feat: Add configurable Bedrock retry parameters

[Dan-Wuensch]

Resolves #4173

Changes included:

• Configurability of AWS Bedrock retry settings via environment variables
• New default retry settings for AWS Bedrock based on observed testing (these help mitigate failures due to running out of retries within the throttling window while don't push it too far)
• Add missing Databricks retry settings to providers documentation

Testing steps performed:

• AWS Bedrock provider retries using defaults
• AWS Bedrock provider retries using set environment variables
• Tried Ollama retries using qwen2.5 and qwen2.5-coder, but this is blocked by the in 1.6.0 bug Unsupported operation: streaming not implemented when listing files #4308
• Use OpenAI provider

[Dan-Wuensch]

I observed the following test is failing, but checked and it is also failing on main:

failures:

---- providers::base::tests::test_provider_metadata_context_limits stdout ----

thread 'providers::base::tests::test_provider_metadata_context_limits' panicked at crates/goose/src/providers/base.rs:564:9:
assertion `left == right` failed
  left: 200000
 right: 128000


failures:
    providers::base::tests::test_provider_metadata_context_limits

[taniandjerry]

Awesome job on your first contribution! Let me tag the dev team to take a look ❤️ @DOsinga @michaelneale @alexhancock @zanesq @jamadeo @katzdave

[DOsinga]

maybe change the default model too?

[Dan-Wuensch]

Does the default Goose install script use latest instead of version pinning?

If so, I worry about disruption to existing users by changing the model since their prompts may work worse (or better!) on a new model. Could require tuning prompt(s) and running evals to adopt in some cases

That upgrade concern probably only applies to pipeline / automatic use cases not desktop CLI/GUI though. Idk what percentage of adoption that represents, and then of course Bedrock would only be a subset of that

What do you think about me logging an Issue for the model update instead, to have that be a separate more visible change?

[DOsinga]

yeah, we're working on that. right now it is a bit of a mess which model a conversation will use. we're working on making this a real property of the session so that if you change the global settings, it won't change it for existing conversations.

either way, changing the default here, should only change which model is suggested when you select the provider

[Dan-Wuensch]

Ah ok! In that case I'll push up a commit for updating default. Want to use latest Sonnet version?

[Dan-Wuensch]

Demo of new default AWS Bedrock model recommendation:

[Dan-Wuensch]

Tangential thought: Does BEDROCK_KNOWN_MODELS affect any validation of the user's model entry? We could look at including some models from other vendors than Anthropic at some point. Not a big concern to me as long as it will accept pasting model IDs from outside the list, which it looks like it does to me.

[DOsinga]

These are just suggested models. @angelahning is looking into showing the full list in a filterable way

[blackgirlbytes]

@Dan-Wuensch there's a DCO check that's failing. Here's how to fix it 😄 https://github.com/block/goose/pull/4316/checks?check_run_id=48836635962

[Dan-Wuensch]

@blackgirlbytes thank you! 🙌 Just force pushed a rebase with the sign-off included for those three commits.

[michaelneale]

looking good a cargo fmt and maybe a script/cargo-lint.sh would help to check it is ok -

[Dan-Wuensch]

Good catch! There was a piece in the model list that needed to be reformatted. Pushed up re-formatted now

[Dan-Wuensch]

@blackgirlbytes checks and approvals are looking good! Except Semgrep appears to have gotten stuck. What's the next step in the process for these changes? 😄

[Dan-Wuensch]

Nice!!