Handbook

LLM providers and Forge Fleet probes

The Studio AI tray consumes the same /api/llm/*, /api/fleet/*, /api/sdlc-copilot/*, and OAuth-like helper routes documented in http-api-and-routes.html. Think of Forge Fleet as discovery + pooling for Forge-hosted…

What it is

The Studio AI tray consumes the same /api/llm/*, /api/fleet/*, /api/sdlc-copilot/*, and OAuth-like helper routes documented in http-api-and-routes.html. Think of Forge Fleet as discovery + pooling for Forge-hosted inference nodes, while the LLM stack tracks provider credentials, catalogs, telemetry, routing previews, and Ollama warm-up scripts.

Loopback-only Studio pane calling Fleet and LLM routes without automatic cloud egress

When to use this guide

  • You flipped LENSES_EXPERIMENTAL_SDLC_COPILOT or Wizard flags but still cannot chat locally.
  • You need to sanity-check Forge Fleet meshes before enrolling Studio clients.
  • You want assurance about what stays on disk (.lenses-local) vs what never leaves 127.0.0.1.

Prerequisites

Requirement Notes
Local Lenses checkout ./scripts/run-lenses.sh on loopback (default 127.0.0.1:8080)
Optional Ollama Set OLLAMA_BASE_URL; fall back to Forge Fleet / hosted models
Optional PAT / OIDC Required when RBAC is enforced (lenses-access.json)

Configure providers (happy path)

  1. Open Forge Studio (/studio/) → Settings → AI.
  2. Pick /api/llm/providers entry that matches your transport (HTTP OpenAI-compat, Forge Fleet pooled node, mocked fixtures for offline demos).
  3. Save via POST /api/llm/settings — response should echo merged JSON with sanitized secrets (tokens never echoed verbatim).
  4. Run POST /api/llm/provider-probe for each profile you rely on nightly.
  5. (Optional Fleet) Populate POST /api/fleet/settings with discovery subnets + bearer references, then hit POST /api/fleet/discover followed by POST /api/fleet/node-detail for each candidate.

Verify success by watching Studio banners fed from GET /api/llm/diagnostics and GET /api/llm/ollama-status.

Next steps: 11-wizard-301_04-cursor-launch-pack.html for packaging outputs, or http-api-and-routes.html for full JSON tables.

Data boundaries (local-first)

Surface Local-only? Notes
Provider secrets Yes Stored under .lenses-local/; never mirrored to Git
Chat transcripts Yes SDLC copilot audit tail sdlc-copilot-audit.jsonl
Fleet discovery cache Mostly Stored alongside LLM blobs; prune via Studio reset
Hosted model traffic Depends Leaves loopback once you intentionally target cloud endpoints

Treat LENSES_ALLOW_ACTIONS=1 as “this host now trusts LAN clients equally with localhost” — do not combine with unsecured Wi-Fi.