fix(anthropic): auto-detect Bedrock model IDs in normalize_model_name (#12295)

[qike-ms]

What does this PR do?

normalize_model_name() unconditionally converts dots to hyphens, mangling Bedrock model IDs like anthropic.claude-opus-4-7 into anthropic-claude-opus-4-7. Bedrock rejects these with HTTP 400/404.

The existing mitigation (_anthropic_preserve_dots() in run_agent.py) only covers the main agent loop. All auxiliary client calls (compression, session_search, vision, flush_memories, title_generation) go through _AnthropicCompletionsAdapter which calls build_anthropic_kwargs() without preserve_dots=True. These calls have been silently broken for Bedrock users -- context compression fails, session search fails, etc.

This fix puts the detection inside normalize_model_name() itself so it covers all call sites without requiring callers to know about Bedrock.

Why fix it here instead of threading preserve_dots through every call site?

There are 10+ open PRs for #12295, all modifying _anthropic_preserve_dots() in run_agent.py. That approach:

• Only fixes the main agent loop, not auxiliary clients
• Requires every new call site to remember to pass preserve_dots=True
• Already has a track record of missing paths (the auxiliary client was missed by [Bug]: Bedrock inference profile model ID dots mangled to hyphens → HTTP 400 "invalid model identifier" #11976)

Fixing at the normalize_model_name level is defense-in-depth: Bedrock IDs are structurally recognizable and should never have their dots mangled regardless of how the caller is configured.

Related Issue

Fixes #12295

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/anthropic_adapter.py: Added _is_bedrock_model_id() helper that detects Bedrock namespace prefixes (anthropic., us., eu., ap., jp., global.). normalize_model_name() now skips dot-to-hyphen conversion for detected Bedrock IDs.
• tests/agent/test_bedrock_integration.py: Updated two canary tests that asserted the broken behavior. Added TestBedrockModelIdDetection class with 10 tests covering bare IDs, regional prefixes, negative cases, and end-to-end build_anthropic_kwargs verification.

How to Test

1. python -m pytest tests/agent/test_bedrock_integration.py tests/agent/test_anthropic_adapter.py -x -q -- 186 passed
2. python -m pytest tests/hermes_cli/test_model_normalize.py -x -q -- 55 passed
3. Verify manually:

from agent.anthropic_adapter import normalize_model_name
# Bedrock IDs preserved (was broken):
assert normalize_model_name("anthropic.claude-opus-4-7") == "anthropic.claude-opus-4-7"
assert normalize_model_name("us.anthropic.claude-sonnet-4-5-v1:0") == "us.anthropic.claude-sonnet-4-5-v1:0"
# OpenRouter dots still converted (unchanged):
assert normalize_model_name("claude-opus-4.6") == "claude-opus-4-6"

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes
• I've tested on my platform: Ubuntu 24.04, Linux 6.x, Python 3.11

Documentation & Housekeeping

• I've updated relevant documentation (docstrings on normalize_model_name and new _is_bedrock_model_id) -- or N/A
• N/A -- no config keys changed
• N/A -- no architecture changes
• N/A -- pure Python, no platform-specific code
• N/A -- no tool schemas changed

[alt-glitch]

Competing fix for #12295 — differs from #12336/#12778/#13112 by moving Bedrock detection into normalize_model_name() instead of threading preserve_dots through call sites. Covers auxiliary clients (compression, session_search, vision) that other PRs miss.

[teknium1]

This fix has already landed on main — closing as implemented.

Automated hermes-sweeper review:

• Commit f2fba4f9a merged this exact change: _is_bedrock_model_id() helper + guarded normalize_model_name().
• agent/anthropic_adapter.py lines 989–1031 now contain the Bedrock detection logic and the early-return guard, matching the PR's stated changes exactly.
• The fix is on main but post-dates tag v2026.4.23; it will ship in the next release.

Thanks @qike-ms for the well-scoped fix and thorough test coverage — the approach of centralising detection inside normalize_model_name() rather than threading preserve_dots through every call site was the right call.