agent-ways

Start the active awareness layer. Launches attend as a persistent Monitor that surfaces environmental changes (git state, peer sessions, process activity) as notifications. Use when the user says "attend", "start awareness", "watch my session", or invokes /attend.

Check how much context window remains in this session — token budget, usage, and room left before compaction. Use when the user asks how much context/room/budget is left, or wants a context-usage gauge.

Land an increment to main through the branch → commit → push → PR → review-gate → merge → cleanup flow. The review gate is a four-square decision (review depth × human gate). Picks up wherever you are in the cycle. Use when the user says "merge this", "land this", "run a review and merge", "ship it", or invokes /merge.

Open a working session — orient to where work was left off, run a structured interview on a greenfield repo, check the context gauge to confirm it's really session-start, then recommend planning with context already warm. The opening bookend to /wrap. Use when the user says "let's start", "pick up where we left off", "what were we doing", "start a session", or invokes /start.

Carry a piece of work through the development loop — pick the loop shape (variable front, stable tail) by where the uncertainty lives, lay the tasklist, then borrow the stage skills rather than reimplementing them. Use when the user says "let's develop this", "work through this feature", "how should we approach this", "what do we do first", or invokes /develop.

Cut a versioned release — infer the version bump, update the changelog, tag, and publish artifacts (GitHub Release, package registries). The heavy sibling of /merge. Use when the user says "cut a release", "publish a release", "tag a version", "ship a release", or invokes /release.

Update the agent-ways installation to the latest version — pull the app source in $XDG_DATA, rebuild the Rust binaries (ways/attend/way-embed) and corpus, and refresh the ~/.claude projection with `ways reconcile`. Use when the user wants to update agent-ways, pull the latest ways/hooks, or refresh their ways framework install. Not for editing or authoring individual ways (that is the ways skill) or upgrading project dependencies.

Wrap up a session at a natural seam — land in-flight work, square the TaskList honestly, write a copy-paste continuation prompt, then hand off to a directed /compact. Use when the user says "wrap up", "wrap this up", "let's wrap", "end of session", "checkpoint and compact", "hand off", or invokes /wrap.

Localize agent-ways into the operator's language — translate every way's matching metadata against the English root, tune it clean, and switch ways (and Claude Code) into that language. Use when the operator asks to run ways in their language ("set up ways in Spanish", "localiza ways al español", "ways auf Deutsch"). Not for authoring or editing individual ways (that is the ways skill), not for English installs (already built), and not for changing Claude Code's language alone (that is a settings.json edit).

Score way matching via embedding similarity, analyze vocabulary, and validate frontmatter. Use when testing how well a way matches prompts, checking cosine similarity scores, inspecting the embedding engine status, or validating way files.

Look up governance controls and their justifications via the governance CLI, to ground a recommendation in a real standard (NIST, OWASP, ISO, SOC 2, CIS, IEEE). Use when you need the control IDs and justification text behind a practice. Not for deciding whether to cite or how to phrase it — that's the governance citation way.

Review what changed in Claude Code upstream and compare against this project's commits and ADRs. Surfaces opportunities, gaps, and ADR status drift. Discovery conversation, not compliance dashboard.

Watch GitHub from a running session by launching a Monitor instead of polling by hand. `ci` watches the current PR's checks until they resolve; `inbox` watches your GitHub notifications while you work. Use when the user says "watch CI", "tell me when checks finish", "watch my PR", "watch my notifications/inbox", or invokes /gh-monitor.

Produce a "catch me up" briefing — a synthesized start-of-day / overnight summary across email, calendar, tasks, and chat, ending in suggested next actions. Use when the user says "catch me up", "what did I miss", "morning briefing", "what's my day", "start of day". Not for acting on findings (sending, scheduling, creating tasks) — it surfaces and suggests; you approve. Not for single-source lookups or deep person/meeting prep.

Run a read-only, tiered trust audit of an unfamiliar or untrusted repository — repo smell, secrets in git history, dangerous source patterns, and dependency vulnerabilities — ending in a go/no-go verdict. Use when vetting a repo before running or installing it, auditing a dependency or fork, or asked "is this repo safe", "audit this repo", "is this package trustworthy". Not for rewriting git history, setting up CI scanning, or fixing findings — it reads and reports only.

Manage Architecture Decision Records using the project's ADR CLI tool. Use when the user wants to create, list, view, lint, or index ADRs, or when working with docs/architecture/ files. Triggers on "create an ADR", "new ADR", "list ADRs", "lint ADRs", "what ADRs exist", "ADR domains".

Set up direnv/.envrc for per-project Claude Code environments — switch API keys, providers, models, and feature flags automatically when entering a directory. Use when the user wants to manage multiple Claude accounts, switch between Anthropic/Bedrock/Vertex per project, or configure per-directory environment variables.

Author and manage documentation-catalog pages using the project's doc CLI tool. Use when the user wants to write, scaffold, list, or lint catalog docs (tutorials, how-to guides, reference, explanation), or asks "write docs for X", "new doc page", "what docs exist", "document this". Pairs with the adr skill — docs and ADRs share domain bands.

Help the operator and Claude jointly compose a well-structured `/goal` condition before it's set — clarifying the end-state, the evidence, and the bounds. Use when the user wants to set a goal, frame a goal, "give Claude a goal", run Claude in goal mode, or asks for help writing a goal condition. Not for clearing a goal (`/goal clear`), not for running multi-stage delivery, and it does not invoke `/goal` itself — it hands you a condition to set.

Structured reasoning strategies (tree-of-thoughts, trilemma, self-consistency, step-back, ReAct). Use when facing complex decisions, weighing trade-offs, comparing multiple approaches, balancing competing objectives, or stuck. `/think <mode>` runs one strategy; `/think` alone shows the menu.