feat: dynamic Bedrock model listing with LiteLLM metadata enrichment

[carTloyal123]

Summary

Replaces the hardcoded Bedrock model list with live AWS API calls, ensuring users only see models they have access to in their account and region. This solves the GovCloud model filtering problem and improves UX for commercial Bedrock users.

Problem

The Bedrock provider returns ~80+ hardcoded models from provider.json regardless of the user's account, region, or enabled models. GovCloud users see commercial-only models they can't use, and commercial users see models they haven't enabled in their Bedrock console.

Solution

Two-source architecture:

1. AWS ListFoundationModels + ListInferenceProfiles -- Account/region-aware APIs that return only accessible models. Filtered to ON_DEMAND, streaming-capable, ACTIVE models.

2. LiteLLM community registry (model_prices_and_context_window.json) -- Enriches each model with context_length, tools_supported, supports_reasoning, supports_vision metadata that AWS APIs don't provide.

Both sources gracefully degrade: AWS API failure falls back to the hardcoded list; LiteLLM failure just means missing metadata.

Changes

File	Description
Cargo.toml	Added aws-sdk-bedrock workspace dependency
crates/forge_repo/Cargo.toml	Added aws-sdk-bedrock crate reference
crates/forge_repo/src/provider/bedrock.rs	+613 lines: control plane client, live model listing, LiteLLM enrichment, FIPS support, 10 new tests
crates/forge_repo/src/provider/provider.json	Added optional AWS_USE_FIPS URL param for GovCloud

Key Implementation Details

• Control plane client (init_control_client()): Separate aws_sdk_bedrock::Client for model listing, same auth mode and region as the chat client
• Model filtering: Only ON_DEMAND, streaming-capable, ACTIVE lifecycle models included
• Inference profiles: Cross-region profiles (e.g., us.anthropic.claude-3-5-sonnet) merged and deduplicated
• LiteLLM enrichment: Fetched and cached with 7-day TTL via existing model cache infra. Tries both raw model ID and Bedrock-prefixed ID for lookup
• Fallback: If ListFoundationModels fails (missing IAM permission), returns hardcoded list with warning log
• FIPS support: Optional AWS_USE_FIPS=true param enables FIPS endpoints on both data and control plane clients

Testing

• 10 new unit tests (72 total pass, 1 ignored):
  • Model mapping: Claude, Nova, Llama foundation models
  • Registry enrichment: Bedrock prefix lookup, raw ID lookup, not-found handling, no-overwrite of existing data
  • Fallback: Hardcoded list returned on API error
  • FIPS: Config parsing for true/false/absent
• cargo check passes across full workspace

GovCloud Usage

forge provider login  # select bedrock
# Choose "AWS Profile (SSO/IAM)"
# Enter GovCloud profile name
# Set AWS_REGION to us-gov-west-1
# Optionally set AWS_USE_FIPS to true
forge list models  # shows only GovCloud-available models

Co-Authored-By: ForgeCode noreply@forgecode.dev

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}