Add AWS Bedrock provider (Converse + ConverseStream)

[ricardosllm]

Closes #254.

Adds a bedrock provider/API type using AWS Bedrock's native Converse and ConverseStream APIs.

Why

Bedrock's OpenAI-compatible endpoint only serves openai.* models — Claude (and other) inference profiles return 404 there, so the existing OpenAI-compat custom-provider path can't reach them. This adds a dedicated handler. (Also supersedes the partial workaround discussed in #329.)

Scope

• Converse (non-streaming) and ConverseStream (streaming) — includes a binary vnd.amazon.eventstream frame decoder
• Bearer-token auth (Authorization: Bearer <token>) — sidesteps SigV4; pairs with AWS_BEARER_TOKEN_BEDROCK
• Tool calling — toolConfig / toolSpec / toolUse / toolResult blocks
• Reasoning — reasoningContent blocks; reasoning_config via additionalModelRequestFields when reason?
• Region is part of the runtime URL; model ids / inference-profile ids are user-configured
• extraPayload is filtered to Converse-valid top-level keys (verified against the botocore service model) so variant payloads aimed at other APIs don't 400
• Retries on transient statuses reuse the existing error classification; modeled stream error frames terminate the turn cleanly (no spurious premature-stop)

SigV4 auth and a model-catalog endpoint are intentionally out of scope for this first cut.

Config

{
  "providers": {
    "bedrock": {
      "api": "bedrock",
      "url": "https://bedrock-runtime.us-east-1.amazonaws.com",
      "key": "${env:AWS_BEARER_TOKEN_BEDROCK}",
      "models": {
        "us.anthropic.claude-sonnet-4-5-20250929-v1:0": {}
      }
    }
  }
}

Also reachable via the /login flow (provider bedrock).

How to verify / test

Automated — bb test (suite green, 523 tests, 2881 assertions). Bedrock-specific: clojure -M:test --focus eca.llm-providers.bedrock-test covers the binary event-stream decoder (crafted frames, truncated/malformed), message normalization, extraPayload filtering, request/response shapes, both tool-call loops, streaming callbacks, and error paths.

Manual against live AWS:

1. Generate a Bedrock API key in the AWS console (Bedrock → API keys) and export it: export AWS_BEARER_TOKEN_BEDROCK=...
2. Build: bb debug-cli → ./eca
3. Add the config above (adjust region in the URL and the model id to one enabled in your account), or use /login → bedrock.
4. Point your editor/client at the built ./eca, select a bedrock/... model, and run a chat. Verify:
   • streaming text renders incrementally
   • a prompt that triggers a tool call round-trips (tool runs, model continues)
   • reasoning content shows when using a reasoning-capable model

This branch was integrated and smoke-tested live (non-streaming text, ConverseStream streaming, multi-turn tool-call round-trip) by a downstream consumer that embeds ECA as a library.

---

• I added an entry in the changelog under the unreleased section.
• This is not an AI slop.

🤖 Generated with Claude Code

[ricardosllm]

I wrote this with Claude Opus 4.7 and reviewed as well with OpenAI GPT 5.5

It works for my usecase, here's what I tested:

Configuration. Bedrock configured as a standard provider entry:

  {:providers {"bedrock" {:api    "bedrock"
                          :url    "https://bedrock-runtime.ap-southeast-1.amazonaws.com"
                          :key    "<bearer token>"
                          :models {"global.anthropic.claude-sonnet-4-6" {}}}}}

• Region ap-southeast-1, runtime URL as above.
• Bearer-token auth (Authorization: Bearer) via config :key.
• Model: global.anthropic.claude-sonnet-4-6 (also configured global.anthropic.claude-opus-4-7).
• Driven through eca.llm-api/sync-or-async-prompt! with :provider "bedrock", the per-call :config above, and :model-capabilities — no
  :models model-config entry needed; the generic branch in llm-api/prompt! keyed on (or model-config model-capabilities) handles it.

Proven working — live against AWS Bedrock:

• Non-streaming Converse — text response, usage parsing.
• ConverseStream — token-by-token streaming, messageStop/end_turn, finish + usage callbacks.
• Tool calling — multi-turn round-trip: model emits toolUse, on-tools-called returns {:new-messages :tools}, ECA re-issues, model incorporates the tool result into the final answer. The on-tools-called contract ([{:id :full-name :arguments}] in, {:new-messages :tools} out) round-trips with no adapter glue.
• provider->api-handler's :else branch correctly routes :api "bedrock" to the native handler.
• Streaming chat and agentic tool workstreams confirmed end-to-end.

Not covered:

• Reasoning/thinking path (reasoningContent / redactedContent)
• Image input (supports-image?)
• :stream false via extra-payload (default streaming + the non-stream handler path were tested)
• Retry/error classification under real 429/503 from AWS
• Mid-stream cancellation
• Parallel tool calls (one tool call per turn tested)
• Long multi-turn conversations
• Non-Claude Bedrock models / inference profiles beyond the global Anthropic ones

If you think this is in the right direction I can try to test what was not covered yet.