feat: enable GitLab Agent Platform with dynamic model discovery

[vglafirov]

Summary

Enable GitLab Agent Platform (Duo Workflow Service) integration with dynamic model discovery, workflow model routing, and local tool execution.

Changes

Package Migration

• Replaced @gitlab/gitlab-ai-provider v3.6.0 with unscoped gitlab-ai-provider v5.1.2
• Updated all imports, lockfile entries, and test fixtures to reflect the new package name

Dynamic Model Discovery (provider.ts)

• Introduced CustomDiscoverModels type and discoveryLoaders registry so providers can dynamically register models at runtime
• Added discoverModels(providerID) function with a per-provider cache that:
  1. Uses GitLabProjectDetector to detect the current GitLab project from process.cwd()
  2. Resolves the project's namespaceId (group)
  3. Calls GitLabModelDiscovery.discover() to fetch selectable, default, and admin-pinned models for that namespace
  4. Queries GitLabModelConfigRegistry for context/output token limits per model
  5. Registers discovered models as duo-workflow-* entries with full capability metadata (reasoning, attachments, tool calls, image/PDF input)
• Honors admin-pinned models: when a model is pinned, only that model is available
• Falls back gracefully when project detection or discovery fails (logs warnings, returns empty)

Workflow Model Routing (provider.ts)

• GitLab custom model loader now routes duo-workflow-* model IDs through sdk.workflowChat() instead of sdk.agenticChat()
• Supports a workflowRef option for selecting a specific upstream model reference
• Static MODEL_MAPPINGS are respected; unknown refs get a generated duo-workflow-{ref} ID
• Model options are merged and passed through to the model loader ({ ...provider.options, ...model.options })

Workflow Tool Executor (session/llm.ts)

• When the resolved language model is a GitLabWorkflowLanguageModel, a toolExecutor callback is wired up
• Tool calls originating from the Duo Workflow Service are executed locally via opencode's tool system
• Results (including metadata and title) are serialized and sent back over the WebSocket connection
• Errors are caught and returned as structured error responses

Server Routes (config.ts, provider.ts)

• Both /config and /provider routes now trigger Provider.discoverModels() for all connected providers before returning responses
• Discovery errors are caught per-provider and logged as warnings without blocking the response

Feature Flags

• duo_agent_platform_agentic_chat and duo_agent_platform are enabled by default for GitLab provider

Testing

• Updated models-api.json test fixture to reflect the new npm package name (gitlab-ai-provider)

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, I found 1 related PR that should be reviewed:

Related PR:

• feat: integrate GitLab Duo Agent Platform (DAP) workflow models #15573 - feat: integrate GitLab Duo Agent Platform (DAP) workflow models

Why it's related: This PR also addresses GitLab Duo Agent Platform (DAP) workflow model integration. Since #18012 is a more recent implementation with dynamic model discovery and workflow routing, you should verify whether #15573 is already merged, closed, or if there's overlap in the implementation approach that should be consolidated.