feat(bedrock): add OpenAI GPT-5.6 family (Sol/Terra/Luna) to Mantle Responses routing

[vinayshah1998]

Summary

GPT-5.6 Sol, Terra, and Luna went GA on Amazon Bedrock on July 13, 2026. Like GPT-5.5, they are served exclusively from the Bedrock Mantle OpenAI-compatible Responses endpoint (https://bedrock-mantle.<region>.api.aws/openai/v1) — the AWS model cards list bedrock-runtime/Converse as unsupported for all three.

This PR builds directly on #53880 (natebransc's Mantle/Responses routing for GPT-5.5) and extends the allowlist to the 5.6 family. The first two commits here are #53880 unchanged (authorship preserved); the third commit is the 5.6 addition. If #53880 merges first, this rebases down to the one commit.

• Add openai.gpt-5.6-sol, openai.gpt-5.6-terra, openai.gpt-5.6-luna to BEDROCK_OPENAI_RESPONSES_MODEL_IDS, so runtime resolution, auxiliary calls, and MoA slots all take the SigV4/bearer Mantle Responses path automatically.
• Surface the family in the curated Bedrock picker list.
• Record the 272K context window from the AWS model cards for all four Mantle OpenAI models (previously these fell back to the 128K BEDROCK_DEFAULT_CONTEXT_LENGTH).
• Generalize the picker tests from hardcoded openai.gpt-5.5 checks to iterate BEDROCK_OPENAI_RESPONSES_MODEL_IDS, so the next Mantle-only model addition is a one-line allowlist change with zero test surgery.

Docs: OpenAI models on Bedrock / GPT-5.6 Sol model card

Why

Bedrock's control-plane discovery (ListFoundationModels/ListInferenceProfiles) does not enumerate Mantle-only models, and the Mantle route is not Converse-compatible, so without the allowlist entry the 5.6 models are neither selectable nor callable from the Bedrock provider. GPT-OSS models remain intentionally excluded — they are Converse-capable and stay on the native path.

Testing

• scripts/run_tests.sh tests/agent/test_bedrock_integration.py tests/hermes_cli/test_bedrock_model_picker.py tests/run_agent/test_moa_loop_mode.py — 100 passed, 0 failed
• python -m py_compile on all touched files; scripts/check-windows-footguns.py — clean (4 files); git diff --check — clean
• Live smoke test against real Bedrock (bearer-token auth, us-east-1):
  • Direct Responses API: openai.gpt-5.6-terra returned bedrock-gpt56-ok
  • End-to-end through the CLI: hermes chat -q ... --provider bedrock -m openai.gpt-5.6-terra returned hermes-gpt56-e2e-ok
  • End-to-end openai.gpt-5.6-sol returned sol-ok
• New tests: Mantle routing for all three 5.6 variants, picker membership, and 272K context-length assertions

Notes

• Depends on / includes feat(bedrock): support OpenAI Responses models #53880. Happy to rebase once that lands.
• Pricing rows for the 5.6 family are not included: aws.amazon.com/bedrock/pricing does not yet publish per-token numbers for Sol/Terra/Luna and usage_pricing.py guidance is to only add rows with a verifiable source. Can follow up when AWS publishes them.

[teknium1]

Thanks for extending the Mantle routing work to the GPT-5.6 family. The premise is still present on current main: hermes_cli/runtime_provider.py:1975-2008 routes every non-Claude Bedrock model to Converse, and hermes_cli/models.py:532-544 does not surface Mantle IDs.

Problems

• The inherited Mantle auxiliary path from 83482bf7 derives its region independently with resolve_bedrock_region(). That disagrees with the main runtime, which prioritizes bedrock.region at hermes_cli/runtime_provider.py:1958-1962; auxiliary calls can therefore leave the primary runtime's configured region. This was also identified in the linked feat(bedrock): support OpenAI Responses models #53880 review.
• website/docs/guides/aws-bedrock.md:9 currently says Hermes does not use the OpenAI-compatible endpoint, but no documentation file is changed here.

Suggested changes

• Preserve the resolved Mantle endpoint/region through auxiliary resolution and cover a bedrock.region versus AWS_REGION mismatch.
• Update the Bedrock guide for the Converse and Mantle routes.

Automated hermes-sweeper review.

[vinayshah1998]

Both review problems addressed in ab2a229ea + 57c7d3758:

1. Auxiliary region resolution now matches the main runtime. Added resolve_bedrock_runtime_region() to agent/bedrock_adapter.py — the canonical implementation of the runtime's region priority (bedrock.region in config.yaml, then AWS_REGION/AWS_DEFAULT_REGION/botocore profile/us-east-1). The auxiliary aws_sdk branch in agent/auxiliary_client.py now uses it, so auxiliary calls (compression, memory, vision) can no longer leave the primary runtime's configured region — for both the AnthropicBedrock/Converse path and the new Mantle OpenAI Responses path. The runtime provider resolver, model validation (hermes_cli/models.py), live discovery (bedrock_model_ids_or_none), and the Mantle URL/SigV4 region fallbacks all route through the same helper, so the priority exists in exactly one place.

Regression tests cover the bedrock.region-vs-AWS_REGION mismatch for both auxiliary paths (test_bedrock_config_region_beats_env_region, test_bedrock_mantle_config_region_beats_env_region in tests/agent/test_bedrock_integration.py) — config region pinned to us-west-2 with ambient env at eu-central-1, asserting the resolved endpoint lands on us-west-2.

2. Bedrock guide updated. website/docs/guides/aws-bedrock.md no longer claims Hermes avoids the OpenAI-compatible endpoint. It now documents the triple routing (Claude via AnthropicBedrock SDK, GPT-5.5/5.6 via the Bedrock Mantle OpenAI Responses endpoint, everything else via Converse), the Mantle auth model (AWS_BEARER_TOKEN_BEDROCK or SigV4 via the boto3 chain), and lists the four openai.* model IDs in the Available Models table with a note that Mantle IDs don't use inference-profile prefixes.

Also folded in a small structural cleanup (57c7d3758): agent_init.py's if "client_kwargs" in locals() guard replaced by unconditional initialization, and the three hand-rolled region-priority copies collapsed into the single helper.

Bedrock-related suites pass locally: test_bedrock_integration.py, test_bedrock_adapter.py, test_bedrock_model_picker.py, test_runtime_provider_resolution.py, test_moa_loop_mode.py — 407 passed, 0 failed via scripts/run_tests.sh.

[vinayshah1998]

Gentle ping @teknium1 - both review concerns are resolved in ab2a229ea + 57c7d3758 (auxiliary region resolution now shares resolve_bedrock_runtime with the main runtime, plus E2E coverage against a temp HERMES_HOME). Happy to rebase if anything has drifted on main since. Ready for re-review whenever convenient.

[GottZ]

This was generated by AI during triage.

Summary

Two open PRs address Bedrock Mantle/OpenAI Responses routing: #53880 introduces GPT-5.5 routing, SigV4 authentication, picker exposure, auxiliary-client support, and MoA provider preservation, while #65076 incorporates that foundation, extends it to GPT-5.6 Sol/Terra/Luna, centralizes region resolution, adds context metadata and regression coverage, and updates the Bedrock guide.

Related pull requests

• feat(bedrock): support OpenAI Responses models #53880 duplicate — (+331/-25) — n/a: Adds the core GPT-5.5 Mantle Responses route across runtime, auxiliary, picker, and MoA paths, but its diff independently resolves the auxiliary region and retains brittle picker assertions identified by the contributor keep-open review. Despite that keep-open review, feat(bedrock): add OpenAI GPT-5.6 family (Sol/Terra/Luna) to Mantle Responses routing #65076 contains the same foundation while addressing the region and documentation concerns and extending model coverage, so feat(bedrock): support OpenAI Responses models #53880 can be consolidated into feat(bedrock): add OpenAI GPT-5.6 family (Sol/Terra/Luna) to Mantle Responses routing #65076.
• feat(bedrock): add OpenAI GPT-5.6 family (Sol/Terra/Luna) to Mantle Responses routing #65076 related — (+536/-36) — n/a: Supersets feat(bedrock): support OpenAI Responses models #53880 with GPT-5.5 and GPT-5.6 Mantle routing, shared config-first region resolution, SigV4/bearer authentication, 272K context metadata, generalized picker tests, MoA coverage, and Bedrock documentation. Consistent with the automated keep-open verdict, the visible diff addresses its cited concerns through resolve_bedrock_runtime_region(), config-versus-environment regression tests, and changes to website/docs/guides/aws-bedrock.md.

Duplicates

#53880 and #65076 implement the same core GPT-5.5 Mantle routing; #65076 is the broader successor and additionally carries the GPT-5.6 family and the documented review fixes.

Suggested consolidation

Keep #65076 open with a salvage path focused on the shared Mantle routing, centralized region handling, regression tests, and documentation; close #53880 as a duplicate of #65076 because its core implementation is incorporated there and its contributor-review concerns are addressed in the successor's visible diff.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup53880 ["PRs duplicating each other"]
        P53880["PR #53880 (open)"]
        P65076["PR #65076 (open)"]
    end
    class P53880 open
    class P65076 open
    class P65076 target
    click P53880 "https://github.com/NousResearch/hermes-agent/pull/53880"
    click P65076 "https://github.com/NousResearch/hermes-agent/pull/65076"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 73 kB of PR diffs, 7 kB of issue/PR text, 5 kB of discussion (5 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[vinayshah1998]

@teknium1 Following up now that the Aug 3 cross-PR triage has run: it marked this PR as the consolidation target for the Mantle Responses routing work and recommends closing #53880 as a duplicate, noting the earlier review concerns (auxiliary region resolution, picker test brittleness, docs) are addressed in the current diff. Happy to rebase on latest main if anything has drifted. Ready for re-review whenever you have bandwidth.