What is this?
CodeMason (编舟) is an autonomous coding agent in the same class as Claude Code or Cline: it understands a requirement, plans, edits code, verifies the result and delivers it, running the whole loop on its own. It is built around the idea that state is never saved — everything is derived from an append-only event log, so every change can be replayed, rolled back or audited.
CodeMason (编舟) is an autonomous coding agent in the same class as Claude Code or Cline: it understands a requirement, plans, edits code, verifies the result and delivers it, running the whole loop on its own. It is built around the idea that state is never saved — everything is derived from an append-only event log, so every change can be replayed, rolled back or audited.Is this a CLI agent like Claude Code, or something else?It is a self-hosted web application, not a single CLI binary: a Python FastAPI + WebSocket backend drives a React 18 'cockpit' dashboard with panes for the event timeline, approval inbox, YAGNI board, cost/context view and a read-only code viewer.
It is a self-hosted web application, not a single CLI binary: a Python FastAPI + WebSocket backend drives a React 18 'cockpit' dashboard with panes for the event timeline, approval inbox, YAGNI board, cost/context view and a read-only code viewer.How do I get it running?The README's Quick Start points to a Docker deployment, so you stand it up with Docker/Compose rather than installing it manually.
The README's Quick Start points to a Docker deployment, so you stand it up with Docker/Compose rather than installing it manually.Do I need my own LLM API key?Yes — CodeMason defaults to iFlytek (讯飞) as its model provider, but a pluggable Provider abstraction lets you swap in OpenAI or Ollama instead, with separate 'architect' (planning) and 'editor' (execution) model roles.
Yes — CodeMason defaults to iFlytek (讯飞) as its model provider, but a pluggable Provider abstraction lets you swap in OpenAI or Ollama instead, with separate 'architect' (planning) and 'editor' (execution) model roles.How reliable is the safety classifier that blocks dangerous actions?On its own 100-command test set it blocks about 89% of dangerous commands and reports a 0% prompt-injection success rate across 720 scenarios, with a false-positive rate under 1% — not perfect, so three consecutive blocks automatically escalate the task to a human approver.
On its own 100-command test set it blocks about 89% of dangerous commands and reports a 0% prompt-injection success rate across 720 scenarios, with a false-positive rate under 1% — not perfect, so three consecutive blocks automatically escalate the task to a human approver.What does the YAGNI engine actually check?After each generation, an independent static-analysis hook runs the staged diff through a seven-level decision ladder (do you really need it, is it in a library, can the stdlib do it, etc.) and reports lines reduced, dependencies avoided, duplicate implementations and a readability score.
After each generation, an independent static-analysis hook runs the staged diff through a seven-level decision ladder (do you really need it, is it in a library, can the stdlib do it, etc.) and reports lines reduced, dependencies avoided, duplicate implementations and a readability score.What sandbox backends are available for running code?A SandboxProvider abstraction supports four interchangeable backends: Firecracker microVM (default, isolated kernel + KVM), a hardened Docker container, gVisor (runsc) as a middle tier, or E2B's cloud sandbox.
A SandboxProvider abstraction supports four interchangeable backends: Firecracker microVM (default, isolated kernel + KVM), a hardened Docker container, gVisor (runsc) as a middle tier, or E2B's cloud sandbox.How does CodeMason's performance get measured against Cline?The same model is run in Cline on an identical 30-task benchmark (bug fixes, features, refactors, tests adapted from SWE-bench and real issues), and results are compared using fail-to-pass rate plus TRAJEVAL's tool-level precision/recall breakdown so only the agent implementation varies.
The same model is run in Cline on an identical 30-task benchmark (bug fixes, features, refactors, tests adapted from SWE-bench and real issues), and results are compared using fail-to-pass rate plus TRAJEVAL's tool-level precision/recall breakdown so only the agent implementation varies.What is the self-evolving engine?A closed loop (Observe→Analyze→Improve→Verify→Persist) that improves memory, skills, planning, tool usage and the harness itself over time, governed by a rate-limiting EvolutionPolicy with staged shadow→canary→full rollout and automatic rollback.
A closed loop (Observe→Analyze→Improve→Verify→Persist) that improves memory, skills, planning, tool usage and the harness itself over time, governed by a rate-limiting EvolutionPolicy with staged shadow→canary→full rollout and automatic rollback.What are similar projects?lean-ctx, utopia, ruflo, Raven, multi-agent-discuss, open-multi-agent, agent-governance-toolkit, openfang
lean-ctx, utopia, ruflo, Raven, multi-agent-discuss, open-multi-agent, agent-governance-toolkit, openfangWhich topics is it in?event-sourcing, self-evolving, tree-sitter, compliance, audit-log, human-in-the-loop, knowledge-graph, fastapi, static-analysis, autonomous-agents, agent-memory, sandbox
event-sourcing, self-evolving, tree-sitter, compliance, audit-log, human-in-the-loop, knowledge-graph, fastapi, static-analysis, autonomous-agents, agent-memory, sandbox