atlas-auditAudits a repo in CODE (quality, security, dead code, docs drift), ARCHITECTURE (maps, duplication), SELF (atlas run health) or ACT (remediate one finding by id) mode. Use when auditing a whole repo or acting on an audit finding; GitHub alerts go to atlas-security.
atlas-autopilotRuns a ready plan or diagnosed fix end to end with consent gates (orchestrate, diff-mode refactor, review, fix) under a budget, stopping at the local commit. Use when a plan should execute hands-off.
atlas-bakeoffDecides a costly, hard-to-reverse technical choice with evidence: compares 2-3 options on named criteria, or grades one Adopt, Trial, Hold or Reject. Use when a second opinion is needed on a library, architecture or build-vs-buy; reversible picks go to atlas-plan.
atlas-brainstormGrills an unclear product idea one question at a time into a requirements doc (default) or a product strategy; writes no product code. Use when an idea needs shaping, before atlas-plan.
atlas-compoundCaptures one verified, non-obvious lesson into docs/lessons/ after an eligibility gate and overlap check. Use when a problem is solved and a future engineer would repeat it.
atlas-crewRuns a parallel tmux + IRC crew, one worker per file with file leases and a check gate, pre-executed from the run flags. Use when a spec or plan.json splits into independent files to build at once.
atlas-db-auditAudits a live database read-only via parallel subagents: schema inventory, schema-to-code reconciliation, privileges, naming. Use when a schema or permission change is coming.
atlas-debugFinds and fixes the root cause of a reproducible bug, exception, stack trace or wrong output, with evidence. Use when a failure's cause is unknown.
atlas-doctorAtlas self-improvement loop: mines cross-session findings from atlas.db, applies accepted changes to atlas itself, records a baseline. Use when atlas telemetry has accumulated or before changing a skill, hook or agent.
atlas-dogfoodDrives the branch or PR diff in a live browser: full mode repairs genuine breaks in a bounded loop, smoke mode only reports Pass, Fail or Skip per route. Use when changed flows need a browser check; whole-app persona passes go to atlas-ux-test.
atlas-explainExplains how or why existing code behaves, in chat, with file:line evidence from an explorer trace; read-only. Use when asked how something works or what code means.
atlas-frontendBuilds, refactors or polishes UI on one design system (shadcn/ui, Tailwind, Radix), verified live in a browser, in build, component or polish mode. Use when creating or reworking screens or components.
atlas-gitignoreGenerates a zero-trust, deny-by-default .gitignore for a named stack, secrets re-excluded last. Use when starting a repo or hardening a .gitignore.
atlas-handoffWrites a dense session handoff so a fresh session resumes without re-discovery. Use when at a checkpoint before context fills, before a break, or when passing work to another session.
atlas-hardenWrites an idempotent endpoint remediation script (CHECK/SET/VERIFY) for RMM or MDM that proves whether it changed state. Use when remediating or enforcing an endpoint setting at scale.
atlas-loopRoutes a recurring or iterative task to a library loop: interval loops use /loop, fan-out loops run as a parallel verify workflow. Use when repeating a task, iterating until a condition holds or sweeping a backlog; watching one job or PR goes to atlas-run-watch.
atlas-optimizeRuns a single-experiment metric loop: baseline, one hypothesis, one change, identical re-measure, real delta reported. Use when a working system's metric (latency, size, build time) should improve; not for bugs or restructuring.
atlas-orchestrateRuns multi-step, multi-layer or whole-repo work (build, fix, audit, refactor, investigate, full-stack features) through subagents with independent verification. Use when a task spans layers or the repo.
atlas-planWrites an implementation-ready plan to docs/plans/ with U-numbered units and a verification contract, from a brainstorm artifact or feature description. Use when requirements are settled.
atlas-promptRewrites a rough prompt or request into a clear, right-sized prompt for a coding agent or chat model, then has a judge confirm every goal and limit survived. Use when asked to rewrite, improve, optimize, tighten, shorten or sharpen a prompt (even a one-line prompt pasted inline), whether a prompt is
atlas-proofPublishes project markdown for human review, records anchored comments in a sidecar file, folds feedback back. Use when sharing a plan, spec or draft for comments; not proofreading.
atlas-prototypeBuilds a throwaway prototype in an isolated run directory to answer one visual, interaction or product-shape question, then promotes decisions or discards it. Use when showing beats arguing.
atlas-pulseGenerates a time-windowed product pulse (usage, performance, errors) from wired telemetry, read-only, into docs/pulses/. Use when a usage and error summary of the running product is needed.
atlas-readmeGenerates an onboarding-grade README.md by inspecting the repo, tracing every claim to a file. Use when a repo has no README or a stale one.
atlas-refactorRefactors, renames or restructures code without changing behavior, or (diff mode, no target) runs a three-reviewer simplify pass over a fresh diff, with before/after evidence. Use when code works but is messy, or right after implementing and before review.
atlas-resolve-pr-feedbackTriages each PR review thread (fix, reply, needs-human), applies approved fixes, drafts replies, confirms before pushing or posting. Use when a PR has open review comments.
atlas-reviewReviews a diff, branch or PR with risk-selected personas; P0/P1 findings verified, report-only verdict (Ready, Ready-with-fixes, Not-ready). Use when a change needs review before merge.
atlas-run-watchWatches a long-running job (PID, log, CI URL, command loop) or one open PR's CI with a tick loop and run ledger; PR mode repairs genuine CI failures in a bounded loop and never merges. Use when a job or PR's CI needs watching.
atlas-securityTriages open GitHub Dependabot, code-scanning and secret-scanning alerts into one grouped fix on a working branch, verified by a fresh verifier. Use when handling vulnerabilities, CVEs or the security tab.
atlas-setupManual atlas lifecycle: onboard (scaffold docs/ and .atlas/), install, connectors, repair, tour, menu. Use when bringing atlas online, fixing it or learning it.
atlas-shipShips finished work: a fresh-context verifier checks a clean tree, then a local commit; stops before any push or PR. Commit-only mode commits a file list. Use when work is ready to commit or open a PR.
atlas-statusBuilds a stakeholder status rollup for a plan or feature from plans, findings, the todo board and commits (claims marked verified or assumed); announce mode drafts release notes. Use when asked for feature status or to announce shipped work.
atlas-sweepTriages GitHub issues into a docs/features/ doc (sweep mode, never closes or replies) or extracts findings from a transcript or notes (analyze mode). Use when user feedback or support tickets need triage.
atlas-test-xcodeTests an iOS app on the simulator: boots it, builds and tests via xcodebuild, captures screenshots and logs. Use when building or testing an iOS or Xcode project; macOS only.
atlas-testingPlans, writes and runs tests in any language: maps gaps across test types (unit through mutation, fuzz, load, security, accessibility), writes the highest-value next test, reports evidence. Use when adding tests or finding what testing is missing.
atlas-tuiOpens the Atlas Command Center, a live terminal UI over the colony's tmux workers (topology, seats, feed, system). Use when watching or attaching to running colony workers.
atlas-ux-testRuns a UX test swarm over a whole web app: generates personas, drives real-browser walkthroughs and fuzzing, gates on client-surface correctness. Use when a full UI/UX pass or persona testing is needed.
atlas-validateVerifies a finished Claude Code plugin: structure, manifest and content quality, pass or fail per check with file:line findings, no auto-fix. Use when a plugin is believed complete.
atlas-wikiGenerates and refreshes docs/wiki/ diagrams from docs/architecture/ using graphify. Use when architecture docs changed, before the completion gate, or when wiki diagrams are stale.
atlas-worktreeGit-worktree primitive: creates or reuses linked worktrees, tears down with dirty-state protection. Use when isolating parallel implementer waves, spikes or bakeoffs, or closing them at wave end.