Koder CLI

Ubuntu 26.04.1 Desktop amd64 walkthrough

Ubuntu / macOS: run these in the system terminal. Skip setup if the model is already configured.

"$HOME/.local/bin/koder" setup
"$HOME/.local/bin/koder"

Enter tasks only after koder > appears. activate enables a background service; it does not open chat. If the full path works but koder does not, follow the PATH walkthrough rather than downloading again. Use the actual installer path for custom installations.

Current release: 0.5.14. Private credential input and Ubuntu login routing.

Credential setup and login · Ubuntu tutorial · All-platform installation · Start here · Download packages · 中文 Ubuntu 教程

English · 简体中文 · v0.5.14

A terminal coding agent that plans, proposes, and applies work on your machine—with your own model API, workspace-isolated history, and local approval before sensitive changes.

Koder CLI is a local task tool from the KARMEN LOOP team. Bring your own provider key or supported local model, choose a project, and work through plans, tools, approvals, and recorded results. Personal, noncommercial local use is free and does not require membership or an invite code. Commercial use and organizational deployment need separate written authorization. Providers may charge for API requests.

Current release: 0.5.14 · Agreement: Koder-EULA-1.1 · Installation: GitHub guide

Upgrading from 0.4.4 or earlier: if an older updater rejects the larger package, run the current official installer once. Configuration and history are retained. Installation and migration.

0.5.3 startup output captured from an isolated macOS PTY and rendered from ANSI. Font rendering varies by terminal.

Download 0.5.14 · Installation · English guide · Chinese guide · Release notes · Command reference

Most terminal agents either hide what they will change, or treat every model suggestion as permission. Teams that already pay for model APIs should not need a second membership gate just to run a local client—and they should see what will run before sensitive edits, deletes, or shell effects land.

Koder keeps four things deliberate:

1. Your model, your bill — Chat Completions, Anthropic Messages, and Responses; catalog entries show current availability next to historical references. A website subscription is never assumed to be API access.
2. Local control plane — planning, permission checks, tool execution, and saved results stay on your machine under the KARMEN LOOP kernel narrative. Execution grants are hashed and scoped; they are not synchronized as “memory.”
3. Smart approval by default — recognized low-risk local checks and ordinary project edits can proceed; code execution, deletion, secret paths, shell effects, unknown commands, and external app actions ask. The decision comes from local policy code, not a model’s claim.
4. Honest boundaries — gray provider rows stay unselectable; native Windows GUI / ConPTY / real mailbox / model-quality paths that were not run stay labeled NOT_RUN in release evidence.

It does not claim that local fixtures certify every PC, provider account, or model quality score.

Three layers stay separate on purpose. The model proposes; the local policy decides; the workspace records.

1. Client & session — interactive chat, editable / prompt choices, mid-run steer-or-queue, multi-window task sessions, and bounded private runtime logs (koder logs, doctor). Queued text cannot approve a pending sensitive prompt.
2. Control plane — plan → local permission policy → tool execution → recorded results. Grants are action-hashed and scoped to a workspace (or an explicit profile with confirm). They live under the local CLI home as execution-permissions.json and are not synced as conversation memory.
3. Your model & tools — Chat Completions, Anthropic Messages, or Responses on endpoints you configure. Optional MCP / Skills / apps attach only after their own setup and approvals. Protocol compatibility does not certify every third-party service.

flowchart TD
  U[You: task in a project workspace] --> P[Plan and tool proposals]
  P --> A{Local approval policy}
  A -->|Low-risk recognized edit or check| X[Execute tools]
  A -->|Sensitive / unknown / shell / delete| Q[Ask: once / always / full / decline]
  Q --> X
  X --> R[Recorded results in workspace-scoped history]
  R --> M[Your model API: completions / messages / responses]
  M --> P
  S[Setup / permissions / privacy scopes] -.-> A
  S -.-> M

Isolation note. The session label default means the default session of that canonical workspace, not one global conversation across every folder. Private transcripts live in the user Koder profile, indexed by workspace and session—not inside a public site output directory.

How the pieces fit together at a glance—control stays local; the model is yours; sensitive apply waits for approval.

flowchart LR
  root((Koder CLI))
  root --> local[Local control]
  root --> model[Your model API]
  root --> work[Workspace work]
  root --> safety[Approval & review]
  root --> docs[Docs & learning]

  local --> L1[Plan → tools → results]
  local --> L2[Runtime logs / doctor]
  local --> L3[Multi-window sessions]

  model --> M1[Chat Completions]
  model --> M2[Anthropic Messages]
  model --> M3[Responses]
  model --> M4[Gray providers blocked]

  work --> W1[Koder Projects]
  work --> W2[Workspace-scoped history]
  work --> W3[ask / run / review]
  work --> W4[Skills / MCP optional]

  safety --> S1[Smart approval default]
  safety --> S2[Diff before apply]
  safety --> S3[execution-permissions.json]
  safety --> S4[EULA & privacy scopes]

  docs --> D1[USER-GUIDE EN/ZH]
  docs --> D2[koder learn 1,800 Chinese lessons]
  docs --> D3[Command reference]

Six real captures from docs/screenshots/, ordered as a first-success path. The Guided walkthrough embeds each image under its step—hero stays lean on purpose.

flowchart LR
  A[1 Install & EULA<br/>cli-install] --> B[2 Session start<br/>cli-start]
  B --> C[3 Home / project<br/>tui-home]
  C --> D[4 CLI diff<br/>cli-diff]
  D --> E[5 TUI diff<br/>tui-diff]
  E --> F[6 Review & apply<br/>cli-review]

Use Windows 10/11, macOS, Linux, or WSL with the matching shell. Normal installation does not require administrator access. Read the EULA, acceptance guide, legal notice, and CLI privacy before typing accept.

Choose one command. Wait for accept / decline:, type accept, press Enter. decline or Ctrl+C cancels. Empty Windows input asks again — it does not accept. License acceptance enables no optional computer, mail, profile-learning, sync, birthday, music, or country-lookup permissions.

The matching screenshot lives in walkthrough step 1.

$ErrorActionPreference='Stop'; $ProgressPreference='SilentlyContinue'; [Net.ServicePointManager]::SecurityProtocol=[Net.ServicePointManager]::SecurityProtocol -bor [Net.SecurityProtocolType]::Tls12; $u='https://github.com/KarmenLoop/koder-cli/releases/download/cli-0.5.14/koder-cli.ps1'; Write-Host ('Koder: loading installer from '+([Uri]$u).Host+'...'); $limit=100000000; if($env:KODER_DOWNLOAD_TIMEOUT_SECONDS){$v=0; if([int]::TryParse($env:KODER_DOWNLOAD_TIMEOUT_SECONDS,[ref]$v) -and $v -ge 1 -and $v -le 100000000){$limit=$v}else{throw 'KODER_DOWNLOAD_TIMEOUT_SECONDS must be 1..100000000'}}; $connect=30; if($env:KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS){$v=0; if([int]::TryParse($env:KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS,[ref]$v) -and $v -ge 1 -and $v -le 3600){$connect=$v}else{throw 'KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS must be 1..3600'}}; $bootstrapLimit=[Math]::Min($limit,$connect); $client=New-Object Net.WebClient; $client.Encoding=[Text.Encoding]::UTF8; $clock=[Diagnostics.Stopwatch]::StartNew(); $next=5; try{$task=$client.DownloadStringTaskAsync([Uri]$u); $wait=([IAsyncResult]$task).AsyncWaitHandle; while(-not $wait.WaitOne(200)){if($clock.Elapsed.TotalSeconds -ge $bootstrapLimit){if($limit -le $connect){throw ('Installer download timed out after '+$limit+'s')}; throw ('Installer bootstrap stalled after '+$connect+'s')}; if($clock.Elapsed.TotalSeconds -ge $next){Write-Host ('Still loading installer: '+[int]$clock.Elapsed.TotalSeconds+'s elapsed. Ctrl+C cancels.'); $next+=5}}; $script=$task.GetAwaiter().GetResult()}finally{$client.CancelAsync(); $client.Dispose()}; Invoke-Expression $script

Paste the complete line at a CMD prompt (not into PowerShell):

powershell -NoProfile -ExecutionPolicy Bypass -Command "$ErrorActionPreference='Stop'; $ProgressPreference='SilentlyContinue'; [Net.ServicePointManager]::SecurityProtocol=[Net.ServicePointManager]::SecurityProtocol -bor [Net.SecurityProtocolType]::Tls12; $u='https://github.com/KarmenLoop/koder-cli/releases/download/cli-0.5.14/koder-cli.ps1'; Write-Host ('Koder: loading installer from '+([Uri]$u).Host+'...'); $limit=100000000; if($env:KODER_DOWNLOAD_TIMEOUT_SECONDS){$v=0; if([int]::TryParse($env:KODER_DOWNLOAD_TIMEOUT_SECONDS,[ref]$v) -and $v -ge 1 -and $v -le 100000000){$limit=$v}else{throw 'KODER_DOWNLOAD_TIMEOUT_SECONDS must be 1..100000000'}}; $connect=30; if($env:KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS){$v=0; if([int]::TryParse($env:KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS,[ref]$v) -and $v -ge 1 -and $v -le 3600){$connect=$v}else{throw 'KODER_DOWNLOAD_CONNECT_TIMEOUT_SECONDS must be 1..3600'}}; $bootstrapLimit=[Math]::Min($limit,$connect); $client=New-Object Net.WebClient; $client.Encoding=[Text.Encoding]::UTF8; $clock=[Diagnostics.Stopwatch]::StartNew(); $next=5; try{$task=$client.DownloadStringTaskAsync([Uri]$u); $wait=([IAsyncResult]$task).AsyncWaitHandle; while(-not $wait.WaitOne(200)){if($clock.Elapsed.TotalSeconds -ge $bootstrapLimit){if($limit -le $connect){throw ('Installer download timed out after '+$limit+'s')}; throw ('Installer bootstrap stalled after '+$connect+'s')}; if($clock.Elapsed.TotalSeconds -ge $next){Write-Host ('Still loading installer: '+[int]$clock.Elapsed.TotalSeconds+'s elapsed. Ctrl+C cancels.'); $next+=5}}; $script=$task.GetAwaiter().GetResult()}finally{$client.CancelAsync(); $client.Dispose()}; Invoke-Expression $script"

With wget installed, paste this complete line. The installer runs only if its download succeeds:

wget --tries=3 --timeout=30 -O /tmp/koder-ubuntu-install.sh https://github.com/KarmenLoop/koder-cli/releases/download/cli-0.5.14/koder-ubuntu-24.04.sh && bash /tmp/koder-ubuntu-install.sh

Use the Ubuntu walkthrough for curl or missing download tools. Installation reports MB, percentage and elapsed time, and supports resuming interrupted transfers. Open a new terminal afterward and run koder --version and koder setup. Desktop operations need separate dependencies and permission.

curl -fsSL https://raw.githubusercontent.com/KarmenLoop/koder-cli/main/install/koder-cli.sh | bash

Downloads show progress and verify SHA-256. Reopen the terminal after success, then:

koder --version
koder setup

koder --version should report 0.5.14. Full platform notes: Installation.

After install and a fresh terminal:

koder setup
koder learn
koder

• koder setup — connect your model API (required before real tasks).
• koder learn — 1,800-lesson Chinese course with the original 30-lesson practice coach; no paid model required to read.
• koder — interactive chat in the current workspace.

Prefer a scripted first success? Follow the Guided walkthrough below.

A first-success path from zero to a conscious approve-or-decline. Deeper workflows (projects, preview, Netlify, mail, residency) live in the English user guide and Chinese user guide—this section alone is enough to install, configure, run one task, read the diff, and decide.

1. Pick one installer command from Install above (PowerShell, CMD, or macOS/Linux/WSL).
2. Wait until progress finishes and the prompt shows exactly: accept / decline:.
3. Type accept and press Enter. Type decline or press Ctrl+C to cancel. On Windows, an empty Enter asks again—it does not accept.
4. When the installer finishes, close and reopen the terminal so PATH refreshes.

Step 1 — installer progress and the EULA prompt. Acceptance enables no optional computer, mail, sync, or celebration extras.

koder --version
koder

Expect 0.5.14. Bare koder opens interactive chat. Prefer a short coach first?

koder learn

Enter continues a lesson, b goes back, q leaves. Offline text editions: koder learn --print --os mac --lang en (also powershell / cmd / linux; --lang zh for Chinese).

Step 2 — session start after a clean install.

koder setup

In setup:

1. Choose an available (non-gray) provider entry.
2. Enter the exact model ID, protocol, and endpoint your account issues.
3. Prefer an environment-variable key reference when you do not want the key value stored in the profile; never paste a key into a task, screenshot, or issue.
4. A website subscription is not automatically API access.

Then create or select a project. New website / slides / app work defaults to Koder Projects under your home directory unless you pass -w or continue an existing workspace:

koder project create "Studio Site" --kind website
koder project create "Quarterly Review" --kind slides
koder project create "Studio App" --kind app
koder project list --json
koder project status
koder chat -w /absolute/path/to/existing-project

Inside chat: /project create "Another Site" --kind website creates and selects; /project open "/absolute/path" returns to an existing folder. Name collisions are rejected—files are not merged.

Step 3 — TUI home after setup; confirm the workspace before the first real task.

Smoke-test the provider when ready (these calls may incur provider charges):

koder model presets
koder model test
koder model test --tools

model test uses a short diagnostic budget; a text reply proves text access only. --tools requests a harmless tool round-trip and continuation.

In chat, type a concrete, bounded ask—for example: explain a file, add a short README note, or repair a failing local check. Or from the shell:

koder ask "Explain the architecture of this project" --lang en
koder run "Add a short note to README about local setup" --lang en

When the agent proposes project edits, read the CLI diff before you approve:

1. Paths — are they inside the intended project?
2. Deletions — intentional, or collateral?
3. Secrets / config — anything near keys, .env, credentials, or deploy config?
4. Scope — does the hunk match the ask, or drift into unrelated files?

Step 4 — CLI diff-oriented review. Approval policy, not an OS sandbox.

If you prefer a full-screen view, open the TUI diff for the pending change. Captions and hunks match the CLI story—same workspace, same policy. Approve or decline with equal care.

Step 5 — TUI diff of a proposed change.

At a sensitive prompt:

Decline is a first-class outcome. An unknown result is not success—verify state before retrying a send, payment, or deployment.

After you accept, the change lands in workspace-scoped history. For a structured review pass:

koder review
koder review --format json

Use koder review when you want findings after apply, before a larger follow-up, or when a CI-style gate should distinguish a clean parse from failure. If a successful model answer still cannot be parsed into complete findings, the CLI exits 3 with REVIEW_UNPARSED—do not treat that as a clean review.

Step 6 — review / fix loop after apply.

First success means: version reports 0.5.14, setup has a working (non-gray) provider, one task produced a readable diff, and you consciously approved or declined. For multi-window sessions, Skills/MCP, residency, and mail inquiry, continue in the user guide.

Day-to-day chat: type / on an empty prompt for editable choices (Tab fills, Esc closes, Enter sends only after fill). /setup opens settings, /model manages models, /new starts a topic, /exit leaves. /menu opens settings; Escape leaves.

Smart approval is the default: recognized low-risk local checks and ordinary project edits can proceed without a question. Code execution, deletion, secret or configuration paths, shell effects, unknown commands, and external application actions ask. The decision is computed by local policy code—not by trusting the model’s description of risk.

This is an approval policy, not an OS sandbox. It does not elevate privileges, bypass accessibility prompts, or replace your provider’s security model.

At a sensitive prompt: y once, a / always for that exact action in this workspace, f / full for this workspace (or --scope profile with explicit confirm), Enter / n to decline.

koder permissions execution status
koder permissions execution ask
koder permissions execution auto
koder permissions execution full --confirm FULL
koder permissions execution full --scope profile --confirm FULL
koder permissions execution revoke all
koder permissions execution reset

Rules live under the local CLI home as execution-permissions.json (host/workspace scope + action hashes—not raw commands or keys). Damaged records fail closed: Koder offers a fresh decision; they do not silently restore broad access. Execution authority is not synchronized as memory.

Privacy and feedback stay offline-readable:

koder privacy consumer
koder privacy statement
koder privacy desktop
koder privacy agreement
koder privacy sources
koder feedback
koder permissions status

koder feedback opens a blank template for info@karmenloop.com — nothing is sent or attached automatically. Strip keys, passwords, private paths, and personal records before you send.

Configure in koder setup or with explicit profile commands. Keep three boundaries clear:

1. Billing — Koder does not replace your provider’s billing. Website membership ≠ API access.
2. Protocols — Chat Completions, Anthropic Messages, and Responses. Catalog rows may show historical references; current availability is shown next to each entry.
3. No silent fallback — independent plan entries keep independent hosts, keys, and billing routes. The internal backup engine stays disabled and gray by default.

koder model presets
koder model models aliyun-coding-plan
koder model test
koder model test --tools
koder model set --model ACTUAL_MODEL_ID --base-url OFFICIAL_API_BASE --protocol openai --key-env MY_PROVIDER_KEY

See PROVIDERS.md and MIMO.md for endpoints and safe setup.

Temporarily unavailable (gray, blocked before network): MiMo Token Plan, Kimi API, Kimi Code, DeepSeek, Alibaba Token Plan, Volcano Coding Plan, and Volcano Ark API. Existing profiles remain for view/remove; activation, test, refresh, and execution stay blocked. No automatic billing-route switch.

Remote Skill discovery now preserves repository-qualified searches such as repo:anthropics/skills. Candidates remain pinned to commits and verified file hashes. The original tool loop can fetch and read a Skill, then continue the unfinished task under existing permissions. Task rounds remain unlimited by default; optional cost budgets and cancellation still apply to the parent and its children.

Hands-on guide · 中文操作教程 · Evidence and retained limits.

Awesome LLM Apps is enabled by default for relevant AI-app, RAG and agent-team tasks. Search 146 pinned examples, inspect their actual source, and build through the existing Koder provider, tools and approvals. Seven adapted methods cover reviews, history, dependencies and delegation. Local retrieval and quote checks operate on explicitly named project files.

koder llm-apps status
koder llm-apps list --query RAG
koder llm-apps plan rag_tutorials/rag_chain
koder llm-apps retrieve policy.md --query "returns" --json

Hands-on guide · 中文教程 · Full catalog. Disable it with koder llm-apps off or Setup → Skills. Bundled examples are reference source; their cloud accounts, Python/Node dependencies and UIs are not automatically installed or certified.

mkdir -p "$HOME/Koder Projects/plate"
cd "$HOME/Koder Projects/plate"
koder cad enable
koder cad verify
koder cad test
koder

Ask for a dimensioned part, a measured volume, an editable .solidcraft design and the required STEP/STL/PNG outputs. Reopen the saved design explicitly in later turns. English tutorial · 中文手把手教程. Native Windows and Intel Mac runs remain beta acceptance; the Apple-silicon check uses real CAD tools with a scripted model.

SolidCraft 0.23.0 is bundled for Windows x64 and macOS Apple silicon/Intel. Enable it per project with koder cad enable, check the connection with koder cad test, and use the built-in Skill to model, measure, save and export. Every turn receives an isolated headless design through the existing Core; CAD delivery checks use files and measurements rather than software-test requirements. Existing mail, Skills, unlimited task rounds, cancellation and cost budgets remain. Read the hands-on CAD guide and release notes.

Remote references stay off by default. Opt in per command with koder resources search "keywords" --remote, or set KODER_REMOTE_RESOURCES=1 for agent-visible remote tools (0 forces off). Import local archives with koder resources import — originals stay local; see REMOTE-RESOURCES and PRIVATE-RESOURCES.

For company licensing and dedicated private repository delivery: koder business or info@karmenloop.com. A separate written agreement defines scope; stars and inquiry emails alone are not authorization (COMMERCIAL-LICENSING).

Eligible managed interactive chats check the official update feed at startup and periodically every six hours and install only when tasks, queues, and drafts are clear. A changed license requires manual review.

koder update --check
koder update
koder update auto status
koder update auto off
koder update auto on

KODER_AUTO_UPDATE=0 disables automatic checks for that process. Scripts and one-shot commands do not auto-install. Details: AUTO-UPDATE.

Release evidence distinguishes source / fixture checks, installed protocol fixtures, hosted native Windows, real provider accounts, and user devices. Local no-key regressions exercise CLI behavior with scripted providers; they do not certify PowerShell 5.1 cold install, ConPTY/IME, Windows-MCP GUI operation, mailbox delivery, or model quality. Those items remain NOT_RUN where the 0.5.14 notes and Windows install checks say so.

Ready-to-use packages, installers, and checksums ship as release assets. New tags may add distribution documentation without publishing proprietary runtime source to this public tree.

Do I need an invite code or paid membership to install?
No. Personal, noncommercial local use is free under Koder-EULA-1.1 and does not require membership. Commercial and organizational use need separate written authorization.

Whose model API does Koder use?
Yours. Configure a provider in setup. Koder does not replace your provider’s billing. Some catalog rows are temporarily gray and blocked (see Models and providers).

Will it change my files without asking?
Recognized low-risk local checks and ordinary project edits may run under smart approval. Sensitive or unknown effects—including deletion, secret paths, and shell side effects—ask first. Review diffs and targets before approving.

Is the public GitHub source the 0.5.14 runtime? No. The public tree keeps historical GPL 0.3.1 material for continuity. The 0.5.14 application runtime is proprietary and distributed via the official installer and release feed (SOURCE-STATUS.md).

Can my company deploy it?
Contact info@karmenloop.com or run koder business. Stars, downloads, and inquiry emails alone are not commercial authorization.

Where do projects and private history live?
New creation tasks default to Koder Projects under your home directory unless you pass -w or continue an existing workspace. Private conversation records stay in the user Koder profile, indexed by workspace and session—not inside a public site output folder.

When should I run koder review?
After apply when you want structured findings, before a larger follow-up, or when a gate must distinguish a clean parse from REVIEW_UNPARSED (exit 3). It is not a substitute for reading the diff before the first approve.

Where is the hand-holding tutorial?
On this page: Guided walkthrough (with all six screenshots). Full depth: USER-GUIDE.en.md · USER-GUIDE.zh.md · koder learn.

• Current distribution: EULA.md (Koder-EULA-1.1) · scope note: LICENSE
• Third-party: THIRD_PARTY_NOTICES.md · OPEN_SOURCE_CREDITS.md · licenses/
• Bundled method packages retain their own licenses (for example the unmodified personal-agent-growth method remains AGPL-3.0-only where shipped)

KARMENLOOP Official Team · Shanghai, China · karmenloop.com · info@karmenloop.com

If Koder is useful, an optional GitHub star helps others find it. Skip freely — stars never affect installation or licensed features.

Classroom edition 2026.10.26.1 adds 1,800 downloadable practice workbooks, twelve domain guides, and an offline exercise helper. Start with the hands-on guide or download the complete classroom. This course revision is separate from CLI 0.5.9.

The 1,800-lesson Chinese classroom is organized into 120 units, with explanations, examples, exercises, quizzes, troubleshooting and developer guidance. Notes and progress stay in the browser and can be exported. The original 30 English lessons and existing practice project remain available. The classroom has shipped since 0.4.12; update the installed CLI to use the added lessons. See the course guide for the two learning tracks and local classroom controls.

Version 0.4.12 introduced koder migrate to Setup and chat. Export an encrypted project snapshot, import it into an isolated profile, or send it through an already configured SSH connection. The destination keeps its existing data and requires fresh credentials and device permissions. This is a data migration, not a transfer of running processes.

Read the step-by-step migration guide or the Chinese classroom guide for supported data, limits, verification and rollback. Real cross-device Windows and SSH acceptance remains pending; see the release notes.

Build from a clear goal, audience and primary action. Koder now bundles 84 complete DESIGN.md references, asks about missing requirements, records project design decisions and continues through the existing implementation, preview and delivery tools. The method is enabled by default and can be disabled in Setup or with koder design off. Guide · 中文教程.

Discuss missing requirements, keep a stable project purpose and build through the existing implementation and preview tools. The original Fullstack Web Delivery 0.9.1 Skill is bundled unchanged and enabled by default. Each new project has its own folder and PROJECT-ANCHOR.md; old-site repairs retain their directory and decisions. Hands-on guide · 中文教程.

Navigate local Markdown, text and PDF trees with the existing Koder model. Read original sections and verify file versions, page/line ranges and exact quotes before answering. The PDF parser ships in the runtime; no Python or new cloud account is required. Enabled by default for document tasks and configurable in Setup. Hands-on guide · 中文教程.

Run koder tools audit --json to inspect mounted channels, permissions, Skill text and parser resources. Scan images and PDF pages with bundled Tesseract.js: koder ocr read invoice.png --json. The 90 professional child-task roles specify inputs, deliverables and checks; they reuse the existing Core and provider. Four children execute concurrently by default, with FIFO admission; 180 task slots are not a paid-provider quota. Headless tasks return explicit completion states, including preflight failures. Execution guide · 中文手把手教程.