feat: add docker agent serve chat command (OpenAI-compatible API)

[dgageot]

Fixes #2502.

Exposes any docker-agent agent through an OpenAI-compatible HTTP server, so any tool that already speaks the Chat Completions protocol (Open WebUI, the official openai SDKs, ad-hoc curl scripts, etc.) can drive an agent without a custom integration.

Endpoints

Method	Path	Notes
GET	/v1/models	Lists exposed agents as OpenAI models
POST	/v1/chat/completions	Runs the agent; supports stream: true (SSE) and false

Usage

docker agent serve chat ./agent.yaml                       # localhost:8083
docker agent serve chat ./team.yaml --agent reviewer       # pin one agent
docker agent serve chat agentcatalog/pirate --listen :9090

curl -sS -X POST http://127.0.0.1:8083/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"hi"}]}'

Design

• The team is loaded once at startup and shared across requests. Each chat completion gets a fresh session and runtime.
• The session is created with ToolsApproved=true and NonInteractive=true — there is no human in the loop. ElicitationRequestEvent is still explicitly declined to avoid hanging on the runtime's elicitation channel.
• The model field of the request can pin a specific agent in a multi-agent team. If it doesn't match an exposed agent (e.g. clients that hard-code gpt-4) we silently fall back to the default agent and echo the requested model name back, so clients matching on the model field stay happy.
• Streaming uses SSE in OpenAI's chat.completion.chunk format and ends with data: [DONE].

Implementation

• New cobra command cmd/root/chat.go (default 127.0.0.1:8083, --agent / --listen flags) wired into cmd/root/serve.go.
• New pkg/chatserver package, split across:
  • server.go — Run, router, HTTP handlers, sseStream, error envelope
  • agent.go — agentPolicy, buildSession, runAgentLoop, sessionUsage
  • types.go — request/response shapes
• Reuses openai.Model from github.com/openai/openai-go/v3 for /v1/models. Other SDK response types serialise too noisily with stdlib encoding/json (the SDK relies on its internal apijson package, which lives under internal/), so the chat-completion shapes are hand-rolled for clean output.

Tests

• Unit tests for session-building, agent-policy resolution, usage extraction.
• HTTP-level tests via httptest for /v1/models shape, the three early-validation paths of /v1/chat/completions (bad JSON, empty messages, history without user), and writeError's status→type mapping.

Validation

• mise lint — 0 issues
• mise test — all packages green
• Manual curl smoke test against examples/42.yaml: /v1/models returns the agent, error paths return correct OpenAI-shaped envelopes.