Development Harness

Works backward from the stated goal through its prerequisite chain — what must exist to reach the goal, then what each of those requires, back to the current state — classifying each prerequisite as AVAILABLE, DERIVABLE, or MISSING, and blocking planning until nothing is MISSING. Use before creating plans, delegating to agents, or defining acceptance criteria. Activates on specs, PRDs, tickets, RFCs, architecture designs, and multi-step engineering tasks. Always pass `#N` or an inline goal — invoking with no argument and no inline context returns BLOCKED. Integrates with CoVe-style planning pipelines.

Routes a Superpowers plan file through the /work-backlog-item pipeline and writes SAM task back-references and chunk annotations into the original plan. Use when given a path to a Superpowers plan file via $ARGUMENTS and needing to create a linked backlog item plus SAM task plan.

Orchestrate parallel agent dispatch as a manager — not a micromanager. Use when coordinating 2+ independent workers, running SAM task waves, relaying discoveries between worker waves, handling blockers, or synthesizing results. Covers both SAM structured dispatch (the task does the work) and ad-hoc dispatch (reference agent-orchestration for prompt template).

Use when Stage 1 Discovery is complete and design must begin — transforms the ARTIFACT:DISCOVERY into an actionable ARTIFACT:PLAN via RT-ICA prerequisite verification. Produces approach, components, success criteria, acceptance tests, and risks. Blocks on missing prerequisites before design proceeds.

Fetch and report current API syntax, changelog entries, and breaking changes for a specific library or protocol version. One research angle within a parallel technical-research set — runs independently and returns a structured cited report. Invoke when a specific library name and version are the target.

Use when starting a new feature, gathering requirements for an unfamiliar domain, refining a vague idea into actionable scope, or when a user request is ambiguous or underspecified. Conducts SAM Stage 1 discovery — structured requirements gathering through user discussion, asking WHO/WHAT/WHEN/WHY and never HOW. Produces the ARTIFACT:DISCOVERY document containing feature requirements, NFRs, goals, anti-goals, references, and resolved questions. Supports backlog item self-initialization via a

Executes SAM Stage 5 — dispatches a single ARTIFACT:TASK to a fresh stateless agent session, runs quality gates, and produces an ARTIFACT:EXECUTION with implementation results and verification output. Use when Stage 4 Task Decomposition is complete and tasks are ready for execution, when re-executing a task after Stage 6 returns NEEDS_WORK, or when dispatching a task to a language-appropriate specialist agent via the development harness pipeline.

Single-verb branch-to-PR quality gate pipeline. Use when the user wants to gate, push, and open a PR for a branch in one command.

Use when executing a SAM task — claims the task via MCP to set it IN PROGRESS, writes active-task context for hooks, loads task-level skills, implements against acceptance criteria, and marks complete via --complete flag. Triggers on task execution within the implement-feature loop or when an agent picks up a specific task from a plan file.

One-line definitions of development-harness (dh) plugin terminology — RT-ICA, ARL, SAM, the S1-S7 pipeline stage names, Impact Radius — each with a pointer to its canonical source file. Use when a dh skill, agent, or workflow step references a term without defining it, before guessing what the term means from its observed outputs, or when asked what a dh concept or acronym stands for.

Rigorous self-assessment checklist before marking any task as complete. Use when about to claim task completion, before final commit, when user asks "is it done?", or when transitioning from implementation to reporting. Prevents premature completion claims by requiring evidence for every assertion.

Development harness plugin documentation index. Use when looking up SAM pipeline, backlog lifecycle, SDLC layers, task/plan schema, plan artifacts, quality gates, or dispatch schema documentation.

Spawn and manage persistent interactive Claude Code CLI sessions with bidirectional communication via tmux. Provides spawn, send, read, status, list, and kill subcommands for orchestrating parallel peer sessions. Uses built-in --worktree and --tmux flags. Sessions stay alive for multi-turn steering. Triggers on "spawn claude session", "launch separate claude", "peer session", "inter-session communication", "shadow clone", "kage bunshin".

Generates one worker task prompt conforming to the CLEAR + selective CoVe task design standard and swarm-task-planner structure. Use when creating or rewriting a single task entry or task block inside a plan — providing a title and brief description as input.

Synthesis step in the multi-angle technical research pipeline. Receives structured outputs from all four research angles (api-state, ecosystem-research, impact-measurement, codebase-auditor), applies cross-angle signal weighting and conflict resolution, and produces a single synthesized Research section. Invoked by the technical-researcher orchestrator after all angle skills complete. Returns content to the orchestrator — does not write to the backlog.

Execute large-scale automated code transformations safely and idempotently. Use when renaming symbols across a codebase, migrating API call-sites, enforcing new patterns at scale, or applying structural edits to many files at once. Triggers on: 'codemod', 'mass rename', 'migrate all usages', 'transform codebase', 'apply pattern at scale', AST-based refactoring, or any task requiring consistent edits across 10 or more files.

Same reverse-thinking method as dh:rt-ica (see dh:rt-ica or dh-glossary for the definition), but localizes any MISSING input to the affected task only, as a non-blocking information-completeness pre-pass before task decomposition and plan generation. Use when grooming backlog items, generating plans, decomposing tasks under uncertainty, or working in brownfield and refactor scenarios. Produces completeness summary (APPROVED-FOR-PLANNING, APPROVED-WITH-GAPS, or BLOCKED-FOR-PLANNING), missing input report with dependency mapping, required unblock actions, and planning annotations for downstream tasks. Use dh:rt-ica instead at the S2 implementation gate where missing inputs must halt execution.

Executes a groomed milestone with parallel kage-bunshin sessions in isolated worktrees. Use when a milestone has been groomed and /groom-milestone has produced a dispatch plan. Reads the dispatch plan, creates an integration branch, spawns one kage-bunshin (independent claude -p process) per wave item in its own worktree — each session is a full orchestrator with the Agent tool. Sequentially merges worktree branches, relays wave discoveries to subsequent waves, then lands the integration branch to main. Takes a milestone number as argument.

SAM-style feature initiation workflow — discovery through codebase analysis, architecture spec, task decomposition, validation, and context manifest. Use when a user asks to add a feature, plan a feature, or convert an idea into an executable SAM plan.

Reviews CLI application code for correctness and quality. Use when reviewing tools that use argparse, click, typer, commander.js, or similar argument parsers — covers exit codes, help flags, stdin/stdout/stderr separation, non-interactive operation, signal handling, argument validation, ANSI color safety, and dry-run support for destructive operations.

Use when reviewing AI/ML code or LLM integration — activates on prompt templates, model selection logic, token budget concerns, or evaluation harness code. Enforces prompt hygiene, model tier matching, context window management, token economics, structured output validation, temperature settings, retry logic, streaming error handling, and PII/safety rules.

Use when reviewing web frontend code — HTML, CSS, JSX, or browser-targeted JavaScript. Enforces accessibility (WCAG AA, aria labels, focus management), XSS prevention (innerHTML, dangerouslySetInnerHTML), performance (layout thrash, CLS, lazy loading), CSS design tokens, form labeling, and event listener cleanup. Loaded by dh:code-reviewer on *.html, *.css, *.jsx detection.

Register a plan artifact via the MCP backlog server. Use when you produce a document or report that downstream agents or worktree-isolated environments need to retrieve — feature-context, codebase-analysis, architect, T0-baseline, TN-verification, or research artifacts. Triggers include "store an artifact", "register a plan artifact", "write a report to the backlog", "upload artifact content".

Use when SAM Stage 5 Execution has completed and task results need independent verification against acceptance criteria. Dispatches a separate reviewer agent to fact-check implementation outputs and returns COMPLETE or NEEDS_WORK with specific findings and remediation tasks.

Grooms a GitHub milestone for parallel execution — batch-grooms ungroomed items, assesses scope gaps, analyzes cross-item dependencies via Impact Radius overlap, builds conflict groups, assigns items to execution waves, and persists the dispatch plan via dispatch_create_plan MCP tool. Calls dispatch_wave_start per wave to register state. Use when preparing a milestone for /work-milestone execution. Pass the milestone number as the first argument. Requires milestone items assigned via /group-items-to-milestone.

Local codebase analysis research angle — derives behavioral contracts, coding conventions, SKILL.md flow insertion points, and agent data availability maps from actual source files. Use when the blocking question is answered by reading the repository: what does this function actually do, what pattern does the codebase use for X, where in this workflow does a new step go, or what data does this agent already have.

Use when the plan_ref returned by add-new-feature is provided. Executes the SAM implementation loop — dispatches ready tasks to specialist agents in parallel, manages bookend tasks (T0 baseline capture and TN verification), tracks concerns and contract violations per task, and relies on hooks to update task status. Manages task batches via sam_plan and sam_task MCP tools.

Use when creating, grooming, planning, or closing a backlog item. Bridges backlog items to SAM planning with issue, project, and milestone tracking against the configured backend. Activates on interactive browsing with no arguments, loading an item by issue reference or title match to run grooming and SAM planning, autonomous unattended runs that substitute evidence-derived decisions for clarifying questions, a quick path for one-file fixes where full grooming is disproportionate, dismissing an item without completion under a required reason (duplicate, out of scope, superseded, wontfix, blocked), marking an item done with an evidence trail and required summary, initializing GitHub tracking infrastructure for a project, or selecting a language/stack profile for planning. Stops when the item already has a plan or is blocked on missing information.

Applies Node.js-specific code review patterns for async I/O, streams, security, process management, and dependency hygiene. Use when reviewing Node.js server code, route handlers, middleware, or any JavaScript file alongside package.json without TypeScript. Triggers on sync I/O in request paths, missing stream backpressure, process.exit misuse, eval/exec injection risks, wildcard version ranges, missing lockfiles, EventEmitter cleanup gaps, and unvalidated environment variables at startup.

Provides Python-specific code review rules for the dh code-reviewer agent. Activates on pyproject.toml or *.py file detection — enforces uv, ruff, ty, pytest, type annotation, error handling, and Python 3.11+ idioms including pathlib, match statements, and modern union syntax.

Research community usage patterns, real-world gotchas, and client compatibility for a specific known library, tool, or protocol feature. Use when a technical-researcher orchestrator needs community-sourced evidence about a named library or feature — bug reports, workarounds, compatibility gaps, and patterns from issue trackers and discussions. Distinct from the broad ecosystem-researcher agent — this skill targets a KNOWN entity and mines what real users have actually experienced, not what exists in a domain.

Certifies that a feature achieves its original objectives via goal-backward verification (SAM Stage 7). Use when all tasks pass forensic review — starts from expected outcomes, works backwards to verify each was achieved, and returns CERTIFIED or NOT_CERTIFIED with specific gaps.

Quantitative cost measurement for technical research — token injection costs, payload sizes, context window consumption, and file-level counts from actual repo files. Use when a technical-researcher orchestrator needs the cost dimension of adding or changing something: how many tokens will it inject, how big are the artifacts, what is the context window impact? This is NOT blast-radius analysis (which files break) — it covers size, tokens, and performance cost only.

Decomposes a contextualized plan into atomic, independently executable tasks with complete embedded context, registered through the plan API. Use after SAM Stage 3 Context Integration produces the contextualized plan artifact — when the plan is ready for task generation with CLEAR ordering, CoVe checks, and dependency graphs for parallel execution.

Use when the PLAN artifact from SAM Stage 2 needs contextualization against actual codebase state — grounds the design plan in reality by performing scope analysis (NEW/MODIFY/COMPLETE classification), conflict detection between plan assumptions and codebase patterns, and resource mapping to concrete file paths and integration points. Produces an updated ARTIFACT:PLAN registered via MCP with a Contextualization section appended.

Routes to the correct dh skill entry point by intent. Use when unsure which development-harness skill to invoke, starting a development workflow, or invoking /dh directly. Covers capture, groom, plan, execute, single task, quality gates, and milestone routing.

Scientific validation protocol for verifying fixes work through observation, not assumption. Use when claiming a bug fix, code change, refactoring, or implementation is complete — enforces reproduce-broken-state then define-success-criteria then apply-fix then verify-outcome. Success means observing intended behavior, not absence of errors.

Use when creating or updating the project skill discovery config — generates or regenerates .dh/skill_discovery.yaml by scanning the repo to infer tech stack, inventorying installed skills via npx skills list, loading candidate skill content before suggesting, and writing a config-driven skill injection file. Triggers on /dh:setup-skill-discovery invocations and programmatic --auto calls from add-new-feature Phase 3.

Use when orchestration or planning agents are producing task plans, task prompts, or TASK.md instructions that must be unambiguous, verifiable, and resistant to hallucination. Applies CLEAR (Concise, Logical, Explicit, Adaptive, Reflective) to structure and write agent task prompts, then adds CoVe (Chain of Verification) checks where accuracy risk is meaningful. Activates on draft task prompts, swarm plans, migration tasks, and multi-step plans requiring independently executable steps.

Provides TypeScript-specific code review patterns covering strict mode, ESM, type safety, branded types, discriminated unions, async patterns, runtime safety, and common anti-patterns. Activates on detection of tsconfig.json, *.ts, or *.tsx files during code review — loaded automatically by dh:code-reviewer.

Manages feature implementation task state via SAM MCP tools. Use when querying task status, listing ready tasks, claiming tasks for execution, updating task timestamps, or coordinating multi-task feature rollout. Activated by the /dh:execution orchestrator to track progress — also activates directly when managing tasks or configuring hook profiles.

Use when all tasks for a feature are marked COMPLETE — runs holistic quality gates including code review, feature verification, integration check, documentation drift audit and update, and context refinement. Creates follow-up plans when issues are found.

Use when a task asks for architecture review, dependency graph visualization, module coupling analysis, or circular dependency detection. Auto-detects scope (git diff → PR diff → full project). Reads project config (pyproject.toml, tsconfig.json, go.mod, Cargo.toml) to establish the intra-project module namespace before parsing imports. Builds a module-dependency graph across Python, TypeScript, JavaScript, Go, Rust, and Java. Detects cycles via graphify output or an executable Python script. Checks Conway's Law alignment against CODEOWNERS and directory structure. For Claude plugin repos, also traces cross-language chains: hook configs → hook scripts, SKILL.md/agent docs → node/uv-run scripts, PEP 723 inline deps, and MCP tool calls. Emits Mermaid flowcharts with severity color-coding (red = circular dep, yellow = high-coupling, green = clean, blue = Conway violation). Applies recursive semantic partitioning for graphs > 40 nodes. Registers each diagram as a codebase-analysis artifact.

Use when a diff needs four parallel perspective reviewers (Security, Performance, Quality, Accessibility). Creates an ephemeral SAM plan, collects structured verdicts, synthesizes them into one deduplicated cross-referenced punch list, prints one summary line per perspective, and exits non-zero if any perspective returns REJECT. SKIP is a passing outcome.

Loaded automatically when reviewing Claude skills or agent definitions — covers SKILL.md structure, frontmatter validity, token budget, description quality, and agent contract compliance.