fix(title): omit unsupported structured output

[EAbaracus]

What does this PR do?

Fixes auxiliary session-title generation failures on providers/models that do not support OpenAI structured-output schemas.

When Hermes generates a session title, it previously always sent:

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "session_title"
    }
  }
}

Nous + deepseek/deepseek-v4-flash-0731 rejects that request with HTTP 400, while the normal chat request succeeds.

This PR makes response_format capability-aware:

• True → send the existing JSON Schema request format
• False → omit response_format
• None / unknown → omit response_format

The existing JSON/prose title parser remains unchanged.

Related Issue

Fixes #

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• Added _title_request_extra_body() in agent/title_generator.py.
• Added tri-state response_format_supported handling to generate_title().
• Omitted empty extra_body entirely when structured output is unsupported or unknown.
• Preserved the existing _TITLE_RESPONSE_FORMAT for explicitly supported routes.
• Added tests covering:
  • unknown capability;
  • unsupported capability;
  • supported capability;
  • plain/prose title parsing without response_format.

Provider/model resolution, retry classification, delegation, and 503 handling are unchanged.

How to Test

Focused title-generation tests:

venv/Scripts/python.exe -m pytest \
  tests/agent/test_title_generator.py \
  -v --tb=short

Result:

37 passed

Gateway regression tests:

venv/Scripts/python.exe -m pytest \
  tests/gateway/test_compression_progress_notices.py \
  -v --tb=short

Result:

38 passed

Main delegation smoke test:

hermes chat -q "Reply with exactly: P0_DELEGATION_OK" \
  --provider nous \
  -m deepseek/deepseek-v4-flash-0731

Result:

P0_DELEGATION_OK

The full auxiliary-client suite was also run:

177 passed, 4 unavailable

The four unavailable tests are async tests requiring pytest-asyncio, which is not installed in the test environment. They were not failures caused by this change.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Windows 10

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A; behavior is covered by the implementation docstring and tests
• I've updated cli-config.yaml.example if I added/changed config keys — N/A; no config keys changed
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — request-shape logic is platform-independent
• I've updated tool descriptions/schemas if I changed tool behavior — N/A; no tool schema changed

For New Skills

• This skill is broadly useful to most users (bundled) — N/A
• SKILL.md follows the standard format — N/A
• No external dependencies that aren't already available — N/A
• I've tested the skill end-to-end — N/A

Screenshots / Logs

Before this change:

Auxiliary title generation failed:
HTTP 400: This request is not valid.
Additional info: Provider returned error

The failing route was:

provider=nous
model=deepseek/deepseek-v4-flash-0731
base_url=https://inference-api.nousresearch.com/v1

The failure occurred only on the title-generation request containing response_format=json_schema.

After this change, the request-shape tests pass for both unknown and unsupported capabilities, while explicitly supported capability still sends the existing schema.

[spfcraze]

This was generated by AI during triage.

Summary:
This PR's new response_format_supported gate is never passed by a live caller, so the JSON-Schema title request is now omitted on every route and structured title output is dropped everywhere, not just on the failing provider.

Problems:

• hermes_cli/sessions_cmd.py:961 calls generate_title(typed) with a single argument, and the auto-titler at agent/title_generator.py:604 passes failure_callback/main_runtime/runtime_validator but not the new response_format_supported.
• With the None default, _title_request_extra_body returns {} and extra_body is omitted, so the branch that would send the JSON Schema is unreachable from production code.

Solution:
Resolve the capability at the call sites from the model's advertised supported_parameters / structured_output (parsed in hermes_cli/models.py and agent/models_dev.py) and pass response_format_supported, so routes that support structured output keep sending the JSON Schema.

Evidence

no deterministic fact backs this claim — model belief, not executed or read evidence

---

Checked against 9d3207b — the PR head when this was written — and fa83af3, main at the same moment.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

Main concern: the new response_format_supported parameter defaults to None, and neither production call site passes it — agent/title_generator.py:604 and hermes_cli/sessions_cmd.py:953 both call generate_title(...) without it. _title_request_extra_body(None) returns {}, so response_format is now never sent: the change silently disables structured output everywhere rather than just "omitting it when unsupported." If the flag is meant to be derived from the runtime/validator, the call sites (or a resolver inside generate_title) need to supply it; otherwise this is a behavior regression beyond the stated intent.

• No test exercises the real caller wiring — only the helper and generate_title with an explicit flag. An integration test asserting the capability flag is actually derived from the runtime would catch the silent-disable regression above.
• Minor: _title_request_extra_body returning {} and the caller's truthiness check works, but returning None when unsupported would make the "omit entirely" intent slightly more explicit.