
Claude-Orchestrator-Skill
Claude Code plugin for token-saving multi-agent orchestration: Opus plans and verifies, Haiku executes, Sonnet escalates. Includes parallel team mode, autopi…
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up Claude-Orchestrator-Skill (claude-plugin project) into my current project. Found on https://claudeers.com/claude-orchestrator-skill Repo: https://github.com/kele-dev/Claude-Orchestrator-Skill Homepage/docs: — Detected install method: claude-plugin → /plugin install claude-orchestrator-skill@kele-dev/Claude-Orchestrator-Skill Category: skills. Platforms: cli, api, desktop. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
⚠ Unverified / not recently updated — review before pasting a run-this config.
/plugin marketplace add kele-dev/Claude-Orchestrator-Skill /plugin install claude-orchestrator-skill@kele-dev/Claude-Orchestrator-Skill
git clone https://github.com/kele-dev/Claude-Orchestrator-Skill
// compatibility
| Platforms | cli, api, desktop |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | MIT |
| Pricing | open-source |
| Language | JavaScript |
Claude Orchestrator Skill
Bahasa Indonesia: plugin ini membalas dalam bahasa pengguna. Lihat ringkasan.
Claude Code plugin for token-saving orchestration: the main session (any model; Opus-class recommended) thinks and writes the PRD, haiku-coder executes, and sonnet-dev handles escalation. The main session never edits code itself.
Runs on Windows, macOS, and Linux.
Installation
Terminal (Windows, macOS, Linux) - one command:
claude plugin install orchestrator --marketplace kele-dev/Claude-Orchestrator-Skill
If your version of Claude Code does not support the command above, use these two commands:
claude plugin marketplace add kele-dev/Claude-Orchestrator-Skill
claude plugin install orchestrator@kele-dev
Inside a claude session you can also run: /plugin marketplace add kele-dev/Claude-Orchestrator-Skill, then /plugin install orchestrator@kele-dev.
Claude Desktop (Code tab) - click + next to the prompt box, select Add plugins, add the marketplace kele-dev/Claude-Orchestrator-Skill, then install orchestrator with User scope.
After installing, start a new session (Opus recommended as the main model). User scope means it applies to all projects.
Update: claude plugin update orchestrator@kele-dev. Uninstall: claude plugin uninstall orchestrator@kele-dev.
Usage
Orchestration mode is active automatically in every session. Manual commands:
/orchestrator:on- turn on (or/orchestrator:on <task>to give a task directly); optionalbudget=<N>k(e.g./orchestrator:on budget=50k <task>) stops with status BUDGET and a checkpoint once the agents' total tokens reach it/orchestrator:off- turn off for the rest of this session (Claude works directly as usual)/orchestrator:team <task>- parallel mode: the main session splits the work into lanes that own separate files, Haiku lanes run at the same time, only failed lanes are retried. Caps: 3 lanes (4 only at plan time, when the work has 4 or more independent, file-disjoint parts), 2 fix rounds per lane, raised automatically up to 4 rounds only while fixes make progress (a cap you set, e.g.lanes=N, is never raised; optionalbudget=<N>kstops with status BUDGET and a checkpoint once the agents' total tokens reach it). Lanes that need each other's files cooperate through the main session. Add--dry-runto see the lane plan first.isolate=worktree(opt-in) runs each lane in its own git worktree; needsworktree.baseRefset to"head"in settings, a clean working tree, andcommit=batch, otherwise it warns and uses the shared tree; passed lanes are copied back into the main tree./orchestrator:autopilot <task>- autonomous mode: one short clarification round, a PRD, then batch after batch without asking in between, until done or a cap is hit. Caps: 5 batches, 2 fix rounds per batch, raised automatically up to 8 batches / 4 rounds only while work keeps passing (a cap you set, e.g.batches=N, is never raised; optionalbudget=<N>kstops with status BUDGET and a checkpoint once the agents' total tokens reach it). Add--dry-runto see the plan first.- Claude suggests
teamorautopilotwhen a task fits, but always starts inonmode; you decide whether to switch. /orchestrator:learn [list | add <lesson> | prune]- project lessons: at the end of a task Claude decides by itself whether to save up to 3 verified lessons (working test commands, root causes, pitfalls) and fixes or removes outdated ones, as local files in the project store (see Data location below), so every later session in that project can read them, also from a git worktree. Nothing is written into the repo. Old.claude/skills/learned-*lessons are still read and copied into the store on the next save./orchestrator:init [--refresh]- maps a new project in one cheap Haiku pass: stack, key folders, and verified test/build/lint commands, saved as thelearned-maplesson. It also writesverify.shin the project store, one command that runs the verified checks and printsVERIFY: PASS|FAIL./orchestrator:resume [<name>]- continues long work from a checkpoint after a context limit, rate limit,/compact, or crash. team/autopilot write checkpoints automatically at stops and between waves/batches. Checkpoints are kept in the project store./orchestrator:status- read-only summary: mode in this session, plugin version, project lessons, checkpoints, and team/autopilot caps.commit=batch(team/autopilot) - one local commit per passed wave or batch, with a conventional message; never pushed.
Data location: lessons, checkpoints, and verify.sh are stored per project in ~/.claude/orchestrator/<project-key>/ (lessons/, checkpoints/, verify.sh). This store is shared by the main checkout and all git worktrees, is never committed, and needs no git ignore entries. Older lessons in the project's .claude/ folders are still read and copied into the store on the next lesson save (per topic missing from the store). Older checkpoints are not migrated: resume and status still read them and mark them as legacy; delete them yourself once that work is finished.
Token-saving defaults
- Concise answers: no preamble, no restated plan, tables or bullets; detail only when you ask.
- Test/build output is trimmed before it reaches the main session; large suites are run by Haiku, which reports only PASS/FAIL and the first error.
- The main session dictates exact text only for small edits; for long content it writes a spec and the executor writes the file.
- One self-review pass before reporting on full-path or multi-file doc/config work, so you need fewer review rounds.
- A
learned-maplesson (max 20 lines) remembers the project layout and commands, so new sessions skip re-exploring. - Claude suggests a new session or
/compactwhen you switch to an unrelated task. - Every report ends with one token line per agent, e.g.
Tokens: audit Explore 96k · batch1 haiku-coder 12k · 3 delegations, total ≈125k. - Code questions and test-failure triage go to Haiku; the main session reads only a short answer with file:line references.
- Briefs start with a type (
TYPE: EDIT | CREATE | RUN | TRIAGE | SCAN | COMMIT | APPLY) that Haiku already knows, so briefs stay short. - Before verifying a larger change, Haiku scans the diff for mechanical mistakes; the main session reads only the flagged hunks and the main logic.
- In a long conversation Claude writes a checkpoint and suggests continuing in a new session.
- Always-loaded rules are smaller:
hooks/rules.mdis shortened, andorchestrator:onis split into a core plus an on-demandreference.md. - A long-session warning appears when the context grows by 120k tokens, again at +250k and +400k (
ORCH_LONG_REPEAT: comma list oroff), suggesting a checkpoint and a new session (from message usage; 2 MB transcript size if usage is missing) or after 30 delegations (ORCH_LONG=off,ORCH_LONG_TOKENS,ORCH_LONG_BYTES,ORCH_LONG_DELEGATIONS);scripts/digest.mjsgives planners a compact conversation summary.
Hooks
- Session info: SessionStart prints the project store path, a one-line index of saved lessons (at most 20), the 5 newest checkpoints, and a note when legacy files exist, so Claude needs no extra tool calls to find them. Needs Node 18+; without Node the skills compute the path themselves.
- Output trim: Bash/PowerShell output is trimmed only when it is over 80 lines AND over 4000 chars. It keeps the first 3 lines, up to 10 error/failure lines, and the last 30 lines before the model sees it.
git diff/show,cat,jq,Get-Content,type, and JSON output are not line-trimmed (a ~26k-char cap still applies to everything). Read, list and search commands at the start of a statement (ls,find,git ls-files,grep,rg,head,tail,sed -n,wc,Get-ChildItem,Select-String) are line-trimmed only above 300 lines. Add# notrimto a command, or set envORCH_NO_TRIM=1, for full output. EnvORCH_TRIM_TAILandORCH_TRIM_ERRORSchange the 30 and the 10. - Edit guard: the main session cannot Edit/Write project files (except
docs/prd/); subagents can. Turn it off with/orchestrator:offor envORCH_GUARD=off. - The edit guard and output trim need Node 18+ on PATH. Without Node they do nothing (fail open). macOS desktop app: Node from nvm or Homebrew may be missing from the app's PATH; check that
/orchestrator:statusshows the store line, or install Node system-wide. - Review guard (
guard-review): a PreToolUse hook on Agent/Task that deniesorchestrator:opus-reviewerwhen its prompt names adocs/prd/*.mdfile that does not exist. It fails open when Node is missing or the check cannot run. Turn it off with envORCH_GUARD=off.
What's new (1.9.0 and later)
1.12.1
- Short-path review rules apply only to non-Opus main sessions; the review guard accepts absolute and quoted PRD paths with spaces.
- Session start prints
Map: nonewhen the project has no learned-map, so Claude suggests/orchestrator:initfirst. - Output trim: read, list and search commands at the start of a statement are line-trimmed only above 300 lines.
scripts/budget.mjs --mainreports main-session token usage (calls, cache read/write, output, current context);/orchestrator:statusshows it.
1.12.0
- Always-loaded rules are split: non-Opus PRD rules moved to
skills/on/non-opus.md, and situational token rules moved toskills/on/reference.md(read on demand, so they may be followed less often). - Validator size budget:
CONTEXT_BUDGET_BYTEScapshooks/rules.md+skills/on/SKILL.md. scripts/eval-compare.mjscompares eval token totals with a stored baseline; the scheduled run is opt-in. The eval baseline is not committed yet, so the CI compare step stays inactive until the owner runs the eval once and commits the--writeoutput.- Stale checkpoints (older than
ORCH_STALE_DAYS, default 14) are marked in the session-info line.
1.11.0
- Long-session reminders repeat at +250k and +400k tokens (
ORCH_LONG_REPEAT), suggesting a checkpoint and a new session. budget=<N>know also works for plain/orchestrator:on.- New
scripts/trace-tokens.mjsplus the measured startup cost: about 10.6k fixed tokens per haiku-coder delegation (no reduction made). - Honest graders: trace-based main-session and write-success checks, and
tool_orderreview ordering. - 5 new evals plus the filled token log.
- Validator checks eval structure, the
onbudget rule, and new smoke tests.
1.10.1
budget.mjs autofinds its own transcript when noDigest:line is printed (the newest transcript for this project).- Non-Opus PRD rule aligned: the planner's PRD is saved to
docs/prd/only when step 2 requires it.
1.10.0
- New
scripts/budget.mjsthat sums agent tokens from the transcript; it prints aBUDGET: spent <X>k (...) of <N>k: UNDER|OVER|UNKNOWNoutput line for team/autopilot. - Review guard (
guard-review, Agent/Task PreToolUse): deniesorchestrator:opus-reviewerwhen its prompt names a missingdocs/prd/*.md; fails open;ORCH_GUARD=offturns it off. - Edit-guard message now says that
docs/prd/is writable. learned-errorslessons trackhits:and prune the least-hit ones (max 15).- Validator checks stop-status names (for example
BUDGET) and the "Baru di" order. - 3 new evals and the token log; the cache claim is softened.
1.9.0
budget=<N>kfor team/autopilot: stops with status BUDGET and a checkpoint.learned-errorslesson, checked before TRIAGE.- Brief order: stable parts first, variable parts last (may help prompt caching; not measured).
Plugin contents
- Skills
onandoff: the PRD -> delegate -> escalate -> review flow, and the switch to turn it off. - Skills
teamandautopilot(user-invoked only): parallel and autonomous modes on top ofon, with hard caps to keep token cost bounded. - Skill
learn: project lessons saved and maintained automatically as local files in the project store; one report line tells you what changed, and the manual commands are optional. - Skills
init,resume, andstatus: one-pass project map, checkpoints to continue in a new session, and a read-only state summary. - Agents
haiku-coder(executor),sonnet-dev(escalation),opus-reviewer(review, read-only; now runs on Opus),opus-planner(hard-design PRD and diagnosis when the main session is not Opus-class, read-only). - Hooks: SessionStart rules (loads the orchestration rules into every session), SessionStart session info (store path, lesson index, open checkpoints), edit guard (PreToolUse), review guard (PreToolUse, Agent/Task), output trim (PostToolUse), long-session reminder (UserPromptSubmit).
Optional: reduce permission prompts
The plugin cannot change permissions. To keep subagents from waiting for approval so often, pick the Accept edits mode when opening a session, or add this to ~/.claude/settings.json:
{ "permissions": { "defaultMode": "acceptEdits", "allow": ["Bash(npm test *)", "Bash(git diff *)", "Bash(git status *)"] } }
/orchestrator:team runs its lanes as background subagents, which cannot show you permission prompts. Use it with Accept edits (or auto) mode, or a lane may fail when an edit or command needs approval.
Moving from the old installer version
If you previously installed via install.ps1, delete haiku-coder.md, sonnet-dev.md, and opus-reviewer.md from ~/.claude/agents/, the folder ~/.claude/skills/orkestrator/, and the BEGIN/END:orkestrasi block in ~/.claude/CLAUDE.md.
Ringkasan (Bahasa Indonesia)
- Pasang:
claude plugin install orchestrator --marketplace kele-dev/Claude-Orchestrator-Skill - Aktifkan:
/orchestrator:on - Matikan untuk sesi ini:
/orchestrator:off - Mode paralel:
/orchestrator:team <tugas>(beberapa lane Haiku jalan bersamaan, maks 3 lane, 2 putaran fix per lane) - Mode otomatis:
/orchestrator:autopilot <tugas>(batch demi batch tanpa bertanya, maks 5 batch, 2 putaran fix per batch) - Tambahkan
--dry-rununtuk melihat rencananya dulu. - Claude akan menyarankan team/autopilot kalau tugasnya cocok, tapi default tetap
on. Batas iterasi naik otomatis (maks 4 putaran fix) hanya selama perbaikannya ada kemajuan. - Belajar: di akhir tugas Claude sendiri yang memutuskan pelajaran mana yang disimpan (maks 3, hanya yang sudah terbukti), serta memperbaiki atau menghapus yang usang, sebagai file lokal di store project (
~/.claude/orchestrator/<project-key>/lessons/). Satu baris di laporan memberi tahu apa yang berubah./orchestrator:learn listuntuk melihat,pruneuntuk merapikan. - Hemat token: jawaban ringkas, output test dipotong, Opus tidak mendikte file panjang, self-review sebelum lapor,
learned-mapuntuk peta project, saran sesi baru saat ganti topik, dan satu baris laporan token di akhir. - Baru di 1.4:
/orchestrator:init(peta project sekali jalan),/orchestrator:resume(lanjut dari checkpoint),commit=batch, riset/triase oleh Haiku, tipe brief, dan scan diff oleh Haiku sebelum Opus memverifikasi. - Baru di 1.5:
/orchestrator:status(ringkasan status read-only), aturan orkestrasi kini juga dimuat saat sesi di-resume, dan validator pluginnode scripts/validate.mjs+ CI. - Baru di 1.6: hook pemangkas output (Bash/PowerShell di atas 80 baris dipotong jadi baris error + 40 baris terakhir;
# notrimatauORCH_NO_TRIM=1untuk output penuh), hook penjaga edit (sesi utama tidak bisa Edit/Write file project kecualidocs/prd/; matikan dengan/orchestrator:offatauORCH_GUARD=off), keduanya butuh Node 18+, dan/orchestrator:initkini menulisverify.sh. - Baru di 1.6.1: lesson, peta project, dan
verify.shpindah dari.claude/di checkout utama ke~/.claude/orchestrator/<project-key>/, supaya sesi di worktree (misalnya desktop app) bisa menyimpannya. Checkpoint baru juga disimpan di sana. Tidak ada yang perlu di-ignore di git. Lesson lama di.claude/disalin ke store saat lesson berikutnya disimpan (per topik yang belum ada di store). Checkpoint lama tidak dipindah, tetap bisa dibaca oleh resume/status (ditandai legacy), dan boleh dihapus setelah selesai. - Baru di 1.7.0: hook info sesi (path store + indeks lesson di awal sesi, tanpa tool call), output trim lebih ringkas (10 baris error + 30 baris terakhir; env
ORCH_TRIM_TAILdanORCH_TRIM_ERRORSmengubah angkanya), aturan bersama team/autopilot disatukan dimodes.md, aturan legacy dipindah kelegacy.md(dibaca hanya bila perlu),inittidak mengulang isi CLAUDE.md,verify.shdiperbarui otomatis saat perintah di map berubah, dan audit/riset read-only memakai agent Explore. - Juga di 1.7.0 (perbaikan audit): guard tidak lagi mati hanya karena teks menyebut
/orchestrator:off, root proyek memakaiCLAUDE_PROJECT_DIR/git, output satu baris yang sangat panjang ikut dipotong, key store sama di macOS, dan aturan eskalasi/caps diperjelas. - Baru di 1.7.1: trim lebih cerdas (tidak memotong diff/JSON, simpan 3 baris awal), indeks lesson/checkpoint dibatasi, guard hanya membaca 512 KB terakhir transcript, brief COMMIT tanpa trailer, SCAN hanya untuk diff >5 file, perbaikan kecil digabung satu delegasi, laporan agent maks 8 baris, agent memakai
effort, dan lesson hanya disimpan setelah ada kegagalan/eskalasi/koreksi. - Baru di 1.7.2: sonnet-dev dan opus-reviewer membatasi output perintah, resume mencocokkan argumen bebas ke baris
Task:dan langsung bilang bila tidak ada checkpoint, status membaca cap dari skill team/autopilot (tidak ditulis ulang). - Baru di 1.7.3: guard-edit mengizinkan PRD di folder worktree, pemangkas output tidak lagi salah mengenali
--type/[INFO], indeks checkpoint mengenali- Task:, validator memeriksa file rujukan, heading, frontmatter, dan smoke test ketiga hook, CI berjalan di Linux/Windows/macOS. - Baru di 1.7.4: verifikasi lane hanya untuk file miliknya, cek file di luar daftar dibandingkan dengan gabungan semua lane, COMMIT hanya menambah path yang dicantumkan, opsi
isolate=worktreedengan briefAPPLY, output perintah gagal wajib dipangkas lewattail, dan sonnet-dev tidak menjalankan perintah git tulis di lane; cek gabungan memakai baselinegit statussebelum wave, urutan wave isolate dibuat eksplisit, dan dua perbaikan hook (pemangkasan[1/12], huruf kecil path di macOS). - Baru di 1.8.0: model sesi utama bebas (Opus disarankan); bila bukan kelas Opus, PRD dan diagnosis dari agent baru
opus-planner, dan diff akhir wajib keopus-reviewer(kini selalu Opus); sesi Haiku menyarankan ganti model;scripts/digest.mjsmeringkas transcript (barisDigest:di awal sesi); peringatan sekali saat sesi panjang (>2 MB atau 30 delegasi;ORCH_LONG=off,ORCH_LONG_BYTES,ORCH_LONG_DELEGATIONS). - Baru di 1.8.1: sesi utama yang bukan kelas Opus memakai jalur pendek hanya untuk 1 file dengan perubahan teks yang jelas (selain itu lewat
opus-plannerdanopus-reviewer); bertanya sekali bila permintaan bisa masuk ke beberapa tempat (memory, skill, config, kode); laporan dianggap belum selesai bila lebih dari 1 file berubah tanpaopus-reviewer. - Baru di 1.8.2: pekerjaan rutin yang jatuh bersamaan (COMMIT, hapus checkpoint, simpan lesson) digabung dalam satu brief
haiku-coder(hemat ~10k token per delegasi); brief COMMIT bisa membawa barisTRAILER:sehingga atribusi commit sesuai model sesi utama;/orchestrator:resumemengecek commit baru di main yang menyentuh file checkpoint sebelum lanjut; eval baruon-sonnet-guardrails; barisLessons:di awal sesi kini menyebut foldernya; laporanhaiku-codertidak lagi memuat catatan token. - Baru di 1.8.3: peringatan sesi panjang kini diukur dari pertambahan token konteks (
message.usage, ambangORCH_LONG_TOKENS, bawaan 120k) alih-alih ukuran file transcript yang membengkak oleh entri lampiran; ukuran 2 MB hanya dipakai bila data usage tidak ada;guard-edithanya mem-parse baris perintah/Skill orchestrator dan mengabaikan<system-reminder>di depan perintah; validator menambah smoke test deteksi/orchestrator:offdan metrik token; penghitung delegasi dan digest melewati baris lampiran/api-request saat parsing; naik versi dan changelog digabung ke batch kode terakhir (modes.md); hitungan delegasi kini inkremental (offset disimpan di file tmp, hanya byte baru yang dibaca); eval CI memakai versi Claude Code yang dipatok. - Baru di 1.8.4: hemat token di sesi utama: ronde perbaikan dikirim ke agent yang sama lewat
SendMessage(tanpa startup ~10k), review diff per batch hanya membaca delta sejak batch lulus terakhir (snapshot tree), verifikasi dan delegasi berikutnya dalam satu giliran tanpa pesan status kosong, audit read-only wajib keExplore(haiku), laporanhaiku-coderyang bersih cukup 3 baris, dan deskripsi skill dipersingkat. - Baru di 1.8.5: hemat token per sesi dan per tugas:
/orchestrator:statusmembacaskills/status/caps.mdyang kecil (dicek validator terhadap Budget team/autopilot),learnmemindahkan perintah Store dan manual keskills/learn/reference.md, teks COMMIT/APPLY dihaiku-coderdan tiga baris aturan sesi dipersingkat. - Baru di 1.9.0: argumen opsional
budget=<N>kuntuk team/autopilot (berhenti dengan status BUDGET dan checkpoint bila total token agent mencapai batas), lesson khususlearned-errors(pola error terverifikasi -> penyebab -> perbaikan, dipakai sebelum TRIAGE), dan urutan brief yang ramah cache (bagian tetap dulu, bagian variabel terakhir). - Baru di 1.10.0: skrip
scripts/budget.mjs(ringkasan token agent, barisBUDGET: spent <X>k (...) of <N>k: UNDER|OVER|UNKNOWNuntuk team/autopilot), review guardguard-review(memblokiropus-reviewerbila prompt menyebutdocs/prd/*.mdyang tidak ada; fail open;ORCH_GUARD=off), pesan guard-edit kini menyebutdocs/prd/boleh ditulis,learned-errorspunya penghitunghits:dan pemangkasan yang paling sedikit dipakai (maks 15), validator memeriksa nama status stop dan urutan "Baru di", 3 eval baru dan log token, serta klaim cache dilunakkan. - Baru di 1.10.1:
budget.mjs automenemukan transcript sendiri bila tidak ada barisDigest:(transkrip terbaru proyek ini), dan aturan PRD non-Opus diselaraskan: PRD planner hanya disimpan kedocs/prd/bila langkah 2 mewajibkannya. - Baru di 1.11.0: pengingat sesi panjang berulang di +250k dan +400k token konteks (
ORCH_LONG_REPEAT) dengan saran checkpoint,budget=<N>kjuga untuk/orchestrator:onbiasa, skripscripts/trace-tokens.mjs(total token dan biaya startup per agent dari trace eval), grader yang lebih jujur (cek trace khusus sesi utama dan keberhasilan tulis checkpoint, urutan review dengantool_order), 5 eval baru dan log token berisi hasil terukur, serta validator memeriksa struktur eval. - Baru di 1.11.1: saat sesi panjang atau tugas berhenti, Claude menulis checkpoint, memberi ringkasan maks 5 baris, lalu langsung menawarkan card sesi baru yang menjalankan
/orchestrator:resume <slug>(bila tool sesi tersedia, mis. desktop app); selain itu perintahnya dicetak untuk dijalankan di sesi baru. - Baru di 1.12.0: aturan dipisah (aturan non-Opus di
non-opus.md, aturan token situasional direference.mddibaca bila perlu), validator punya batas ukuran konteks,scripts/eval-compare.mjsmembandingkan token eval dengan baseline (jadwal CI opsional; baseline belum di-commit, jadi langkah compare di CI tetap nonaktif sampai pemilik repo menjalankan eval dan meng-commit output--write), dan checkpoint basi ditandai sesuaiORCH_STALE_DAYS(default 14). - Baru di 1.12.1: aturan jalur pendek hanya berlaku untuk sesi utama non-Opus dan guard review menerima path PRD absolut atau berkutip yang berspasi; sesi dimulai dengan
Map: nonebila learned-map belum ada sehingga Claude menyarankan/orchestrator:initdulu; trim output hanya memotong perintah baca, listing, dan cari di atas 300 baris; danscripts/budget.mjs --mainmelaporkan pemakaian token sesi utama (panggilan, cache baca/tulis, output, konteks saat ini) yang ditampilkan/orchestrator:status. - Mode orkestrasi aktif otomatis di setiap sesi.
Maintainers
Releasing: bump version in .claude-plugin/plugin.json and marketplace.json on every release so claude plugin update picks it up. Run node scripts/validate.mjs before pushing (Node 18+, no dependencies); CI runs the same check on every push and pull request. Add the new Baru di X.Y.Z line after the newest one at the end of the list (the validator requires ascending order).
License
MIT, see LICENSE.
// faq
What is Claude-Orchestrator-Skill?
Claude Code plugin for token-saving multi-agent orchestration: Opus plans and verifies, Haiku executes, Sonnet escalates. Includes parallel team mode, autopilot, checkpoints, and self-maintained project lessons.. It is open-source on GitHub.
Is Claude-Orchestrator-Skill free to use?
Claude-Orchestrator-Skill is open-source under the MIT license, so it is free to use.
What category does Claude-Orchestrator-Skill belong to?
Claude-Orchestrator-Skill is listed under skills in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/claude-orchestrator-skill)
// retro hit counter
[](https://claudeers.com/claude-orchestrator-skill)
// reviews
// guestbook
// related in Claude Skills
An agentic skills framework & software development methodology that works.
AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs,…
💫 Toolkit to help you get started with Spec-Driven Development