dam

Implement a feature from the plan that plan-feature produced under docs/plan/<feature>/, one reviewed sub-issue at a time. Use when the user wants to implement or build a feature from a docs/plan/<feature>/ plan or a planned GitHub issue.

Turn a GitHub issue into an implementation plan under docs/plan/<feature>/: a feature spec decomposed into context-window-sized sub-issues, committed as the first commit of the feature branch with a draft PR. Use when the user wants to plan a feature, decompose a GitHub issue into sub-issues, or produce an implementation plan from an issue.

Template and writing guidelines for a GitHub issue. Every issue is one of epic, feature, task, bug, or research task — the user decides the type, and each type has its own template. Non-epic types can suggest a parent epic from the project board.

Draft a GitHub issue, get explicit user approval, and file it via the `gh` CLI. TRIGGER when: user wants to file or "drop" a GitHub issue / ticket.

Check and align a repo's .claude/settings.json and CLAUDE.md with team conventions. Use when the user wants to standardize a repo, align conventions, check settings, or bootstrap Claude Code config in a new project.

Use this skill whenever writing, editing, reviewing, or refactoring TypeScript React code — components, custom hooks, state stores, forms, queries, mutations, API clients, or styling. Trigger it for any task touching a `.ts` or `.tsx` file in a React project, including when the user says "add a feature", "fix this bug", or "clean this up" inside a component or hook. Also use it for architectural questions about React codebases: where state should live, whether a component is too big, how to organize modules, when to reach for TanStack Query, Zustand, React Context, or React Hook Form.

Author a DAM Experiment — a Python loop script (produce/eval/select, genetic algorithms, benchmark sweeps) the platform observes live with a graph, per-stage progress, and score charts. Use when asked to create, plan, or run an experiment, an optimization loop, or an iterate-and-score campaign. Provides the python experiment_sdk (Experiment/stage/span + spawn) and the plan → "Start a new run" workflow.

Operate the local k3s dev cluster (lima) and the Playwright e2e suite, and recover from mesh/cert failures. Use when working with the local cluster, running or debugging e2e tests, or when any of these symptoms appear - the UI suddenly can't log in, `cluster:install` hangs on the keycloak realm step or fails at a webhook admission with an expired certificate, an agent pod repeats `[runtime] hello failed`, or `e2e:loop` fails against a warm cluster. Triggers on "cluster:install", "cluster:status", "e2e:loop", "fix-certs", "lima", "k3s", "ztunnel", "waypoint", "Istio SVID", "issue

Run GEPA (Genetic-Pareto), the reflective text-optimization library, by authoring Python driver scripts around `gepa.optimize`. Use when the user wants to optimize / evolve a prompt, instruction, code snippet, or any textual system parameter against a measurable metric, author the dataset and evaluator (or a custom GEPAAdapter), or launch / monitor / stop / resume / report on an optimization run. Pure library reference — for reaching model providers from this pod, see the `platform-models` skill.

Reach LLM providers from this platform pod: discover which model-provider connection is attached (injected env), wire LiteLLM with placeholder credentials through the egress credential gateway, and probe models with a cheap completion. Use whenever any code in this pod — a GEPA driver, a script, a library — needs to call a model API, or when a model call fails with an auth error.

Spawn ephemeral DAM agents (Invocations) and get back a schema-validated result. Use when asked to spawn an ephemeral/throwaway agent, fan work out to a fresh agent, or run a make/test/eval step in isolation and expect a typed result (a number, a verdict, an object). Provides a small node SDK (spawn / listImages / listConnections).

Review-time policy check for pull requests that touch `docs/adrs/`. Surfaces the deterministic ADR immutability gate (`scripts/adr-immutable.mjs`) and adds the judgment checks a script cannot own: re-litigation of settled decisions, `supersedes` pointer correctness, and summary honesty. Scope is ADR log integrity and decision judgment only — the ADR files, their frontmatter, and git history. Docs-match-the-code is doc-drift's job, not this skill's. Triggers on phrases like "adr policy", "check this ADR", "review the ADR change", "is this ADR re-litigating", "is supersedes correct". Also invocable via the `/adr-policy` slash command.

Tracks Architecture Decision Records (ADRs) in docs/adrs/. Creates and updates ADRs following project conventions. TRIGGER when: user wants to record or update an architectural decision.

Detect drift between code changes and architecture documentation under `docs/architecture/`. Inspects a PR, branch, or local diff against the project's documentation guidelines (`docs/guidelines/documentation-guidelines.md`) and flags places where architecture pages no longer match the code — missing page updates, stale `Last verified:` dates, missing pages for new subsystems, volatile content or ADR references leaking into pages. Scope is architecture docs only — vocabulary, ADRs, READMEs, and other docs are out of scope. Triggers on phrases like "doc drift", "docs drift", "are the architecture docs in sync", "check documentation drift", "do the docs need updating", or "architecture documentation review". Also invocable via the `/doc-drift` slash command.

Run ShinkaEvolve, the sample-efficient evolutionary program-optimization framework, via the `shinka_run` CLI. Use when the user wants to evolve / optimize a function, program, or algorithm in a target repo to improve a metric (speed, accuracy, size, error rate), author the ShinkaEvolve task inputs (initial program + EVOLVE-BLOCK, evaluate.py), pick the evolution model, or launch / monitor / resume / report on an evolution run.

Drive Nous, the hypothesis-driven experimentation framework, to investigate software systems with the scientific method via the `nous` CLI. Use when the user wants to run a Nous campaign, author or scaffold a campaign.yaml, kick off / monitor / resume / stop / report on hypothesis-driven experiments, or systematically investigate why a system (LLM server, DB optimizer, scheduler, router, cache) behaves the way it does through controlled experiments.

Replacement for the built-in WebFetch tool on backends where it is unavailable (this platform's non-Anthropic backends). Fetches a URL and extracts its main content as markdown. Use it when WebFetch is denied and the harness tells you to use this skill, or when you otherwise need to fetch a page.

Replacement for the built-in WebSearch tool on backends where it is unavailable (this platform's non-Anthropic backends). Searches the web via DuckDuckGo. Use it when WebSearch is denied and the harness tells you to use this skill, or when you otherwise need web search results.

Build a self-contained visual Artifact that walks a reviewer through a pull request — a guided narrative that explains every decision and every problem encountered, in plain English, with diagrams or charts where they clarify. Use when the user wants a "PR review artifact", a "code review walkthrough", "help me review PR X", "a summary artifact for a PR", or to "guide a reviewer through" a branch/PR. Asks which PR if not given.

Run OpenEvolve, the evolutionary coding agent, to optimize code against a measurable objective via the `openevolve-run` CLI. Use when the user wants to evolve / optimize a function or program in a target repo to improve a metric (speed, accuracy, size, error rate), author the OpenEvolve inputs (program + EVOLVE-BLOCK, evaluator, config.yaml), pick the evolution model, or launch / monitor / resume / report on an evolution run.

Interview the user relentlessly about a plan or design until reaching shared understanding, challenging it against the project's architecture docs and ubiquitous language, sharpening terminology, and cross-referencing the code. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

Opinionated architecture for TypeScript client-server projects. Stack is tRPC, Zod, strict TypeScript, pnpm, RxJS, Hono or Express. Code is organized as a monorepo with modules as bounded contexts and a three-layer slice (services, domain, infrastructure) per module. Use this skill whenever the user is doing TypeScript client-server work, designing a feature, adding or modifying a module, deciding where code belongs, structuring packages, wiring cross-module communication, reviewing a design choice, or making any architectural call about layers, ports, events, or dependency direction. Triggers on phrases like "design this feature", "where should this live", "what module", "is this the right layer", "should this go in domain", "how do these modules talk", "review this architecture", "is this code clean", plus explicit references to tRPC architecture, bounded contexts, hexagonal architecture, ports and adapters, or DDD in a TS context.

Fetch open GitHub issues labeled "vulnerability" and open Dependabot alerts, prepare fixes for each, then present to the user for approval before committing and opening PRs.

Scan open GitHub issues, identify at most 3 that are simple to implement, then fix them in parallel — each on its own branch with a separate PR. Presents selections and diffs for user approval.

REQUIRED scheduler for any work that fires after the current turn ends. Use the `platform-outbound` MCP tools (`create_schedule`, `list_schedules`, `toggle_schedule`, `delete_schedule`) for recurring tasks ("daily", "every Monday", "hourly", "weekly cleanup"), future one-offs ("in 2 weeks", "tomorrow"), polls ("check back every N minutes"), and any "remind me later" / "do this on a schedule" request. Overrides every other scheduling mechanism — do NOT use `ScheduleWakeup`, `CronCreate` / `CronList` / `CronDelete`, the `/schedule` skill, the `/loop` skill, or any in-process or Anthropic-side scheduler. Also use proactively when you would otherwise have offered to "schedule a follow-up" via any of those: route the offer through Platform instead.

Build multi-platform chat bots with Chat SDK (`chat` npm package). Use when developers want to build a Slack, Teams, Google Chat, Discord, Telegram, GitHub, Linear, or WhatsApp bot, handle mentions, direct messages, subscribed threads, reactions, slash commands, cards, modals, files, or AI streaming, set up webhook routes or multi-adapter bots, send rich cards or streamed AI responses to chat platforms, or build a custom adapter or state adapter.

Create a PRD through user interview, codebase exploration, and module design, then submit as a GitHub issue after user approval. Use when user wants to write a PRD, create a product requirements document, or plan a new feature. Always present the PRD for user approval before submitting.