Spaces:
Running
Running
Commit History
-
cors_proxy: allow loopback origins for local development
0f0c10f
-
Add Dyad-style preview backend
231ed61 -
Sync latest from GitHub 8c6d660: llama.cpp, registry routing
142602d
-
Restore static Space: frontend/ at root (undo wrong-repo push)
7751dab
-
Per-node triage incidents, CODE audit tab, GPU meter; remove MASTERING panel
CRANE +1: Claude Sonnet 5.5- Triage: one incident (own clock, error code, log) per red node; never reset - New CODE tab: read-only file tree + editor, generated pipeline code, triage log - Backend: /api/reports/triage, /api/source (secrets redacted) - GPU meter with vendor, uptime and cost; collapsible node inspector - Remove MASTERING panel (AvatarFace extracted); IPS Mirror Loop client-side - scripts/start_crane_api.sh loads ~/.hermes/.env before starting the API - HANDOFF: exact problem statement and open items Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>01eb548 -
fix: key script reads from /dev/tty so it always stops and asks
CRANE +1: Claude Sonnet 5.5Reading stdin swallowed pasted lines; now reads the keyboard directly, rejects placeholder text, re-asks until a real value is entered. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>13edc41 -
fix: args-based key entry, no interactive prompts
CRANE +1: Claude Sonnet 4.6Script now takes keys as positional args instead of read prompts — interactive read breaks in Claude Code terminal and non-TTY shells. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3c775be -
fix: remove pipefail, bulletproof all subcommands in enterprise key script
CRANE +1: Claude Sonnet 4.6- Removed set -euo pipefail (was silently killing script on any failed grep/curl) - Replaced source with safe while-read loop for env file loading - Every curl/python3 call now has || true or explicit error handling - read -rs protected with || fallback - All steps soft-fail: script always runs to completion Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1aa1da0 -
feat: enterprise NGC + HuggingFace key setup script
CRANE +1: Claude Sonnet 4.6- Asks [1/2] NGC Enterprise key, [2/2] HF token — one at a time - Auto-detects NGC key format (nvapi / hex / oauth / user:pass) - Tests enterprise NIM inference + pulls full enterprise model catalog - Flags exclusive models (Nemotron-4 340B, Cosmos, Edify, Parakeet) - Writes nim_enterprise_pipelines.json with enterprise-tier entries - HF: tests whoami, shows account name + plan (Free/Pro) - Patches opencode.jsonc HF apiKey if changed - Stores all in vault if CRANE server is running - Prints full key status table at end Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1ab3b60 -
fix: collect all credentials upfront, fix tokkio NGC key prompt
CRANE +1: Claude Sonnet 4.6setup_ngc_nim.sh — full rewrite to batch credential collection: - Prints one clear header: "Paste all three tokens now — then sit back" - [1/3] NGC API key, [2/3] Hostinger email, [3/3] SMTP password, [opt] API key - ALL setup/test/export steps run only AFTER all inputs are collected - No mid-run interruptions setup_tokkio_ace.sh: - No longer exits with error if NGC_API_KEY missing - Prompts for key inline + saves to ~/.hermes/.env so re-runs don't re-prompt Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a5aa04e -
feat: NGC NIM + Tokkio ACE + Hostinger email agent scripts
CRANE +1: Claude Sonnet 4.6setup_ngc_nim.sh: - Stores NGC API key in ~/.hermes/.env + CRANE vault - Smoke-tests NIM chat endpoint (Nemotron 70B) + streaming SSE - Lists available NIM models - Writes nim_pipelines.json (Nemotron 70B, Llama 3.1 8B, Mistral Nemo 12B) - Hostinger SMTP/IMAP setup: stores creds, tests SMTP, tests API - Writes hostinger_email.json with agent task registry - Stores email creds in vault setup_tokkio_ace.sh: - --nim-only: registers full Tokkio NIM cloud stack (ASR/Dialog/TTS/A2F-2D) - --local: NGC Docker login, clones NVIDIA Digital Human blueprint, docker compose up, writes tokkio_local.json - Both modes write tokkio_nim/local.json with latency budgets - Baseline: Tokkio NIM = 1450ms total; CRANE target to beat = <900ms Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>064879a -
feat: JEV integration, CRANE Engineers, MultiSuite TipsOverlay, Eve rename
CRANE +1: Claude Sonnet 4.6JEV / TypeSafe System One (8 strategic nodes): - jev.py: perceive, performance, proprioception, judge_turn, pick_speaker, route_task, rerank/retrieve_ranked, memory_worth/remember, triage_log, review_verdict; circuit breaker (3 fails → 60s), LRU cache, sem(6), fail-open - main.py: _HoldDone proxy holds WS "done" until JEV perf+proprio complete; AVATAR_FACTS for proprioception; jev_perception/jev_perf/jev_proprio WS events - chat.py: allow_tools gate, rerank flag, style_hint injection - multiavatar.py: jev floor policy, per-turn non-blocking judge, avg_jev in summary - vault.py: _test_jev now pings real TypeSafe endpoint (not HF) CRANE Engineers (4-role disciplined team): - engineers.py: architect/engineer/verifier/reviewer roles, METHOD string, 3-round loop with JEV routing/verdict gates; /ws/engineers WebSocket - scripts/sync_engineers_to_opencode.py: exports roles as OpenCode subagent files Avatar face choreography (MasteringPanel.tsx): - Cues type, EMO_BASE, AvatarFace dri…829b495 -
feat: Multi-Avatar Testing Suite with Tips overlay
CRANE +1: Claude Sonnet 4.6- multiavatar.py: WebSocket engine with 1-4 seats, converse/arena modes, round-robin/random/mention floor policies, barge-in, loop guard (3×60% repetition), per-pipeline asyncio.Semaphore, quality scoring, JSON run records, headless run_headless() API - MultiSuitePanel.tsx: full React UI — cluster toggle (shared/isolated), pipeline health chips, per-seat AvatarFace tiles with TTFT/tok/s metrics, arena leaderboard with EXPORT RUN, host interjection box - TipsOverlay.tsx: 7-topic side-nav docs overlay; TipButton ? chips placed at every friction point (header, cluster, modes, floor, persona, results, pipelines); Esc / click-outside to close - App.tsx + MasteringPanel.tsx: wired MULTI-SUITE nav tab and exported AvatarFace - tests/test_multiavatar.py: 10 test groups, all pass - tests/scenarios/: head_to_head (arena) and four_way_conversation (Beryl/Eve/Nova/Sage) - scripts/run_scenario.py: headless CLI scenario runner with leaderboard output - Vite build clean (tsc + bundle) …751a6da -
Kaggle GPU v2: Qwen3-Coder 30B agent, watchdog, ops helper
CRANE +1: Claude Sonnet 4.6- kaggle_ollama_server.ipynb v2: zstd fix, qwen3-coder:30b as crane-agent (32k ctx, tool_call), dolphin3:8b as crane-chat; OLLAMA_KV_CACHE_TYPE=q8_0 + OLLAMA_SCHED_SPREAD for 2xT4; Cell 5 watchdog restarts Ollama/tunnel - kaggle_ops.sh: push/status/url/sync commands — all driven from the API token - kaggle_set_url.sh: patches opencode.jsonc with new tunnel URL in one cmd Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>ee94d41 -
Fix Kaggle notebook: install zstd before Ollama, fail loudly on install error; add Kaggle token setup script
CRANE +1: Claude Sonnet 5.5Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>12bb21a -
Add Kaggle Ollama tunnel notebook for remote GPU access
CRANE +1: Claude Sonnet 4.6Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>0d1b384 -
Update handoff: USB pool, OpenCode migration, BitNet, system cleanup
CRANE +1: Claude Sonnet 4.6- HANDOFF.md: full session 2 update — system map corrected, USB pool documented (85GB LVM across 4 drives at /mnt/usbpool), Ollama install pending with models going to USB, Heretic abliteration noted for non-uncensored models, BitNet setup instructions added - setup_usbpool.sh: LVM linear pool script (new file) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>41f7514 -
Add CRANE MCP server, CodeRAG routes, and Ollama model setup script
CRANE +1: Claude Sonnet 4.6- mcp_server.py: SSE-based MCP server exposing Second Brain tools (search_knowledge, read_note, knowledge_status) to OpenCode and any MCP client via GET/POST /mcp - main.py: include coderag + mcp_server routers; add /ws/rag-chat and /ws/avatar-chat WebSocket endpoints; proxy /api/services/speaches/tts - setup_ollama_models.sh: pulls dolphin3:8b (uncensored lead) and three CodeRAG models (qwen2.5-coder:3b, phi3.5, starcoder2:3b) via Ollama Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>d1a164f -
Add OpenCode config + connect_berylize.sh; migrate from Hermes
CRANE +1: Claude Sonnet 4.6- scripts/connect_berylize.sh: one-shot tunnel script — kills stale tunnels, opens 8010+8013, starts vLLM + speaches if down - ~/.config/opencode/opencode.jsonc: berylize (14B vLLM) + coderag (Ollama 3B models) + CRANE MCP server configured - HANDOFF.md: documents Hermes→OpenCode migration and GPU fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>accdf24 -
Add setup_jev.sh — one-shot JEV pairing script
CRANE +1: Claude Sonnet 4.6Asks for API key once (hidden input), then: - writes to ~/.hermes/.env (Hermes credential store) - injects connection block into ~/.hermes/config.yaml - saves to ~/.crane_vault.env - exports in ~/.bashrc, ~/.zshrc, ~/.profile - restarts Hermes gateway Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>70aa25a -
Add Mastering Suite: nav bar, live talking avatar, /ws/avatar-chat + speaches TTS
CRANE +1: Claude Sonnet 4.6- App.tsx: replace header segmented-control with crane-nav bar (BUILD / MASTERING / BACKEND / REPORTS tabs); BUILD tab retains view toggles / mode / run controls - MasteringPanel.tsx: animated SVG avatar with idle/listening/thinking/speaking states, mouth lip-sync driven by audio amplitude, browser SpeechRecognition STT, SpeechSynthesis + Kokoro + speaches TTS hierarchy - styles.css: crane-nav, mastering-overlay, avatar animations, audio-bar, mic-pulse keyframes - main.py: /ws/avatar-chat dedicated WebSocket (isolated history for avatar), /api/services/speaches/tts proxy (port 8013) - gpu.py: idle_timeout 0 = disabled; GPU VRAM activity resets idle clock - scripts/start_berylize_node.sh: updated to vllm serve (removes deprecation blocks) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>527ba4c -
feat: add Berylize node startup script for AWQ model
CRANE +1: Claude Sonnet 4.6Updated GPU node setup to use Qwen/Qwen2.5-Coder-32B-Instruct-AWQ (INT4 AWQ, ~18 GB VRAM) instead of FP16 32B which exceeds the L4's 23 GB. The AWQ weights have been downloaded to the nvme0n2 extra-storage disk. A fresh vllm-berylize venv is being set up alongside. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b39a1ea -
feat: Second Brain (Obsidian vault + BM25/link-graph retrieval), Berylize persona, Berylize Creatives rename
CRANE +1: Claude Sonnet 5.5- knowledge/ is now an Obsidian vault: 25 HF papers, 10 GitHub projects (+5 honorable), 24 concepts, 5 playbooks, persona/expertise notes, user research - knowledge.py: heading chunks, BM25 + wikilink graph expansion, off-topic gating; /api/knowledge/* - chat.py: Berylize persona + SECOND BRAIN context injection; fixes KeyError that broke every chat turn - BackendPanel: Second Brain React Flow node (recall edge to Berylize), MiniMax shown as Berylize Creatives - phase2: correct mislabeled prototype file Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>a133806