devops

Codebase knowledge graph via the external graphify CLI — default-on, opt-out via `{"consent":false}` in `.claude/graphify.json` or `~/.claude/graphify.json`. Auto-installs graphify, keeps the graph fresh windowlessly, answers codebase questions with `graphify query`, and hard-gates broad raw-file searches toward the graph. Triggers: "knowledge graph", "graphify", "code graph", "/auto-graph". Do NOT trigger for simple single-file lookups.

Manually update the devops plugin to latest from GitHub. Delegates to ss.plugin.update hook (pull + cache + registry), then adds changelog and verification report. Triggers on: "update plugin", "plugin updaten", "self update", "devops update", "neue version". Explicit user request only.

Fetch live token usage for completion card battery line. Reads the native statusLine-written usage file first (no scrape), falls back to the Edge CDP scraper or cached data. Run silently pre-card, or manually: "refresh usage", "wie viel hab ich verbraucht", "token budget".

Collect mode — a UserPromptSubmit hook parks each prompt in `.claude/batch.md` instead of executing it (no model turn), until an execute marker merges the whole set into ONE feasibility-checked plan. Triggers on "/claude-batch", "sammelmodus", "collect mode", "batch mode", "erstmal sammeln", "nicht sofort umsetzen". Do NOT trigger for normal work, for backlog execution (/run-backlog), or for issue creation (/setup-issue).

Interactively scaffold or adapt a project-level extension for any plugin skill. Lists available skills, checks for existing extensions, and creates or opens the correct files. Triggers on: "extend skill", "customize skill", "skill extension".

Capture a long-term learning/correction and route it to the correct project- specific instructions (skill, skill-extension, deep-knowledge, scheduled-task runbook, or as a last resort CLAUDE.md) — NOT to personal feedback memory. Also prunes now-duplicate `feedback_*.md` entries with confirmation. Routing is one decision matrix (Step 2); per-branch execution lives in `deep-knowledge/routing-details.md`. Triggers ONLY on explicit invocation: "/claude-learn", "devops learn" (also misspelt — "devos learn", "/devops-learn" — and also when it sits at the END of a task prompt: the task runs AND the capture runs, in the same turn, never dropped), "lerne das", "merk dir das fürs Projekt", "remember this for the project", "capture learning". Do NOT trigger for one-off conversational corrections or for personal feedback memory.

Strict mode — the deliverable is exactly what the prompt names, nothing wider; attributes the prompt leaves open are Claude's call and every such choice is reported. Propagates to every skill, agent, /concept iteration and autonomous resume the turn starts. `/claude-strict <task>` for one prompt, `/claude-strict on|off` for the current worktree + branch — never project-wide. Triggers on: "/claude-strict", "strict", "strikt", "genau so und nicht mehr", "nur das ändern", "nichts anderes anfassen". Do NOT trigger for: ordinary requests without a literal-scope signal, /tune-harden or /tune-polish scope fences (they stay as they are), or TypeScript `strict` compiler options.

Generate an interactive HTML page for analysis, plans, concepts, prototypes, comparisons, or creative work — open it in the browser and monitor user decisions (toggles, selections, comments) to feed them back into the workflow. Triggers on: "concept", "concept page", "interactive plan", "show me this as a page", "visualize this". Also auto-suggest when Claude completes analysis, planning, comparison, or concept work that would benefit from interactive decision-making. Do NOT trigger for: simple code explanations, debugging (use /fix), or static documentation (use /setup-readme).

Read session logs, runtime errors, and crash output to diagnose and fix the current issue — root-cause analysis first. Use when something is broken or behaving unexpectedly. Triggers on: "debug", "this is broken", "doesn't work", "error", "crash", "blank screen", "warum geht das nicht", "funktioniert nicht". Also triggers on pasted error/stack traces or when post.flow.debug hook detects repeated Bash failures. Do NOT trigger for general code questions.

Deliberate channel promotion for the ring model: move an already-shipped version alpha→beta→stable (or fast-track alpha→stable) by re-tagging the SAME commit via the ship_promote MCP tool. Never rebuilds, never bumps versions, never runs autonomously — promotion is always a user decision. Triggers on: "release", "promote", "promotion", "channel release", "auf stable heben", "promote to beta", "promote to stable". Do NOT trigger for shipping new work (use /ship) or plugin updates (/auto-update).

Full-ceremony multi-agent orchestration: plan → user confirm → execution mode → waves → inter-wave gates → synthesis. The EXPLICIT path for the Full-ceremony tier of the always-on delegation policy (3+ domains, a feature end-to-end, high-risk change); the lower tiers never go through this skill. Triggers on: "run agents", "use agents", "orchestrate", "parallel agents", "multi-agent", "delegate to agents", "agent workflow", or the user saying yes to Claude's offer. Do NOT trigger for: simple edits, quick fixes, explanations, or single-agent research.

Fully autonomous agent orchestration for when the user is away from the PC. Runs agents without supervision (implementation, desktop interaction, live browser/app tests) and can OPTIONALLY SHUT DOWN THE PC after completion. Triggers: "autonomous", "run autonomous", "run this while I'm away", "afk mode", "autopilot". Do NOT trigger when the user stays present — use /run-agents instead.

Milestone-centric backlog runner: picks open GitHub milestones (or loose issues), then refines, implements, tests and ships each item unsupervised while the user is away, closing issues and milestones as they land. Use when the user wants the planned backlog worked off without babysitting, or "let old ideas and fixes get worked off while I'm gone". Triggers: "backlog abarbeiten", "arbeite den backlog ab", "backlog runner", "run the backlog", "milestones abarbeiten", "arbeite die milestones ab", "arbeite den milestone ab". Do NOT trigger for maximizing token budget (that is /run-burn, explicit-only), nor for a generic single ad-hoc AFK task (that is /run-autonomous).

Autonomous task runner that consumes the remaining weekly budget before it resets — as depth per task (better models, higher effort, extra review passes) first, added parallel lanes second — and lands every task individually so a mid-run limit costs one task, not the whole run. EXPLICIT INVOCATION ONLY — the user must type /run-burn to activate. Do NOT trigger on any phrasing, keyword, or intent — not on "burn", "budget", "limit", "aufbrauchen", "maximize", "alles verbrauchen", "token", or ANY other wording. Only the literal /run-burn command.

Analyze repository branch hygiene: unmerged branches, stale locals with deleted remotes, active sessions (worktrees), open PRs that still need to land, verify work landed in main. Results: interactive concept page with filters, 2-state delete controls, a ship queue for open PRs (each landed via /ship, one after another, as if shipped from its own session), inline detail expand, and an Apply-Manifest + Dry-Run-Confirm before executing anything. Triggers on: "repo health", "branch cleanup", "branch hygiene", "offene PRs landen", "open PRs shippen". Explicit user request only.

Create GitHub issues — or refine / update existing ones — with enforced title format, labels, and optional milestone and project board integration. Also handles milestone creation and naming. The single owner of every issue write in this plugin: skills, agents and hooks that need an issue created or edited delegate here instead of calling `gh issue create` / `gh issue edit`. Use when the user wants to create a GitHub issue, refine or change an existing one, plan a milestone, or manage issue lifecycle. Triggers on: "neues Issue", "create issue", "Issue erstellen", "mach ein Issue", "new issue", "refine issue", "Issue refinen", "Issue ergänzen", "Issue anpassen", "update issue", "milestone planen", "plan a milestone". Do NOT trigger for: PR creation (use /ship), plain commits, or code implementation.

Audit or initialize a project's repository hygiene: .gitignore, LICENSE, README, .editorconfig, .gitattributes, and AI tooling config. Also scaffolds plugin skill extensions for the project. Triggers on: "set up this project", "init repo", "audit gitignore", "add license", "fix gitignore", "repo hygiene". Do NOT trigger for README generation (/setup-readme), CLAUDE.md edits, or source code changes.

Generate a polished, modern README.md for any project. Use when the user explicitly calls /setup-readme, AND when Claude is about to create, rewrite, or substantially update a README.md. Triggers on: "create a readme", "update the readme", "improve the readme", "README erstellen", "README aktualisieren". Do NOT trigger for minor one-line edits like bumping a version number.

Full end-to-end shipping pipeline using MCP tools: ship_preflight, ship_build, ship_version_bump, ship_release, ship_cleanup, render_completion_card, then silent memory consolidation. Supports hierarchical merges (sub-branch → feature → main). Use when work is ready to land. Triggers on: "ship it", "push and merge". Do NOT trigger during coding/debugging or for commits without shipping.

Stabilization pass: run the full test suite, fix bugs autonomously, identify architecture smells, write regression + missing-coverage tests, apply consistency fixes (spacing, tokens, typography, icons, colors, state-visuals) across the codebase. Does NOT introduce new UI elements, new buttons, re-arrangements, or substantial position changes — those belong to /tune-polish. Triggers on: "harden", "stabilize", "härten", "stabilisieren", "bug pass", "consistency pass", "lint und fix". Skips all confirmations when invoked with --autonomous. Do NOT trigger for: feature work, new UI structure, theme changes.

UI refinement pass: visual consistency (spacing, tokens, typography, icons, colors), state-visuals, UI-side functionality checks, the standing UI rules from deep-knowledge/ui-defaults.md (tooltips, dropdowns, spacing, hotkeys), and small demonstrably UI-related backend fixes. Structural UI changes only with user approval; `--autonomous` skips prompts but keeps the "structural changes flagged not applied" rule. `--invoked-by=ship` is the narrow rules-only path /ship calls: static checks on the diff, no agents, no browser, report-only. Triggers on: "polish", "ui polish", "ui angleichen", "design konsistenz", "feinschliff", "visuell aufräumen", "design pass". Do NOT trigger for: backend-only work, feature implementation, theme/style overhaul.

Strategic reset for stuck development: iterations circle without reaching the goal, or a part of the app is functionally and visually not where it should be and incremental fixes stopped helping. Re-derives the goals from repo + live-app evidence, generates fresh approaches via code-blind lens agents, lets the user decide on a concept page, then hands off to run-autonomous. Triggers on: "festgefahren", "stuck", "unstuck", "wir drehen uns im Kreis", "neu denken", "rethink", "frischer Ansatz", "fresh approach", "komplett neu denken", "das führt zu nichts". Do NOT trigger for: debugging/errors (use /fix), incremental consistency or UI passes (/tune-harden, /tune-polish), or normal feature work.

Live tutorial in the user's own Edge tab for what Claude Code cannot do on a website itself (log in, generate an API key, create an OAuth app, accept terms, change an account setting): a step panel overlay guides one step at a time and takes values back. Triggers on: "guide me through", "führe mich durch", "web guide", "zeig mir auf der Website", "ich muss das auf der Website machen", "walk me through the site", "API key anlegen", "help me set up on <site>". Do NOT trigger for testing the project's own app (use the browser tools directly), for scraping/reading a page, or for local-app tutorials.