fix(bedrock): serve 1M context window for Claude Opus 4.6+/Sonnet 4.6

[sudhraja]

Summary

Bedrock Claude Opus 4.6/4.7/4.8 and Sonnet 4.6 serve a 1M context window, but Hermes was budgeting them at 200K — so long sessions compressed at ~half the real capacity.

The two halves of the code disagreed:

• On the wire: build_anthropic_bedrock_client (agent/anthropic_adapter.py) attaches the context-1m-2025-08-07 beta header to every Bedrock Claude request, unlocking the 1M window at the API.
• In the budget: get_model_context_length resolves Bedrock via the static BEDROCK_CONTEXT_LENGTHS table (agent/bedrock_adapter.py), which capped every Claude model at 200K.

opus-4-8 had no entry at all, so global.anthropic.claude-opus-4-8 substring-matched the generic anthropic.claude-opus-4 → 200K key. Result: 1M unlocked on the wire, 200K enforced by the agent.

Fix

Set the 1M-capable Bedrock Claude models to 1_000_000 in BEDROCK_CONTEXT_LENGTHS, matching the beta header and the existing DEFAULT_CONTEXT_LENGTHS entries in agent/model_metadata.py:

Model	Before	After
anthropic.claude-opus-4-8	200K (fell through to opus-4)	1M
anthropic.claude-opus-4-7	200K (fell through)	1M
anthropic.claude-opus-4-6	200K	1M
anthropic.claude-sonnet-4-6	200K	1M
Sonnet 4.5 / Opus 4 / Sonnet 4 / Haiku 4.5 / 3.x	200K	200K (unchanged)

Longer keys win the substring match in get_bedrock_context_length, so the generic anthropic.claude-opus-4 fallback still catches unversioned IDs at 200K.

Test Plan

• Updated the 3 tests in tests/agent/test_bedrock_adapter.py that encoded the stale 200K assumption (opus-4-6, sonnet-4-6, inference-profile), added an opus-4-8 case and a Sonnet-4.5-stays-200K guard.
• scripts/run_tests.sh tests/agent/test_bedrock_adapter.py → 137 passed, 0 failed.

[alt-glitch]

This was generated by AI during triage.

Duplicate of open #24059: the live diff makes the same Bedrock Claude 4.6/4.7/4.8 and Sonnet 4.6 context-table change. #24059 remains the earlier canonical PR.

[sudhraja]

Closing — incorrect approach. The 1M long-context beta on Bedrock is a per-account/org entitlement, not a model-level property. BEDROCK_CONTEXT_LENGTHS is a global static table shipped to all users, so hardcoding these models to 1M would make Hermes budget to 1M for accounts without the entitlement and hit Bedrock ValidationExceptions near the top of the window. The correct, entitlement-safe mechanism already exists: users who have the beta set model.context_length in their own config, which wins at step 0 of get_model_context_length(). No global change needed.