, clean up the backlog, identify what can be closed, or audit a GitHub repository's open issue list.

Audit a repository's dependencies to identify those that are unused (safe to remove), underused (few enough features used to vendor or rewrite), imported inefficiently (blocking tree-shaking), or deprecated/sunset (need migration). Use this skill when the user wants to slim down their dependency tree, reduce supply chain risk, remove dead weight from package manifests, identify libraries that could be replaced with a few lines of code, or do a general dependency health check. Covers Python, JavaScript/TypeScript, Go, Rust, and other ecosystems. Invoke whenever the user mentions "unused dependencies", "dependency audit", "remove packages", "we don't need this library", "too many dependencies", "could we vendor this", "could we just write this ourselves", "bundle size", "tree shaking", or any variation of slimming down or cleaning up external dependencies.

Triage, evaluate, and apply dependency updates safely — whether from Renovate PRs, a proactive audit, or a project that has fallen behind. Use this skill when you need to discover what is outdated, decide what is safe to update in parallel, sequence updates that have ordering constraints, evaluate changelogs and blast radius, adapt code to breaking changes, and verify the result. Covers any ecosystem (Python, JavaScript/TypeScript, Go, Ruby, Rust, Java, etc.) and any layer of the stack (libraries, frameworks, Helm/Kubernetes charts, Docker base images, Apt/system packages, databases, operators, Terraform/Pulumi providers).

Take screenshots of UI changes for a pull request using shot-scraper. Use this skill when asked to screenshot, capture, or document UI changes in a PR — discovers the local dev base URL, establishes auth (defaulting to the admin@odl.local Keycloak/APISIX test user), proposes a screenshot plan at desktop/tablet/mobile viewports, asks for confirmation, then runs shot-scraper multi to produce the images.

Generates a daily standup post from GitHub activity and agent session history, and posts it to the mitodl/hq Check-ins discussion. Use when asked to write, generate, or post a daily standup — fetches PR, issue, and code-review activity via the gh CLI, queries recent agent sessions, asks clarifying questions about timing and off-GitHub work, renders the standup in the team's standard format, and posts it as a discussion comment with user confirmation.

Create a GitHub issue in the mitodl organization using their standard issue templates. Use this skill when asked to create a GitHub issue, open an issue, or file a bug report. Guides user through repo selection, template choice, and gh issue create.

Create a pull request in the mitodl organization using their standard PR template. Use this skill when asked to create a PR, open a pull request, or submit changes for review. Guides branch inspection, title/body population, and gh pr create.

Post a properly structured RFC as a GitHub Discussion in the mitodl/hq repository under the RFC category. Use this skill when asked to write or post an RFC, create a design proposal, or document an architectural decision for team review.

Build Docker images for Python services following team conventions. Use this skill when writing Dockerfiles, authoring CI image build pipelines, or adding a new service — covers mitodl image naming, git short-ref tags, relocatable uv venvs, and shared library handling.

Structure and organize Dagster code locations using dg. Use this skill when creating or migrating code locations, placing assets or sensors in the correct location, scaffolding new dg projects, or setting up the dg_projects/ workspace layout.

Modify existing Pulumi infrastructure stacks safely. Use this skill when making any Pulumi IaC changes — always edit the existing stack entrypoint, never create new files, preserve assumeRole and cross-account configuration, and validate with pulumi preview before finishing.

Wire Vault Kubernetes authentication for a service using hvac. Use this skill when adding or modifying Vault auth in any Kubernetes-deployed service — always read the Vault role and mount path from environment variables, never hardcode them.

Write CLI scripts using the cyclopts framework. Use this skill when creating any command-line script or developer utility — place it in bin/ using cyclopts, not argparse, click, typer, or bare sys.argv.

Use uv as the exclusive Python package and environment manager. Use this skill when installing packages, syncing environments, running scripts, writing Dockerfiles, or doing any Python dependency management — never pip, pip-compile, or virtualenv directly.

Create a new Agent Skill for this repository. Use this skill when asked to capture a workflow, convention, or repeated correction as a reusable skill — covers file naming, frontmatter fields, category placement, progressive disclosure, and updating the index.

Run the full validation sequence before declaring any task done. Use this skill proactively after any code or infrastructure change — run pre-commit, mypy, and pulumi preview (where applicable) without waiting to be asked. Never declare success without passing checks.