Synthetic Philosophy and Deductive Engineering
Task branches and worktrees are in ampd004.md.
A development task is done by a local agent on this DGX Spark, in the worktree and on the branch for that task. One task is in play in a worktree. The agent is told the role, the outcome, and the files or formal target it may change.
The Spark’s services listen on loopback, so access them from another machine through an SSH tunnel. OpenHands Agent Canvas listens on port 3000; the vLLM OpenAI-compatible API listens on port 8000. The two ports serve different clients.
From the client machine, forward both ports in one SSH session:
ssh -N -L 3000:127.0.0.1:3000 -L 8000:127.0.0.1:8000 <user>@spark-95fb
Keep the SSH session open while using either service. Replace <user> with the Spark login; use the hostname or SSH alias configured for the Spark if spark-95fb does not resolve.
Open http://127.0.0.1:3000 in a browser on the client machine. OpenHands is already configured to use the vLLM service on the Spark’s Docker network, so it does not need to connect through the client’s port 8000 tunnel.
VS Code Chat can use the Spark model as a custom endpoint. In VS Code, run Chat: Manage Language Models, choose Add Models, then Custom Endpoint. Select Chat Completions and configure the endpoint as http://127.0.0.1:8000/v1/chat/completions. When prompted for an API key, use local-llm; vLLM does not enforce API-key authentication, but the setup expects a value. Set the model ID to the service’s current ID from http://127.0.0.1:8000/v1/models—currently gpt-oss-120b—and enable tool calling to use the model in agent mode. Then select the added model in Chat’s model picker. After a model switch, update the custom endpoint’s model ID to the new VLLM_SERVED_MODEL_NAME as described in amms010.md.
This configures the chat and agent experience, not Copilot inline code completions. Copilot Business or Enterprise administrators may need to enable BYOK for the organization.
Copilot CLI can use the same vLLM endpoint as a custom OpenAI-compatible provider. Keep the SSH tunnel to port 8000 open, then in the client terminal where you will run Copilot CLI set:
export COPILOT_PROVIDER_BASE_URL=http://127.0.0.1:8000/v1
export COPILOT_PROVIDER_TYPE=openai
export COPILOT_PROVIDER_MODEL_ID=gpt-4.1
export COPILOT_PROVIDER_WIRE_MODEL=gpt-oss-120b
export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=122880
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=8192
copilot
Here, gpt-4.1 is the recognized model profile Copilot CLI uses for its agent configuration; gpt-oss-120b is the actual model ID sent to vLLM. Setting only COPILOT_MODEL=gpt-oss-120b made this CLI send a custom-tool schema that the current vLLM rejects. The explicit model-profile/wire-model pair above was tested with both a completion and a shell tool call. Leave COPILOT_PROVIDER_WIRE_API unset (its default is Chat Completions); the current vLLM does not accept Copilot CLI’s custom tool type through its Responses API either.
The token limits sum to the current vLLM context limit of 131072 tokens. After switching the served model, set COPILOT_PROVIDER_WIRE_MODEL to the ID reported by /v1/models and adjust the token limits to that model’s configured context. vLLM does not require an API key, so COPILOT_PROVIDER_API_KEY can be left unset. Run copilot help providers for provider configuration details.
Grok and Copilot in the cloud are used when a frontier model is likely to do the task better than the local models then available. The project manager applies that test to the task.
Cases that meet it at the start of this arrangement:
main (Copilot code review).The test is reapplied as local models, including models fine-tuned for SPaDE, become able to hold the role. Admin documents are written by the project manager. That role may be filled in a Grok session while the work is still the liaison itself, and delegated to a local agent when the task is bounded.
mainWork that is to land on main is progressed on the task branch in its worktree and submitted as a pull request. Copilot code review is included on that pull request. The principal accepts the merge.
Detail of the review step is in ampd005.md. A pull request authored by the Copilot coding agent is reviewed by the principal. Copilot does not independently review a change it authored.
A bounded specialist task is commissioned as a GitHub Issue. The issue names the role, the outcome, the files or formal target, and the acceptance check. The worker is a local agent, or the Copilot coding agent when the frontier-model test above says so. Copilot delegation, when chosen, follows ampd001.md and ampd008.md.
The issue is the record of the subcontract. The worktree is the workplace. The pull request is the review. The issue links the branch.
Exploratory philosophy, architecture, and strategy are developed in conversation and recorded as tlph, tlad, and ampl documents, following ampd003.md. They are not opened as Issues in order to be discovered.
GitHub Projects, as in ampl003.md, may still group issues. The issue is the unit that is handed to a specialist.
Document ID: ampd009 Author: Grok Build (Grok 4.7) Status: In progress Chat log: amcl003.md