Tracker: future adapter opportunities

## Summary

Tracker for evaluating and shipping additional adapters beyond the
built-ins.

Organised by tier: Tier 1 needs a deliberate fit-for-shape evaluation
before scheduling; Tier 2 is mechanical (~30 LOC each via
`OpenAICompatibleRuntime`); Tier 3 is the decline list captured here
so we don't re-evaluate.

## Recently shipped

- [x] **BedrockRuntime** (provider ID `bedrock`) — shipped in
PR #10 (merge commit `148bc74`). Wraps AWS Bedrock's
[Converse API](https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html)
via `aioboto3`. Twelve `Feature` flags True (structured output,
streaming, cancel, reasoning effort + budget on Anthropic-on-Bedrock,
vision, file, function tools, permission, lifecycle hooks, both
budget caps). MCP transports + session resume are permanent
declines (Converse has no wire equivalent). Landed across six
independently-mergeable iterations A–F per
[`dev-docs/bedrock-adapter-plan.md`](../blob/main/dev-docs/bedrock-adapter-plan.md).
Reserved `"bedrock-agents"` for a future sibling wrapping
`bedrock-agent-runtime` (Knowledge Bases, action groups).
- [x] **OpenRouterRuntime** (provider ID `openrouter`) — shipped in
PR #10. Targets `https://openrouter.ai/api/v1`, the
multi-vendor router fronting 200+ models. Vendor-prefixed model IDs
(`anthropic/claude-3.5-sonnet`, `google/gemini-pro-1.5`,
`meta-llama/llama-3.1-70b-instruct`). Per-model feature
heterogeneity documented — silent JSON-schema degradation is the
failure mode on some routes.
- [x] **OpenCodeGoRuntime** (provider ID `opencode-go`) — shipped in
PR #7 (commit `7b6ff98`). Targets `https://opencode.ai/zen/go/v1`
flat-fee subscription endpoint, distinct from the per-token
`OpenCodeZenRuntime` (`opencode-zen`) at `https://opencode.ai/zen/v1`.
14-model subscription catalog with `cost_usd=0.0` per call (flat-fee
at the margin). Revealed a pattern: **same vendor + multiple billing
paths → multiple adapters**.
- [x] **OpenCodeZen rename** — `PROVIDER_ID` changed from bare
`"opencode"` to `"opencode-zen"` for symmetry with `"opencode-go"`.
Pre-v1 cleanup before the asymmetry calcified.

## Tier 1 — substantive evaluation needed

- [ ] **Google GenAI** (`google-genai` SDK)
- Plan: [`dev-docs/google-genai-adapter-plan.md`](../blob/main/dev-docs/google-genai-adapter-plan.md) (6 iterations)
- Provider ID: `google` (or `gemini` — TBD in plan)
- Just needs scheduling
- [ ] **OpenAI Responses API** (`openai.responses.*`)
- Fills the reserved `"openai"` provider ID — direct-API-billed path
- Distinct from `CodexRuntime`, which is the **subscription** path (ChatGPT Plus / Pro via `codex login`) and runs the local sandboxed CLI
- Unlocks `SESSION_RESUME` + `TOOLS_FUNCTION` + `TOOLS_MCP_HTTP` for the OpenAI provider that chat-completions / Codex can't deliver
- Likely a new class, not a subclass of `OpenAICompatibleRuntime` (chat-completions and responses diverge on session shape)
- [ ] **OpenAI Agents SDK** (`openai-agents`)
- Needs a "fits airframe shape?" memo first — same shape question we wrote for Google ADK
- Closer to LangGraph (orchestration framework) than to a vendor runtime
- High mindshare; worth a deliberate yes/no rather than reflexive inclusion
- [ ] **Bedrock Agents** (`bedrock-agent-runtime`)
- **Sibling to the now-shipped `BedrockRuntime`** — provider ID `bedrock-agents` (reserved).
- Different endpoint (`bedrock-agent-runtime` vs `bedrock-runtime`), different request shape (Knowledge Bases, action groups, server-side orchestration).
- Worth shipping once consumer demand for retrieval-grounded Bedrock workflows surfaces.

## Tier 2 — OpenAI-compatible HTTP siblings

Each is a ~30-line subclass of `OpenAICompatibleRuntime`. Good
"first issue" tier for outside contributors via the
[third-party adapter pattern](../blob/main/docs/adapters/third-party.md).

Already named in docs as future siblings:
- [ ] Together AI
- [ ] Groq
- [ ] Fireworks

Additional easy candidates:
- [ ] Mistral (also has a native SDK if we want both)
- [ ] xAI / Grok
- [ ] DeepSeek
- [ ] Perplexity
- [ ] Cloudflare Workers AI
- [ ] NVIDIA NIM
- [ ] HuggingFace TGI / Inference Endpoints

## Tier 3 — declined (wrong shape)

These live *above* `AgentRuntime` (they'd consume airframe, not be
consumed by it). Captured here so we don't re-evaluate.

- **Google ADK** — orchestration framework (evaluated already)
- **LangChain / LangGraph** — orchestration framework
- **AWS Strands Agents** — Bedrock-adjacent orchestration framework
- **HuggingFace Smolagents** — orchestration framework
- **AutoGen / CrewAI** — multi-agent orchestration

## Recommended order

1. **Google GenAI** — plan already exists, ship it
2. **OpenAI Responses** — fills the reserved `"openai"` provider ID; unlocks capabilities Codex can't deliver
3. **OpenAI Agents SDK evaluation memo** — write before committing
4. **Bedrock Agents** — opportunistic, demand-driven sibling to `BedrockRuntime`
5. **Tier 2 siblings** — opportunistic / contributor-driven