LM Studio + OpenCode Backup Profiles

This folder backs up local OpenCode agentic coding profiles. The default profile is still the 48 GB Apple Silicon MacBook Pro LM Studio setup. The Ubuntu RTX 3090 Lucebox setup is opt-in and must be selected with the profile/config environment variables shown below.

Default active profile:

• LM Studio local server at 127.0.0.1:1234
• Fast default OpenCode model: lmstudio/local-fast
• Qwen coding OpenCode model: lmstudio/local-coder
• Fast model source: mlx-community/Llama-3.2-3B-Instruct-4bit, loaded as local-fast
• Chat/coding model source: froggeric/qwen3.6-27b-mlx-8bit, loaded as local-coder
• Local model path: froggeric/Qwen3.6-27B-MLX-8bit
• MLX/safetensors 8-bit, about 28 GB, about 35 GB minimum system memory
• OpenCode Qwen context/output: 32768 / 4096
• OpenCode fast context/output: 32768 / 1024
• Embedding model: text-embedding-nomic-embed-text-v1.5
• OpenCode MCPs: context7, gh_grep, local_code_index, local_dev_tools
• Default agent: fast, a no-tool quick-answer route that prevents local Qwen from hanging on trivial turns
• Qwen agents: build, plan, indexer, codebase-researcher, debugger, test-runner, code-reviewer, doc-researcher, security-auditor
• Slash commands: /explain, /research, /debug, /test, /review, /docs, /security, /index, /implement
• LSP entries: typescript, eslint

Model weights are not copied here. The repo stores model identifiers, runtime package metadata, scripts, MCPs, and docs; LM Studio and Lucebox model directories remain the places that own model downloads and updates.

Source Of Truth

Default Mac profile variables live in:

/Users/oiqbal/AndroidStudioProjects/llm-opencode/config/profile-48gb.env

Edit that file first when changing the target username/home, chat model, model path, context, output tokens, LM Studio host/port, embedding model, OpenCode paths, RAG database, launch env, LSP commands, or LM Studio MLX runtime package URLs/checksums.

Then update config/opencode.json to match and run:

/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/validate-profile-sync.sh

The old 24 GB profile is preserved as reference at config/profile-24gb.env and reference/PLAN-24GB-legacy.md.

An opt-in Ubuntu/NVIDIA profile is available for an i9 13th gen, 96 GB RAM, RTX 3090 workstation:

• Setup plan: PLAN-96GB-UBUNTU-NVIDIA.md
• Profile: config/profile-96gb-ubuntu-nvidia.env
• OpenCode config: config/opencode-96gb-ubuntu-nvidia.json
• Main model runtime: Lucebox DFlash through an autowake proxy on 127.0.0.1:18080
• Main model: lucebox/luce-dflash backed by Qwen3.6-27B-Q4_K_M.gguf + dflash-draft-3.6-q4_k_m.gguf
• Rollback model: lmstudio/local-coder, kept configured but not auto-loaded in normal mode
• Embeddings: LM Studio text-embedding-nomic-embed-text-v1.5 on 127.0.0.1:1234, loaded with --gpu off --ttl 3600
• Default Lucebox context: 49152
• Default Lucebox output cap: 2048
• Lucebox prefix cache: disabled with --prefix-cache-slots 0 for stability
• Idle behavior: the proxy starts the real Lucebox backend on demand at 127.0.0.1:18081 and stops it after 1 hour without API traffic
• Request guard: the proxy clamps oversized chat max_tokens values to the configured Lucebox output cap

Use it explicitly with:

OPENCODE_BACKUP_PROFILE="$PWD/config/profile-96gb-ubuntu-nvidia.env" \
OPENCODE_BACKUP_CONFIG="$PWD/config/opencode-96gb-ubuntu-nvidia.json" \
scripts/validate-profile-sync.sh

The default backup remains the 48 GB Mac profile.

Lessons Baked In

• Do not manually swap the loaded LM Studio model for OpenCode. The setup requires stable aliases: local-fast, local-coder, and the embedding model must all be visible from /v1/models.
• Qwen stays the protected 48 GB local-coder implementation model, but it is no longer the default interactive agent. The default fast route handles short no-tool turns quickly.
• The build agent denies MCP tool namespaces by default to avoid repeated local_code_index and local_dev_tools loops. Use /research for MCP-backed discovery and /implement or Build for Qwen-backed edits.
• Qwen must receive the <|think_off|> prelude from config/qwen36-instructions.md; /no_think was not enough for this LM Studio template.
• /explain is intentionally fast and no-tool. For exact repository evidence, use /research.

Restore On Ubuntu RTX 3090

Follow the full runbook in PLAN-96GB-UBUNTU-NVIDIA.md. The short path is:

cd /home/ohmz/StudioProjects/llm-opencode
export OPENCODE_BACKUP_PROFILE="$PWD/config/profile-96gb-ubuntu-nvidia.env"
export OPENCODE_BACKUP_CONFIG="$PWD/config/opencode-96gb-ubuntu-nvidia.json"

scripts/install-opencode-config-linux.sh
scripts/ensure-lucebox-linux.sh
scripts/install-lucebox-service-linux.sh
scripts/ensure-lmstudio-models-linux.sh

Useful health checks:

curl -sS http://127.0.0.1:18080/health | jq .
systemctl --user status lucebox-dflash-proxy.service lucebox-dflash.service --no-pager
lms ps
opencode run --model lucebox/luce-dflash --format json 'Reply exactly: OK'

Normal Ubuntu mode should show only the embedding model in LM Studio. The LM Studio Qwen chat model is rollback-only and should be loaded only with LMSTUDIO_LOAD_CHAT_ROLLBACK=1.

Restore On The 48 GB Mac

/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/install-opencode-config.sh
/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/ensure-lmstudio-models.sh
/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/seed-rag.sh

Restart OpenCode Desktop after installing the config.

If OpenCode is stuck in a long running turn, quit it before reinstalling. Then run ensure-lmstudio-models.sh and verify that /v1/models includes local-fast, local-coder, and text-embedding-nomic-embed-text-v1.5 before reopening OpenCode.

If LM Studio's MLX runtime is missing its signed CPython vendor layer, run:

/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/repair-lmstudio-mlx-runtime.sh

Validate

Static validation, useful before running on a 48 GB host:

jq . /Users/oiqbal/AndroidStudioProjects/llm-opencode/config/opencode.json >/dev/null
/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/validate-profile-sync.sh
SMOKE_SKIP_HARDWARE=1 SMOKE_SKIP_LMSTUDIO=1 SMOKE_SKIP_LAUNCHCTL=1 /Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/smoke-test.sh

Full validation on the 48 GB Mac:

/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/smoke-test.sh

The full smoke test follows the active profile. On the Mac profile it loads the 32K local-fast and local-coder aliases, verifies fast no-tool generation, and keeps direct Qwen generation opt-in with SMOKE_QWEN_GENERATION=1. On the Ubuntu Lucebox profile it checks the Lucebox autowake proxy and chat path, keeps LM Studio embedding-only, exercises local MCPs, checks remote MCP reachability, and validates the OpenCode agent/command/LSP shape.

For the Ubuntu RTX profile, always pass the profile and config explicitly:

OPENCODE_BACKUP_PROFILE="$PWD/config/profile-96gb-ubuntu-nvidia.env" \
OPENCODE_BACKUP_CONFIG="$PWD/config/opencode-96gb-ubuntu-nvidia.json" \
scripts/smoke-test.sh

Ubuntu RTX Troubleshooting

• prompt + max_tokens exceeds context window: the Ubuntu profile should be 49152 context and 2048 output. Reinstall with scripts/install-opencode-config-linux.sh and restart the stuck OpenCode session if the TUI already received the error.
• OpenCode appears stuck, GPU usage drops, and Lucebox logs show snapshot_longer_than_prompt: prefix cache must be disabled. Confirm the backend command includes --prefix-cache-slots 0, then reinstall/restart with scripts/install-lucebox-service-linux.sh and retry the OpenCode prompt.
• First prompt after idle is slow: expected. /health does not wake the model, but /v1/* and /props start lucebox-dflash.service and reload the model onto the RTX 3090.
• LM Studio is using VRAM in normal mode: run scripts/ensure-lmstudio-models-linux.sh. It unloads local-coder unless LMSTUDIO_LOAD_CHAT_ROLLBACK=1 is set and reloads only the embedding model with --gpu off --ttl 3600.
• Local indexing is noisy: set OPENCODE_INDEX_BACKGROUND=0 in the active profile to disable background indexing, or increase OPENCODE_INDEX_BACKGROUND_SECONDS and OPENCODE_INDEX_AUTO_SYNC_SECONDS above 300.
• Repeated shell warning ohmzai: syntax error: unexpected end of file is an unrelated shell environment issue. It is noisy but did not block the OpenCode/Lucebox validation.

Update This Backup From A 48 GB Live Setup

/Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/backup-current-setup.sh

That command refuses to copy a live OpenCode config whose model does not match the active profile unless BACKUP_ALLOW_PROFILE_MISMATCH=1 is set.