Code Repo: https://github.com/dhanuka84/local-ai-development-platform/tree/main
Large repositories expose the weakness of a purely conversational coding assistant: the model can read only a fraction of the code at once, and a plausible answer is not necessarily grounded in the branch actually being changed. This article builds a different workflow on a single NVIDIA GB10 workstation. We will start the platform locally, index the main branch of the public upstream Flowable Engine repository, retrieve revision-specific code context, and use a local Ollama model through Codex to explore a BPMN 2.0 designer UI.
The result is a practical local-hybrid loop:
- Ollama performs local model inference and creates embeddings.
- PostgreSQL stores the authoritative repository catalog, active code graph, provenance, workflow state, and indexing outbox.
- Milvus discovers semantically relevant code entities.
- Apache AGE can project active graph topology when available; recursive PostgreSQL traversal remains the safe fallback.
- Codex supplies the interactive coding interface.
- MCP provides a typed boundary for health, retrieval, graph traversal, and governed writes.
- Cloud review is conditional and policy-controlled. An allowed RAG miss uses read-only Codex review; a strong approved RAG hit or protected-data task stays local. It is never a silent fallback from Ollama.
This is an engineering walkthrough, not a claim that the prototype shown here is a production-ready Flowable designer. The useful outcome is the repeatable method: establish exact repository identity, retrieve grounded context, make a bounded change, and validate it without quietly changing model or trust boundaries.
What “hybrid” means in this workflow
The word hybrid is overloaded, so it helps to be precise.
The default local path combines local inference with several deterministic data systems. A question is embedded by Ollama, Milvus finds semantic seeds, PostgreSQL hydrates authoritative records, and the exact graph expands around those seeds. The answer is still produced by one local model. A separate cloud Codex review is conditional: a policy-allowed RAG miss enters the read-only cloud lane only after selecting and sanitizing the context that may leave the workstation. A strong approved RAG hit and protected-data work stay local.
Codex CLI supports connecting a local client directly to an MCP server, and its TUI exposes active servers through /mcp, as described in the official OpenAI MCP documentation. That connection does not, by itself, prove that every open-weight model will select every custom MCP tool correctly. In the tested configuration, qwen3.6:35b could connect to the MCP server and invoke generic MCP discovery, but it did not reliably select the custom platform_status tool. For that reason this walkthrough uses deterministic make mcp-call retrieval before starting the local Codex session. OpenClaw remains the durable controller for queued tasks and invokes cloud Codex only on the governed review route.
That limitation is important: this is a working local-hybrid engineering pipeline, not silent local/cloud model mixing and not a claim of universal tool-calling compatibility.
Why Milvus and PostgreSQL/AGE, not Qdrant or Neo4j?
This project uses Milvus rather than Qdrant because the current platform is already built around Milvus collections, PostgreSQL UUID hydration, local Ollama embeddings, and an asynchronous indexing worker. Milvus is the semantic discovery layer; it is not the authority.
The graph authority is PostgreSQL. Apache AGE is used as a rebuildable Cypher projection when available, while recursive PostgreSQL traversal remains the safe fallback. That means Neo4j is not required for the local deployment. PostgreSQL answers what is official, approved, revision-specific, and current. Milvus answers which approved item or code entity is semantically similar to a question. AGE helps answer what is connected to what.
Keeping those responsibilities separate is the core trust rule: semantic search can suggest where to look, but the platform hydrates results from PostgreSQL before it follows relationships or asks an agent to change code.
Hardware and software used
The measured run used an ASUS/NVIDIA GB10 workstation with 128 GB unified memory. The platform itself is containerized and multi-architecture, but GPU mode requires a working NVIDIA driver and NVIDIA Container Toolkit.
The host needs:
- Docker Engine with Compose v2;
- Git and GitHub CLI where organization synchronization is required;
- Codex CLI;
- sufficient disk for source checkouts, container images, Maven dependencies, PostgreSQL, Milvus, and Ollama models;
- a clean Git checkout for reproducible indexing.
All operational examples below use checked-in Make targets. Run them from the local-ai-development-platform repository root. From its parent directory, use the equivalent make -C local-ai-development-platform <target> form.
1. Create protected local configuration
Create .env without overwriting an existing configuration:
make env-initReview .env before starting the stack. In particular:
- set
CODEGRAPH_HOST_ROOTto the narrow parent directory containing the repository checkout; - keep
CODEGRAPH_ENABLED=true; - keep the generated authentication and database secrets private;
- confirm
LOCAL_CHAT_MODEL=qwen3.6:35bandOLLAMA_EMBEDDING_MODEL=embeddinggemmaunless deliberately testing another compatible model; - size
CODEGRAPH_MAX_FILES,CODEGRAPH_MAX_ENTITIES, andCODEGRAPH_MAX_RELATIONSfor the repository being analyzed.
The code root is mounted read-only at /workspace inside the analyzer-enabled gateway. Compiler-backed JVM analysis runs from a disposable copy and uses a separate dependency cache.
Validate configuration and host prerequisites:
make mcp-preflight
make preflight2. Start the local GPU platform
Start the NVIDIA profile while retaining existing volumes:
make up-gpuPull the larger local coding model separately; the initial stack start pulls the embedding model:
make pull-local-model
make models-listVerify the gateway and its dependencies:
make mcp-status
make doctor
make mcp-call MCP_TOOL=platform_status MCP_ARGUMENTS='{}'The expected platform_status result reports PostgreSQL, Ollama, Milvus, and the graph backend health. In the default AGE profile this includes apache-age; when Cerbos dependency reporting is enabled, it also includes cerbos. AGE may intentionally use PostgreSQL fallback when the extension is unavailable; PostgreSQL remains authoritative either way.
Finally, verify the local Codex/Ollama route:
make codex-local-check
make codex-local-smokeThe smoke test proves that Codex is using the configured Ollama provider. It does not prove that the model can choose every custom MCP tool, which is why the MCP check above is separate.
3. Prepare the upstream Flowable Engine main checkout
Use the original public Flowable Engine repository and its main branch. Pinning the analyzed commit still makes every retrieval and audit record reproducible even as upstream main advances.
The single-repository synchronization target clones a missing checkout or fast-forwards an existing clean checkout. It refuses a dirty tree, a different origin URL, and an unsafe branch name:
make repository-sync-one \
REPO=/home/<user>/projects/open-source/flowable-engine \
REPOSITORY_URL=https://github.com/flowable/flowable-engine.git \
REPOSITORY_BRANCH=mainPoint CODEGRAPH_HOST_ROOT at /home/<user>/projects/open-source for that layout, then recreate the GPU stack with make up-gpu so Compose applies the mount.
The catalog keeps three facts distinct:
| Fact | Canonical field | Meaning |
|---|---|---|
| Forge default branch | default_branch | Remote catalog metadata, such as main. |
| Analysis branch | branch | The checked-out branch actually scanned, here main. |
| Analysis commit | revision | The exact full Git commit analyzed. |
The concepts sometimes called branch_name and git_commit are therefore present, but the implemented API and SQL names are branch and revision.
4. Scan and index one repository
Run the complete single-repository workflow:
make repository-index-one-all \
REPO=/home/<user>/projects/open-source/flowable-engine \
REPOSITORY_PROJECT=local-developmentThe target performs four bounded operations:
- resolves and validates the real checkout path below
CODEGRAPH_HOST_ROOT; - refuses an uncommitted tree and reads the remote, default branch, active branch, and exact commit;
- invokes
code_repository_indexthrough the authenticated MCP boundary; - waits for asynchronous semantic projection and prints the authoritative active snapshot.
The Java/Kotlin path uses SCIP output from compiler-aware analysis. PostgreSQL commits the analysis run, occurrences, exact relations, and active-head update atomically. Only after that transaction succeeds does the outbox worker embed selected entity summaries and upsert them into Milvus.
The completed command prints a snapshot with this shape:
repository: flowable-engine
catalog branch: main
analysis branch: main
revision: <exact 40-character commit from the checkout>
entities: <count for that commit>
relations: <count for that commit>
analysis time: <measured duration on this workstation>Record the actual values in the run audit; do not copy example counts from a different branch or commit. Network state, Maven caches, model placement, container storage, and the exact revision can change the result substantially.
Verify the snapshot again without reindexing:
make repository-verify-one \
REPO=/home/<user>/projects/open-source/flowable-engine \
REPOSITORY_PROJECT=local-development
make repository-org-queue-statusThe first command should show both the catalog and analysis branches as main, plus the exact commit. The second should show zero pending code projection events before feature work begins.
5. Retrieve feature context before asking a model to edit
The feature question is deliberately broad:
Can we create a BPMN 2.0 designer UI for this repository?
A model can answer “yes” without understanding the repository. The retrieval step turns that vague question into evidence attached to the indexed branch and commit.
Start with semantic symbol discovery:
make mcp-call \
MCP_TOOL=code_symbol_search \
MCP_ARGUMENTS='{"project_id":"local-development","query":"REST deployment process definition BPMN XML repository service","limit":10}'Then request bounded graph context scoped to the repository:
make mcp-call \
MCP_TOOL=graph_context_search \
MCP_ARGUMENTS='{"project_id":"local-development","repository":"flowable-engine","query":"Where should a browser BPMN modeler load, validate, save, and deploy BPMN XML through the REST application?","max_hops":2,"seed_limit":8,"max_nodes":40,"max_edges":80}'Milvus supplies likely starting points; the returned code records and edges are hydrated from the active PostgreSQL snapshot. Semantic similarity is treated as discovery, not proof. Before changing a file, inspect the exact graph around the best symbol:
make mcp-call \
MCP_TOOL=code_graph_get \
MCP_ARGUMENTS='{"project_id":"local-development","repository":"flowable-engine","symbol":"<qualified-symbol-from-search>","depth":2}'The official Flowable documentation describes deployments and process definitions as repository concerns and exposes REST resources for listing definitions and retrieving deployment resources. Its REST endpoints require authenticated access and recommend HTTPS with Basic Authentication. A browser designer should therefore use a same-origin server adapter rather than embedding credentials in JavaScript. See the Flowable REST API documentation.
6. Shape a bounded BPMN designer experiment
For a proof of concept, bpmn-js is a sensible modeling surface. Its official walkthrough describes it as a browser-based BPMN 2.0 viewer and modeler that can be embedded in a web application. See the bpmn-js walkthrough.
The experiment should be split into independently testable pieces:
- Modeling surface — embed the BPMN modeler, create a blank diagram, and import/export BPMN 2.0 XML.
- Application shell — add a toolbar, palette, canvas, properties region, status/error area, keyboard handling, and responsive layout.
- Validation — report XML/import failures and distinguish browser-side modeling checks from server-side engine validation.
- Flowable adapter — list process definitions, retrieve BPMN resources, and deploy new XML through an authenticated same-origin backend.
- Security — keep credentials and tokens out of browser storage, require HTTPS outside loopback development, constrain CORS, enforce CSRF protection, and log accountable deployment actions.
- Quality — cover import/export round trips, keyboard accessibility, representative BPMN constructs, REST failures, and rollback behavior.
A static directory inside a distribution module can demonstrate the UI, but it is not automatically the right product boundary. Retrieval should first confirm how the chosen Flowable application packages static resources, exposes REST endpoints, and applies authentication. A production implementation may belong in a dedicated frontend module with an explicit server adapter instead.
7. Start local Codex in the target repository
Launch Codex with Ollama as the model provider:
make dev-session-local-repo \
REPO=/home/<user>/projects/open-source/flowable-engineThis local session is used after deterministic make mcp-call retrieval has already collected the required graph context. It is not relying on Qwen to choose custom MCP tools inside Codex.
The session should report qwen3.6:35b, not a cloud GPT model. A useful first prompt is:
We are evaluating a BPMN 2.0 designer UI on the indexed main revision.
Use the retrieved code-symbol and graph-context evidence supplied with this
task. Inspect the current repository before proposing files. Design a bounded
proof of concept using bpmn-js with import/export, validation, and a
same-origin Flowable REST adapter. Do not place credentials in browser code.
First return the architecture, affected modules, security constraints,
acceptance criteria, and deterministic validation plan. Do not edit until the
plan is approved.After approving the bounded plan, allow the agent to create only the agreed files. Avoid granting broad permanent command prefixes merely to remove friction; approve the narrow operations needed for this experiment.
The screenshot captures the local coding loop identifying the Java/Spring and REST context, choosing a browser BPMN modeler, and expanding the work into UI, integration, export, validation, and documentation steps. It is evidence of a prototype session, not evidence that all acceptance criteria have passed.
8. Observe local resource use
During the captured run, the GX10 dashboard showed approximately 90% GPU utilization and 49.81 GB of 128 GB unified memory in use:
This is a point-in-time observation rather than a benchmark. The useful operational lesson is that the 35B local model leaves meaningful memory headroom for PostgreSQL, Milvus, Ollama, containers, analyzer caches, and the desktop. GPU saturation during inference is expected; sustained out-of-memory pressure, queue growth, or repeated model eviction is not.
Check platform health during a long session with:
make mcp-status
make repository-org-queue-status9. Validate before calling the feature complete
The prototype is complete only when behavior, integration, and security are verified. At minimum, require:
- a BPMN XML import/export round trip with no semantic loss;
- creation and editing of representative events, tasks, gateways, sequence flows, pools, and lanes;
- deterministic error reporting for invalid XML;
- load and deployment through a server-side authenticated adapter;
- no credentials in source, generated bundles, local storage, screenshots, or captured prompts;
- keyboard navigation and basic accessibility checks;
- documented configuration and rollback;
- repository-native tests at the exact base revision.
For a governed patch, encode the allowed files and exact checks in a work packet, evaluate it, and validate the resulting patch in a disposable clone:
make dev-policy-check \
PACKET=/absolute/path/to/bpmn-designer-work-packet.json
make dev-patch-verify \
PACKET=/absolute/path/to/bpmn-designer-work-packet.json \
PATCH=/absolute/path/to/bpmn-designer.patchThe verifier applies the patch to the declared revision, enforces scope and size limits, and runs the packet’s explicit command-and-argument checks without modifying the original checkout. The lower-level workpacket-evaluate and workpacket-verify targets remain useful aliases, but the role-level development commands make the intended workflow clearer.
10. Add conditional read-only cloud review
Local implementation and cloud review are separate trust lanes. Queue each atomic task with execution_mode=auto (the default). At activation, a strong approved RAG hit skips cloud review. A policy-allowed RAG miss automatically packages only the bounded diff, relevant interfaces, sanitized test output, and specific questions for read-only Codex review. Do not send the entire indexed repository or raw database context merely because the model has a large context window.
Before relying on this route, prove its fail-closed boundaries with:
make hybrid-verifyThe cloud reviewer runs with a read-only sandbox and repository mount. Any useful finding remains evidence until Ollama reproduces it locally, validates it, and an accountable actor approves the generalized lesson.
Queue the feature as three atomic tasks
Create one governed workflow and retain the returned workflow UUID:
make mcp-call \
MCP_TOOL=workflow_run_create \
MCP_ARGUMENTS='{"project_id":"local-development","kind":"software-development","risk":"medium","data_classification":"public","request":"Design, implement, and test a BPMN 2.0 Design Console for the indexed Flowable Engine main revision","idempotency_key":"flowable-main:bpmn-designer:v1"}'Then queue design, implementation, and browser-test work. Replace <workflow-uuid> with the returned value:
make mcp-call \
MCP_TOOL=workflow_task_begin \
MCP_ARGUMENTS='{"workflow_id":"<workflow-uuid>","task_key":"designer-architecture","title":"Design the BPMN 2.0 Design Console architecture","task_type":"design","execution_mode":"auto","rag_query":"validated Flowable BPMN browser designer architecture and same-origin REST adapter","idempotency_key":"flowable-main:bpmn-designer:design"}'
make mcp-call \
MCP_TOOL=workflow_task_begin \
MCP_ARGUMENTS='{"workflow_id":"<workflow-uuid>","task_key":"designer-implementation","title":"Implement the BPMN 2.0 Design Console","task_type":"implementation","execution_mode":"auto","rag_query":"validated bpmn-js Flowable Design Console implementation lessons","idempotency_key":"flowable-main:bpmn-designer:implementation"}'
make mcp-call \
MCP_TOOL=workflow_task_begin \
MCP_ARGUMENTS='{"workflow_id":"<workflow-uuid>","task_key":"designer-ui-tests","title":"Create the Playwright UI automation suite","task_type":"testing","execution_mode":"auto","rag_query":"validated Playwright UI tests for BPMN import export validation and REST failures","idempotency_key":"flowable-main:bpmn-designer:testing"}'PostgreSQL accepts all three submissions. Only the FIFO head activates; the others remain queued rather than being rejected. Each task performs its RAG lookup when it activates, so implementation can reuse the approved design lesson and testing can reuse both earlier lessons.
The enforced checkpoint sequence is:
local_execution
-> cloud_review_required on an allowed RAG miss
-> local_revision_required
-> validation_required
-> promotion_required
-> rag_readback_required
-> completed -> activate next queued taskIn the cloud route, review_record stores the exact Codex response and sanitized context manifest in the artifact store. CLOUD_REVIEW_RECORDED accepts their hashes only when PostgreSQL has a matching row for the same candidate, workflow, provider, and model. Codex cannot revise the pending candidate: Ollama must apply accepted findings, rerun deterministic checks, and record the local revision. Accountable approval and successful Milvus read-back are still required before the next task starts.
Reproduce the platform verification
Run the implementation checks and the disposable fresh-image database test:
make fmt-container
make check-container
make authz-policy-test
make contracts-check
make openclaw-plugin-check
make openclaw-config-check
make integration-test-fresh
make hybrid-verifymake integration-test-fresh pulls fresh pinned PostgreSQL/Apache AGE and Go images, applies every migration to a disposable database, executes the PostgreSQL adapter tests, and removes its containers, network, and data. make hybrid-verify separately proves both model lanes fail closed and writes a hashed JSONL audit containing role, provider/model, repository identity, network destination, input diff hash, and before/after working-tree hashes.
What this experiment demonstrates
The important result is not that an agent can generate HTML and CSS. It is that a local workstation can maintain a revision-specific understanding of a large Java repository and use that evidence in a governed development loop.
The workflow preserves several invariants that ordinary chat-based coding does not:
- repository, branch, and commit attribution are explicit;
- a failed scan cannot replace the active graph;
- exact code relations stay in PostgreSQL even when semantic search is used;
- vector indexing is asynchronous and rebuildable;
- the local model and conditional read-only cloud reviewer are visibly different routes;
- a single user can perform development, QA, product-owner, and operations roles in solo mode while each transition still records its acting role;
- validation evidence, not model confidence, determines whether the change is complete.
Closing thought
“Can we build a BPMN 2.0 designer?” is an easy question to answer with a demo. The harder and more valuable question is whether we can build it against the right branch, through the right integration boundary, with reproducible evidence and without leaking the codebase to an unintended model provider.
On one GB10 workstation, the combination of Ollama, PostgreSQL, Milvus, compiler-aware indexing, MCP, and Codex gets us much closer to that standard. The remaining local-model tool-selection limitation is visible and testable, which is exactly how a trustworthy hybrid architecture should fail: explicitly, without pretending that registration is the same as correct execution.
