Agent-Field/agentfield

Split out of #620, where the reporter asked for "mypy and ruff for strict checks of async / blocking code". The ruff half shipped in #812 (ASYNC ruleset, enforced in CI on 3.10/3.11/3.12). The mypy half was never adopted: sdk/python/mypy.ini exists but nothing in .github/workflows/sdk-python.yml runs it, and mypy --config-file mypy.ini agentfield/ currently reports ~194 errors across ~28 files. What "done" looks like A mypy step in sdk/python CI that is required to pass. A per-module allowlist ([mypy-agentfield.<module>] with ignore_errors = True) for modules that are not yet clean, so the gate can be turned on immediately and the allowlist shrunk module by module. No # type: ignore sprinkling to get to green — fix the type or allowlist the module. Suggested order Land the CI step with everything allowlisted (zero behaviour change, proves the wiring). Clean the async-heavy modules first, since that is the reporter's concern: run_async.py, lock_utils.py, async_lifecycle.py, result_

Summary Opencode users likely don't want their opencode AGENTS.md injected, this bloats the system prompt for all uses of opencode. A better approach is to use the skills feature as was done with claude code. Proposed solution mkdir -p ~/.config/opencode/skills ln -s ~/.agentfield/skills/agentfield/current ~/.config/opencode/skills/agentfield Alternatives considered none Additional context none

Summary A harness provider needs an external coding-agent to be present at runtime (a CLI binary and/or a Python/JS SDK wrapper, plus auth). Today there is no first-class way to (a) install the providers you intend to use, or (b) verify they are actually installed and usable before you run a real workflow. Failures surface late, mid-run, as a subprocess error inside app.harness(...). This is especially painful in containers/CI where you want the image build to fail fast if a required harness is missing. We want a clean, bloat-free DX for declaring and verifying harness provider availability, consistent across the Python, TypeScript, and Go SDKs. Why "just add pip/npm extras" is not enough The real dependency is usually a binary, not a pip/npm library: Provider Python install TS/JS install Underlying binary claude-code claude-agent-sdk (PyPI) @anthropic-ai/claude-agent-sdk (npm) bundles the Claude Code CLI codex openai-codex (PyPI) or raw CLI @openai/codex-sdk (npm) codex (Rust

Umbrella tracking issue from #654. #654 ships a deliberately provider-agnostic session DX — provider is an explicit, required knob and the transport matrix is multi-provider — but OpenAI WebRTC is the only fully-wired realtime path today. realtime-offer hard-gates to provider=openai+transport=webrtc and posts to api.openai.com/v1/realtime/calls; the declared openrouter: audio_turns pair has no handler. This issue tracks growing sessions into genuinely multi-provider realtime ingress, so AgentField is the provider boundary for more than just OpenAI. Candidate providers / transports OpenRouter (voice-capable models) — tracked in #665. Azure OpenAI Realtime — near drop-in; Azure documents the same SDP-proxy pattern (/openai/v1/realtime/calls, ephemeral via /openai/v1/realtime/client_secrets). Mostly an endpoint + auth swap in createOpenAIRealtimeCall. Google Gemini Live — different protocol (WebSocket bidi streaming, not WebRTC/SDP); needs a dedicated control-plane adapter. Dee

Follow-up from #654 / discussion. The session transport matrix already declares openrouter: {audio_turns}, but audio_turns has no handler anywhere in the control plane — realtime-offer hard-gates to provider=openai+transport=webrtc and posts to api.openai.com/v1/realtime/calls. So today the matrix advertises an OpenRouter capability that does nothing. This issue tracks making OpenRouter a real, working session provider, including its voice-capable models. Why #654 ships a deliberately provider-agnostic session DX (provider is an explicit, required knob) but only OpenAI is actually wired. OpenRouter is the declared second provider; it should either work or not be advertised. OpenRouter brokers many voice/realtime-capable models; supporting it makes AgentField sessions genuinely multi-provider rather than OpenAI-locked. Scope / open design questions Decide what audio_turns actually is (turn-based audio request/response) vs. whether OpenRouter exposes a realtime/WebRTC path, and impl