v26.323.10

Cerebellum offline diagnostics & uninstall scripts

Cerebellum diagnostics

• startCerebellum() now returns structured { ok, reason } with detailed diagnostics
• TUI displays the actual failure reason (Docker not running, permission denied, container logs, gRPC errors)
• Docker binary detection checks common locations (/usr/bin/docker, /usr/local/bin/docker, /snap/bin/docker) when not in PATH
• Auto-starts Docker service via systemctl start docker when daemon is installed but not running
• Onboarding now starts and enables the Docker service after installation

Uninstall scripts

• uninstall.sh (Linux + macOS): stops container, removes config/npm package, optional --purge for Docker image/volumes
• uninstall.ps1 (Windows): same flow adapted for PowerShell
• Flags: --yes (skip prompts), --purge (remove model weights/checkpoints), --keep-config (preserve ~/.cereworker)

v26.323.9: Instinct pillar — auto fine-tuning & training improvements

Instinct Pillar Improvements

Auto Fine-Tuning via Heartbeat

• Registers a scheduled heartbeat task after Cerebellum connects
• Subscribes to heartbeat stream and auto-triggers training on configured schedule (auto/hourly/daily/weekly)
• The finetune.schedule config option (set during onboarding) is now actually consumed

/finetune Subcommands

• /finetune or /finetune status — show current training status
• /finetune start — manually trigger fine-tuning
• /finetune config — show method + schedule config
• /finetune config method <auto|lora|qlora|full> — change method at runtime
• /finetune config schedule <auto|hourly|daily|weekly> — change schedule at runtime
• /finetune history — list past completed/failed jobs

Agent Fine-Tune Awareness

• System prompt now includes a Fine-Tuning section when enabled
• Agent proactively reports training progress, completion, and failures

Training Data Quality

• Minimum 5-pair threshold before starting training (data preserved for next run)
• Min 10-character length for both instruction and response
• Deduplication against existing pending pairs

Conversation Extractor

• New ConversationExtractor extracts user/assistant message pairs from conversation history
• Combined with memory-based curation for richer training data

Cerebellum Offline Error

• StatusBar shows red bold OFFLINE instead of yellow
• Sticky error message with troubleshooting steps on startup failure
• System prompt instructs agent to treat offline Cerebellum as critical

384 tests passing (15 new)

v26.323.8

Fixes

• Docker permissions: install.sh now ensures the user is in the docker group on every install/upgrade, not just fresh Docker installs. On startup, CereWorker auto-fixes group membership if passwordless sudo is available.
• TUI log leak: Logger no longer writes warn/error JSON to stderr in TUI mode. Log output only goes to stderr in headless mode (cereworker serve).
• Docker pull progress: Onboarding now shows Docker's native pull progress bars instead of a silent wait.

v26.323.7

What's New

Worker Profile

• Onboarding: New profile step — set your worker's name, role, and personality traits
• cereworker configure profile: Update your profile after initial setup
• System prompt: Profile (name, role, traits) injected into the agent's identity

Memory Initialization

• memory_read on a fresh install now auto-creates a starter MEMORY.md instead of returning an error

Simplified Installation

• Removed setup.sh and cereworker setup subcommand
• Docker installation inlined into install.sh and the onboarding wizard
• No more build-essential — not needed since the switch to node:sqlite and Docker Hub

README Updates

• Prerequisites updated to Node.js 22.5.1+
• Documented worker profile, cereworker configure profile, and new tools (editFile, searchFiles, glob, httpFetch, webSearch, browser CDP)

v26.323.6: Add 7 new agent tools

New tools

Builtin (executed directly by AI SDK):

• editFile — find-and-replace partial file editing (exact match, must be unique)
• searchFiles — content search via rg/grep with pure-TS fallback
• glob — file pattern matching via fs.globSync

External (orchestrator-registered, Cerebellum-verified):

• httpFetch — HTTP requests with private URL blocking and response truncation
• webSearch — DuckDuckGo search, no API key required

Browser CDP:

• browserConnect — attach to existing Chrome via Chrome DevTools Protocol
• browserDisconnect — detach without closing Chrome
• Unified session management supporting both launch and connect modes

Config: New tools.http, tools.web, tools.browser config sections.

Internal: Switched to workspace:* protocol for internal dependencies.

Total tools: 27 named + dynamic proxy tools.

v26.323.5: Fix builtin tool execution

Fix "Unknown tool: shell" runtime error. Builtin tools (shell, readFile, etc.) are executed by the AI SDK directly — the stream handler no longer redundantly dispatches them to the orchestrator's tool registry.

v26.323.4

Remove @mariozechner/pi-ai from bundled dependencies (last source of deprecation warnings). OAuth users can install it on demand. npm install -g @cereworker/cli now has zero deprecation warnings.

v26.323.3

Auto-install wechaty during onboarding when WeChat channel is selected. Users who don't use WeChat get zero deprecation warnings.

v26.323.2

Remove wechaty from optionalDependencies (npm installs them by default). Lazy-load node:sqlite to suppress ExperimentalWarning. Zero deprecation warnings on npm install.

v26.323.1

Migrate from better-sqlite3 to Node.js built-in SQLite (node:sqlite). Eliminates the last npm deprecation warning, removes the need for C++ build tools, and simplifies onboarding.