Issue 2026-06-21
v0.8.61
CodeWhale
any model, open models first
Open source · multi-model · local-first

An agentic terminal for the best available models.

CodeWhale is a community-built terminal agent that runs locally and works with the models you actually use — Arcee Trinity, GLM, DeepSeek, Kimi, MiniMax, OpenRouter, and more. A full tool surface, approval gates, snapshots you can roll back, sub-agents, and sessions you can resume.

Mission idea put in this versionStart with an accountable self — this terminal, this repo, this session. Add a constitution for conflict. Then make tools, approvals, rollback, and handoffs leave local evidence. The exact shape will keep iterating.
Maintained by Hmbown0.8.6325 providers
one-line install
# npm — no Rust toolchain required
$ npm install -g codewhale
$ codewhale

# Cargo / Homebrew / direct download: # /install
needs Node or Rust 1.88+other ways →
Stars · 星标
38.8k
Forks · 复刻
3.3k
Contributors · 贡献者
160
Latest · 版本
v0.8.63
New in v0.8.63

Multi-agent runtime, model routing, open provider surface.

0.8.61 focuses on the runtime control plane: sub-agents stop freezing the TUI, provider/model routes can vary by role, route state is isolated per session, and Arcee AI joins the public provider catalog for open-model work.

#3216 / #3266
Report →

Sub-agents without TUI stalls

sub-agent progress is coalesced off the render path, completion resumes the parent turn, and multi-worker stress coverage guards input/render/cancel liveness.

#2027 / #1768
Report →

Per-role model routes

sub-agents can run with explicit provider/model, strength, and thinking controls, so scouts, verifiers, and synthesis work can use different routes.

#3071 / #3076
Report →

Arcee and open-model routes

the provider catalog now includes Arcee AI alongside DeepSeek, GLM/Z.ai, Kimi, MiniMax, OpenRouter, local vLLM/SGLang/Ollama, and more.

See how it decides

Every agent claims to be aligned and trustworthy. CodeWhale can prove it — the Constitution shows up in the model's reasoning, citing "Article II" and "Article V" as it decides. These are real traces from an actual session, paired with the decision each produced.

codewhale — thinking
reasoning trace
When the prior session's summary didn't match the real repo state…
 Let me be careful and truthful here. Per Article II (truth) and V (verification), I need to report what's actually in the tree — not what the previous agent claimed.
Article II · TruthArticle V · Verification
Reported the uncommitted work the summary left out, instead of repeating the claim.
codewhale — thinking
reasoning trace
When a “pause” signal conflicted with the user's explicit request to keep going…
 The user's current request outranks the runtime statute — so I keep executing the task, not stopping on a technicality.
Article III · User agencyArticle VII · Hierarchy
Finished the work the user actually asked for.
codewhale — thinking
reasoning trace
When trimming the README, tempted to drop the contributor wall…
 Deleting community credit would break the project's own ethos. Relocate it instead — zero names lost.
Community ethos
Moved 200 lines of credits to docs/CONTRIBUTORS.md and linked from the README.
Faithful excerpts from real reasoning — not invented. The constitution isn't a poster on the wall; it's the order the model actually follows when it decides.

The judgment stack

A model generates text; an agent leaves consequences. CodeWhale puts the structure for those consequences around the model: self, authority, evidence, and tools, each in its place.

01 · self-model

Responsibility needs an address

The agent is not a model card or leaderboard score. It is an instance in this terminal, this workspace, this session. Give it an address before you ask it to act.

02 · nested constitution

Conflict has law

Global Constitution, current user request, runtime statutes, repo-local law, live evidence, memory, personality, and old handoffs each have a rank. When they conflict, the agent has an order to follow.

03 · local execution

Evidence stays local

Files, shell, git, web, MCP, sub-agents, rollback, and diagnostics are runtime surfaces. The model does not claim verification as a mood; it has to leave evidence.

Built-in providers
DeepSeekNVIDIA NIMOpenAI-compatibleAtlasCloudWanjie ArkVolcengine ArkOpenRouterXiaomi MiMoNovita AIFireworks AISiliconFlowSiliconFlow CNArcee AIMoonshot/KimiSGLangvLLMOllamaHugging FaceTogether AIOpenAI CodexAnthropicZ.aiStepFunMiniMaxDeepInfra
maintainer loop

Long work has to be inheritable

The hard part is not making the model do more at once. It is keeping intent, authority, evidence, and contributor context clear enough for the next maintainer to inherit.

Good agent work is not a pile of activity. It is a reviewable line: request, investigation, patch, verification, credit.

The harness carries the heavy parts: constitutional prompts, structured trust, feedback loops, rollback, and handoffs that survive the session. The model can keep its attention on the task itself.

CodeWhale / review loop
maintainer workflow
01

Intent

The current user request defines the goal, constraints, and success bar.

02

Investigation

The agent reads code, uses tools, delegates when useful, and keeps the evidence visible.

03

Patch

The output stays small, reproducible, testable, and respectful of branch and contributor context.

04

Review

The maintainer sees what changed, what was verified, what risk remains, and who deserves credit.

How it works

Today's Dispatch
2026-06-21
DeepSeek V4-Flash · 6h cron

Quiet release week — install paths and contributor guides up to date.

This dispatch is regenerated by DeepSeek V4-Flash on a six-hour cron. Until the cron fills in fresh repo activity, the static links below are the best starting points.

  • shipped
    Install

    npm, Cargo, Homebrew, or a direct download.

  • opened
    Open issues

    Start with anything labelled good first issue.

  • discussion
    Roadmap

    What is confirmed, what is being weighed, what is ruled out.

v0.8.63 credits

Every patch and report counts

Join in

A small project. Your patch matters.

No CLA. No sponsor lockouts. The maintainer reads everything personally, issues are triaged in the open, and releases cut from main.

version0.8.63
providers25
licenseMIT