feat: Add bedrock provider package

[jsonbailey]

Summary

• Adds launchdarkly-server-sdk-ai-bedrock provider package at packages/ai-providers/server-ai-bedrock/.
  • Model runner wraps boto3 bedrock-runtime.converse(...) (async via asyncio.to_thread).
  • Agent runner integrates the Strands Agents SDK, exposed through the optional [agents] extra so the package works without it installed.
  • Agent-graph support intentionally deferred — create_agent_graph(...) raises NotImplementedError. Follow-up work will land that piece.
• Mirrors the OpenAI provider layout: BedrockRunnerFactory, BedrockModelRunner, BedrockAgentRunner, bedrock_helper (message/tool conversion, metrics extraction, map_provider), and matching tests.
• Registers the new package in the workspace pyproject.toml, root Makefile, release-please-config.json / .release-please-manifest.json, and .github/workflows/{ci,release-please}.yml.
• PyPI name launchdarkly-server-sdk-ai-bedrock is already reserved at 0.0.0 (placeholder). First real release will be 0.1.0 via release-please.

Depends on #191 (bedrock tracker-method removal). #191 should merge first.

feat: Add bedrock provider package for AWS Bedrock + Strands

Test plan

• make test in the new package — 42 passed
• make test across the workspace — 377 passed (201 server-ai + 88 langchain + 46 openai + 42 bedrock)
• make lint across all four packages — clean
• make build in the new package produces wheel + sdist; twine check passes
• Reviewer: confirm strands-agents version floor (currently >=0.1.0)
• Reviewer: confirm CI workflow matrix additions (linux + windows for the new package)

Notes for reviewers

• Structured output: implemented via system-prompt JSON-schema injection + json.loads. Bedrock's native toolConfig is a more reliable structured-output mechanism but doesn't map cleanly to the OpenAI provider's parsed shape. Reasonable follow-up.
• tool_calls from the agent runner: Strands aggregates per-tool call counts rather than preserving invocation order; the runner emits each name repeated by call_count. Model runner returns names in response order.
• LDAIMetrics.success is derived strictly from HTTPStatusCode == 200, matching the prior track_bedrock_converse_metrics semantics.

🤖 Generated with Claude Code

---

Note

Medium Risk
Adds a new Bedrock provider implementation and wires it into CI and automated PyPI publishing, so failures could impact release automation and provider selection paths.

Overview
Adds a new packages/ai-providers/server-ai-bedrock distribution (launchdarkly-server-sdk-ai-bedrock) implementing a Bedrock AIProvider with a BedrockModelRunner (boto3 converse via asyncio.to_thread, toolConfig support, and schema-injection structured output) plus an experimental BedrockAgentRunner backed by optional strands-agents; create_agent_graph is explicitly unsupported.

Updates repo plumbing to include Bedrock everywhere: uv workspace membership, root Makefile install/test/lint/build targets, CI jobs for Linux+Windows, and release-please configuration/manifest plus a new release job to publish the Bedrock package to PyPI with provenance attestation.

^{Reviewed by Cursor Bugbot for commit e678074. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default mode and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit dffd3b8. Configure here.}

[cursor]

Metrics drop tool_calls on empty-content error paths

Low Severity

When the Bedrock model responds with only toolUse content blocks and no text (a valid scenario when tools are configured via parameters), extract_content_from_response returns empty string. The code then constructs a new LDAIMetrics passing tokens and duration_ms from the original metrics but omitting tool_calls. This loses diagnostic information about which tools the model attempted to invoke. The same pattern appears in the structured-output JSON decode failure path.

Additional Locations (1)

• packages/ai-providers/server-ai-bedrock/src/ldai_bedrock/bedrock_model_runner.py#L160-L169

 

^{Reviewed by Cursor Bugbot for commit dffd3b8. Configure here.}