fix(bedrock): recognise au./apac. inference profiles to enable prompt caching

[jaketracey]

Problem

Claude models invoked on Amazon Bedrock through the Australia (au.) or Asia-Pacific (apac.) cross-region inference profiles — e.g. au.anthropic.claude-haiku-4-5-20251001-v1:0, apac.anthropic.claude-sonnet-4-6 — run without prompt caching, making multi-turn / agentic sessions far more expensive than they should be.

Root cause

is_anthropic_bedrock_model() decides the routing fork: Claude models go through the AnthropicBedrock SDK path (full feature parity — prompt caching, thinking budgets, adaptive thinking), everything else through the Converse API path (no cache_control support). It detects Claude by stripping a regional prefix and checking for anthropic.claude:

for prefix in ("us.", "global.", "eu.", "ap.", "jp."):

This list is missing au. entirely, and ap. does not match apac. ("apac.anthropic…".startswith("ap.") is False). So au.* and apac.* Claude profiles fall through to the Converse path and silently lose prompt caching.

Fix

Add "apac." and "au." to the prefix list (apac. before ap. for clarity; they don't overlap functionally):

for prefix in ("us.", "global.", "eu.", "apac.", "ap.", "au.", "jp."):

Impact

Validated live on a real agentic session (SEO audit, au.anthropic.claude-*) before vs after:

	cache_read_tokens
Before (Converse path)	0
After (AnthropicBedrock path)	80,000+

i.e. the static system+tools prefix was being re-billed at full input price every turn instead of the ~10×-cheaper cache-read rate — a large, silent cost regression for AU/APAC Bedrock users.

Tests

Adds test_au_inference_profile and test_apac_inference_profile to TestIsAnthropicBedrockModel, matching the existing test_eu_claude style. No behaviour change for non-Anthropic models (Nova / DeepSeek / Llama / Mistral) or the existing us. / global. / eu. profiles.

Notes

Discovered while deploying Hermes on Bedrock in ap-southeast-2 (Australia) for an AU-data-residency deployment. AWS added the au. (Sydney + Melbourne) inference profiles relatively recently; this just brings the detector in line. Worth auditing other prefix lists for the same au./apac. gap.

🤖 Generated with Claude Code

[teknium1]

Thanks for identifying a real Bedrock routing gap. Current main confirms that agent/bedrock_adapter.py:451 omits both au. and apac., while hermes_cli/runtime_provider.py:1979-2000 uses that result to choose the AnthropicBedrock caching path.

Problems

• The new routing is incomplete for auxiliary Bedrock calls. Once this detector accepts AU/APAC, agent/auxiliary_client.py:5085-5095 selects AnthropicAuxiliaryClient; its request adapter calls build_anthropic_kwargs() without preserve_dots (agent/auxiliary_client.py:1278-1286). _is_bedrock_model_id() still omits these prefixes (agent/anthropic_adapter.py:1571-1575), so normalize_model_name() converts their structural dots into hyphens (agent/anthropic_adapter.py:1597-1604).
• The mirrored Bedrock pricing prefix list also omits AU/APAC (agent/usage_pricing.py:680-697), although the Bedrock pricing lookup calls it at agent/usage_pricing.py:730-737.

Suggested changes

• Share or consistently extend the regional-prefix list in the detector, _is_bedrock_model_id(), and _normalize_bedrock_model_name().
• Add routing, outgoing-model-preservation, and pricing-normalization tests for both prefixes.

Automated hermes-sweeper review.

[teknium1]

Merged via PR #68005 — your commit was cherry-picked onto current main with your authorship preserved (rebase merge). You were the earliest submission for the au./apac. gap (June 14), and during salvage we found two more sibling sites with the same stale prefix list (the Bedrock-ID detector in anthropic_adapter and the reasoning-timeout resolver) — all four now share one roster. Thanks for the fix, prompt caching now works down under!

#68005