xm

Agent primitives and autonomous behaviors — fan-out, delegate, broadcast, research, solve, consensus, swarm for Claude Code native multi-agent orchestration

Repository-grounded adaptive execution with deterministic quality gates; planning and legacy orchestration are paid for only when needed

Web dashboard for .xm project state — start, stop, open in browser

Agent output quality evaluation — multi-rubric scoring, strategy benchmarking, and A/B prompt experiments

Session handoff — save comprehensive session state for cross-session continuity

Session restore — resume from last handoff, inject context automatically

Remove AI writing patterns — detect and rewrite AI-generated text into natural, human-sounding prose. English + Korean pattern detection. Based on Wikipedia's "Signs of AI writing" guide and Korean AI-slop conventions. Also fixes translationese in Korean UI and CLI strings (labels, descriptions, buttons, menus, CLI help and status lines) via `ui` mode.

Structured retrospective — reflect on failures together, find root causes, explore alternatives, and grow

Receiving side of /xm:toss — list, take, resolve, or drop cross-project bug reports that landed in this project's .xm/inbox/. Use when the user asks to check the inbox, "받은 편지함 봐줘", "what did other projects toss at us", or wants to start/finish/dismiss a specific handed-off item.

x-mesh toolkit — list available tools and their status

Backlog for off-scope work — capture drive-by bugs, cleanup, and ideas without derailing the current task; promote when ready

Prepare and execute an audited fix in another registered project on the same host. Use when a bug belongs to a locally available project and the user wants it fixed now in an isolated target worktree while retaining toss/inbox and mem-mesh history; use toss instead for remote or asynchronous handoff.

Cross-session decision and pattern memory — persist learnings, auto-inject relevant context on session start

Run mutation testing on changed lines to check whether existing tests detect small source changes, using cargo-mutants (Rust), StrykerJS (JS/TS), gomutants (Go), and Muter (Swift). Use for /xm:mutate, $xm:mutate, $xm-mutate, mutation testing, surviving mutants, or requests to find test detection gaps. Does not generate test code.

Strategy orchestration — 17 strategies including refine, tournament, chain, review, debate, red-team, brainstorm, distribute, council, socratic, persona, scaffold, compose, decompose, hypothesis, investigate, monitor

Cross-vendor entry point + adversarial panel engine. `/xm:panel <verb>` routes multi-model work to the matching consumer in --cross-vendor mode (review→x-review, plan(brainstorm)/debate/council→x-op, solve→x-solver, eval→x-eval, consensus→x-build, fan-out→x-agent); `/xm:panel <target>` runs the panel engine itself (N model CLIs refute each other → consensus verdict); bare `/xm:panel` is an interactive picker; cross/detect/doctor/preflight/types/models are engine utilities (preflight = live model check before a run). Use for "panel review", "다른 모델들로 같이 리뷰", "여러 LLM으로 적대 리뷰", "다중모델로 토론/문제해결/평가", "panel 돌리기 전에 모델/프로바이더 상태 점검", or /xm:panel.

Produce an implementation plan from requirements, show it as readable Markdown, and persist a machine-readable PlanEnvelope under .xm/plan. Use for planning without x-build lifecycle or execution.

Premise validation — challenge assumptions, kill bad ideas early, earn the right to build

Cross-session artifact index. Find and read outputs from previous sessions — reviews, op strategy results, plans/PRDs, eval scores, probe verdicts — stored under .xm/. Use when the user asks to recall/find/look up a past review/op/plan, "what did the last review say", "show me the previous op result", "pull up the last plan", or wants a prior artifact handed to another tool/session (Codex, Cursor). Also exposes a tool-neutral handoff (HANDOFF.md).

Discord에서 Linux host의 managed Claude/Codex session을 모니터링하고 steer/interrupt/decision 입력을 전달한다

Multi-perspective code review orchestrator — PR diff analysis with severity-rated findings and LGTM verdict

Release automation — commit squash, version bump, changelog, push, GitHub release. Works with any project.

Structured problem solving and bug diagnosis — iterate (diagnose → hypothesize → falsify → fix → prove), decompose, constrain, or auto-pipeline

Multi-machine .xm/ state sync — server start/stop, push, pull, setup, status

File a reproducible bug report directly into another registered x-kit project's inbox — captures a repro command + output (secret-redacted), a fix direction, writes a durable local outbox record, and best-effort notifies the target via mem-mesh. Use when a repro found while working in project A actually implicates project B's code, or the user says "toss this to <project>" / "이거 다른 프로젝트 버그같은데 던져줘".

Agent execution tracing — timeline, token/cost tracking, replay, and diff for multi-agent observability

Write GitHub-facing and release documents from repository evidence — PR and issue bodies (published with gh after one confirmation), release notes, changelog entries, and commit messages (text only, for xm:ship). Concise maintainer prose, no invented claims.

Session worktree — isolate the whole current session in a git worktree, then land it back to the parent branch. Use for "worktree로 작업", "이 세션을 worktree에서", "격리해서 작업하고 부모로 머지", "/xm:wt", or when the user wants the ENTIRE session (not one task) to run in an isolated worktree and later merge back.