fix(research): inherit main model's Bedrock region prefix in sub-agent

[taxfree-python]

Fixes #184.

Problem

The research sub-agent hard-codes bedrock/us.anthropic.claude-sonnet-4-6, which AWS Bedrock rejects with "The provided model identifier is invalid." whenever the caller's region/account cannot reach the us.* cross-region inference profile (e.g. ap-northeast-1). The main agent itself is unaffected because the user-chosen profile id is passed straight through to LiteLLM — only the sub-agent overrides it.

Fix

Inherit the inference-profile prefix from the main model. If the user is running bedrock/jp.anthropic.claude-opus-4-7, the sub-agent now picks bedrock/jp.anthropic.claude-sonnet-4-6; same for global., apac., etc. The sub-agent stays on the cheaper Sonnet model — only the prefix follows the main agent's choice, matching the project's existing convention of letting the user's explicit profile selection drive Bedrock routing.

if main_model.startswith("bedrock/") and "anthropic" in main_model:
    head = main_model.split("/", 1)[1]
    region_prefix = head.split(".", 1)[0]
    return f"bedrock/{region_prefix}.anthropic.claude-sonnet-4-6"

The existing us. test case is preserved (the new logic still produces us. for a us. main model), and a new test covers global. / apac. prefixes.

Verification

Direct LiteLLM call from ap-northeast-1 before/after, against the same model the function returns:

bedrock/us.anthropic.claude-sonnet-4-6  -> BedrockException "The provided model identifier is invalid."  (the reported bug)
bedrock/jp.anthropic.claude-sonnet-4-6  -> OK   (what the patched function emits for the user's main model)

Unit tests pass:

$ pytest tests/unit/test_cli_rendering.py
======================== 7 passed, 5 warnings in 6.51s =========================

Caveats

This makes the sub-agent work for every prefix under which claude-sonnet-4-6 is published on Bedrock (currently us., jp., global.). If a user is on an exotic prefix that has no sonnet-4-6 variant, the sub-agent will still fail on that environment — but that's also true of the main agent on that prefix today, and addressing it would expand the scope beyond this fix.

[taxfree-python]

Note on the failing review check: it appears to be the standard anthropics/claude-code-action@v1 permission issue affecting external-fork PRs (Actor does not have write permissions to the repository). I see the same failure on other recent fork-based PRs (e.g. #180, #181, #182), while internal-branch PRs pass. The change here is unrelated — happy to rebase or adjust if you'd like a fresh run after any workflow tweak.

[aryan5v]

@taxfree-python the failed Claude check is expected for external fork PRs. It only runs automatically for contributors with write access to avoid potential abuse of the review workflow and repo tokens. Maintainers can still trigger it

[fglogan]

closed per maintainer request