pipecrew

Launch the machine-wide PipeCrew fleet dashboard — a live view of EVERY Claude Code session on this machine (not just the current /deliver run), each card showing token usage, sub-agents, and a "needs approval" badge. PipeCrew /deliver sessions render with the same pharaoh crew icons as /site-view. Backed by the standalone `pipecrew-siteview` tool (npm); this skill just finds and starts it. Use to watch several sessions/runs at once, or a plain Claude Code session with no PipeCrew run.

PipeCrew delivers a feature end-to-end across all workspace repos. Orchestrates: requirements → architecture → spec editing → parallel implementation (backend + frontend + mock + infra) → cross-repo assessment. Usage: /deliver <feature description>. Reads workspace config from {workspace_root}/{workspace}/config.json.

PipeCrew site-view — a live browser UI of the construction site where the agent crew works. Opens at http://127.0.0.1:5173 showing characters as queued / working / done, pyramid tiers rising as agents complete, fed by the current run's scratchpad + checkpoints + awaiting-input flag over Server-Sent Events. Auto-started at Pre-flight by the /deliver skill. Full reference at docs/site-view.md.

Learn from user feedback — from a merged PR, a recorded /deliver run, a branch diff, a Claude Code session transcript, or free-form text — and propose scoped updates to the workspace's durable context docs (platform.md § Established Patterns, repo CLAUDE.md / DESIGN_SYSTEM.md / agent-context/). Presents findings tier-classified (repo / workspace / plugin-level) with before/after diffs; user approves per finding; approved changes are applied. After the docs are saved the user can optionally dispatch the per-repo implementer agents to apply the same findings to an existing branch as a fix round (so the branch the feedback came from gets brought in line with the new conventions, not just future work). Every run is logged to history/learn-log.md for institutional memory.

Discover a new project for the PipeCrew. Inspects repos, interrogates the domain, generates workspace config + platform.md + CLAUDE.md files + domain-specific agents. Run once per project, then use /deliver to ship features.

Standalone cross-repo assessment. Checks a feature branch across all repos for cross-repo integration: wire-shape agreement, requirement enforcement symmetry, event/infra wiring. Use for verifying manually-implemented features or pre-merge validation.

Audit or refresh PipeCrew context docs at three scopes: a single repo (its agent-context/, CLAUDE.md, and DESIGN_SYSTEM.md if frontend), the workspace (platform.md), or everything (workspace + every repo). Audit mode reports staleness only; refresh mode updates docs to match current code.

Manage a workspace's GitHub-backed shared memory directly — without running a full pipeline. Subcommands: status (freshness / ahead-behind / unpushed), pull (fetch the team's latest now), sync (redact + commit + publish any pending durable-doc changes / retry a failed push), enable (turn memory on for an already-onboarded workspace and bootstrap its private repo). The automatic syncs in /discover, /learn, /context-refresh, and /deliver still run on their own; this skill is the manual + operational control surface over scripts/sync-memory.js. Named memory-sync (not memory) to avoid Claude Code's built-in /memory command.

Cross-repo incident triage. Takes a one-line symptom + optional flags, runs a hypothesis-driven investigation against the workspace's logs and recent diffs (using the OBSERVABILITY routing table from platform.md), and writes a structured report. READ-ONLY by hard rule — the troubleshooter agent's system prompt forbids mutating commands, and an optional `PreToolUse` hook backed by `scripts/troubleshooter-bash-guard.js` enforces the same allowlist at tool-dispatch time.

Standalone code review. Diffs a feature branch against main in one repo, dispatches the appropriate tech-stack reviewer agent, and returns a structured report with findings. Use for PR review or reviewing code not produced by the /deliver pipeline.

Lightweight, memory-backed change skill. Applies small, well-specified, recurring changes (audit-finding fixes, codemods, convention enforcement, mechanical migrations) using reusable fix RECIPES instead of the full /deliver pipeline. No product-owner, no architect, no per-repo paired reviewers — design and review criteria come from recipes + ADRs + per-repo checklists. One approval gate. Read-only verification. Use when the WHAT is already known and the blast radius is small; bounce to /deliver when the change needs requirements, UX, or cross-repo contract design.

Generate or refresh architecture diagrams for a workspace. Default mode regenerates the two canonical Mermaid files (architecture-overview.mmd + architecture.mmd) for an onboarded workspace by re-dispatching the solution-architect in discovery mode. Optional --topic flag produces a focused diagram on a specific concern (auth flow, event flow, deploy topology, etc.) without touching the canonical files. Reuses the same conventions documented in rules/discovery-diagrams.md that /discover uses. Standalone — does not run the rest of /discover.

Kill stale PipeCrew site-view servers. Discovers every running site-view via the same probe as /sites, then kills their PIDs. Supports --keep-port, --keep-run, --keep-latest, and --dry-run to preview before killing. Defaults to --dry-run for safety.

List every PipeCrew site-view server currently running on localhost. Probes ports 5173-5195 via the /state endpoint and reports port, PID, workspace, run-id, character count, and whether any server is awaiting user input. Use when you've started multiple runs and need to see what's alive, or before killing a stale server.

Generate the demo workspace at {workspace_root}/simulate-run-demo/ with full /discover + /deliver + /learn artifacts following the latest plugin schema (Phase 8 PR publish, pr_urls.json, learn runs, stacks docs, architecture diagrams). Each invocation wipes and recreates the workspace in place — fixed run IDs are reused, so no timestamped run dirs accumulate. Spawns the site-view pointed at the demo by default; pass --no-ui for headless generation. Zero agent cost.

Scaffold new repos from scratch or from an example. Two modes: --from-scratch uses a PROJECT_BRIEF to create repo skeletons with git init + baseline CLAUDE.md; --from-example clones structure from a reference repo. Standalone skill — can be called directly or by /discover --greenfield.