npx
npx -y clavue@8.22.5 --version
npx -y clavue@8.22.5Clavue 8.22.5 · Planner + Cache Visibility
Inspect, edit, run, verify in one loop. Planner role runs high-capability planning in isolated cache-stable sessions (protects your main DeepSeek/Grok reasoner prefix). Live cache hit % after every compact. Native-perfect tool calling and reasoning for Grok + DeepSeek on /v1/chat/completions.
Install and first run
Run once with npx, install globally for daily use, or pin a release with the installer script.
npx -y clavue@8.22.5 --version
npx -y clavue@8.22.5npm install -g clavue@8.22.5
clavue install --force
clavuecurl -fsSL https://unpkg.com/clavue@8.22.5/install.sh | bash -s -- 8.22.5Clavue Agent SDK 2.2.0
Clavue CLI remains the primary coding product. clavue-agent-sdk is the parallel TypeScript SDK for teams that want to embed agent loops inside their own apps: no daemon, no subprocess, just run(), query(), and createAgent() in the host process. Version 2.2.0 adds synthesis risk-tier routing for synthesized tools and skills.
npm install clavue-agent-sdk@2.2.0
import { createAgent, inferSynthesisRiskTier } from 'clavue-agent-sdk'
import { routeSynthesisCandidates } from 'clavue-agent-sdk/contracts'Chinese software · dragon spirit
The dragon visual system represents focus, strength, continuity, and verified delivery. Planner sessions protect high-model prefix cache for DeepSeek/Grok. Live cache-hit visibility + native chat perfection. Provider compatibility is runtime; Clavue is the product and workflow identity.
Inspect files, search code, edit source, and run checks from the CLI instead of staying in a detached chat surface.
Save profiles, assign model slots (main, haiku, sonnet, planner), validate reachability, and repair drift. planner slot for cache-isolated high-capability planning.
Planner role runs in separate session (no prefix pollution for DeepSeek/Grok reasoners). Every compact surfaces live "缓存命中 XX%" + real post-token count. 65% proactive for 512k+ windows.
Forced to /v1/chat/completions with complete tool_calls (id+type+function, streaming deltas, XML fallback) and reasoning_content → thinking. No lossy anthropic_messages emulation.
Report concrete checks first. 10min compact timeout + emergency truncate guarantee. Micro-compacts, pinned context, and boundary markers keep 200k+ sessions stable and cheap.
Inspect team readiness, active routes, permissions. MCP tools, skills, worktrees, /mao review, /retro all in one verified loop. Reasonix-inspired discipline for cost and stability.
Provider routing
Configure official login, custom API profiles, compatible gateways, and model slots (mainLoop / summarize / planner / smallFast). Planner slot for isolated cache-stable planning. Validate reachability and repair drift without changing your daily coding loop.
/provider
/provider list
/provider current
/provider doctor
/provider validate
/provider repair
/provider use gpt54-mainSet CLAVUE_API_DIALECT=openai_responses only when a gateway explicitly requires the Responses adapter.
Best use cases
Inspect the failing test around provider routing, make the smallest source fix, run the targeted test, then run the relevant verification command before reporting back./mao adversarial-review --base main provider routing and permissionsOpen provider setup, create a saved profile for this API route, validate it, and make it the default if validation passes./team check