stage-batch32: v0.51.150 / Release DV — single-PR reasoning-effort agent metadata

[nesquena-hermes]

stage-batch32 — v0.51.150 (Release DV)

Single-PR reasoning-effort agent metadata integration.

PR

• fix(reasoning): use agent metadata for effort support #3017 Use agent metadata for reasoning-effort support — starship-s

Changes resolve_model_reasoning_efforts() in api/config.py to consult Hermes Agent's models.dev metadata (agent.models_dev.get_model_capabilities()) between the exact provider-specific resolvers and the broad heuristic fallback. Closes the gap where native/provider-specific catalogs (e.g. xAI OAuth grok-4.3) were reasoning-capable per agent metadata but reported no efforts by WebUI because no local prefix heuristic matched.

Order of precedence (verified by Opus)

ACP / Copilot subset / OpenAI Codex / LM Studio probe  →  (authoritative)
NEW: models.dev metadata
  • supports_reasoning=True  → full effort list
  • supports_reasoning=False → [] (authoritative — known non-reasoning variants stay hidden)
  • Unknown / None          → fall through
Prefix-based heuristic fallback                        →  (unchanged)

The new metadata layer sits exactly where the old prefix heuristic was the only answer for non-special-cased providers. Existing exact resolvers return BEFORE the metadata lookup, so they cannot be overridden by stale agent metadata.

Pre-merge gate

• Targeted-test sweep (41 reasoning-adjacent tests): 41/41 passed
• Full sequential pytest: 6657 passed, 0 failures, 12 skipped (156s)
• CI green on PR branch: test 3.11/3.12/3.13 SUCCESS
• Opus advisor deep review (claude-opus-4-7): SHIP-IN-NEXT-LOW-RISK-BATCH
  • Brick-class assessment: ~2/10 — additive + fallthrough-safe
  • Import is bare-except Exception-guarded (older hermes-agent without models_dev module = identical behavior to current master)
  • supports_reasoning=False authoritative path is bounded: a single model becomes unaccessible-for-reasoning at worst, not a global brick
  • The models.dev call is wrapped in a second try/except Exception; no raise can escape

Drive-by fix called out by Opus

The LMStudio branch now passes hinted_model (the provider-stripped value) to lmstudio_model_reasoning_options() instead of the raw model arg. Previously the unstripped @xai-oauth:... style hint was passed through to LMStudio's probe — that was a real bug that this PR incidentally fixes.

Test coverage

tests/test_models_dev_reasoning.py (110 LOC, 6 tests) covers:

• supports_reasoning=True → full effort list
• supports_reasoning=False → authoritative empty
• Unknown / None → falls through to prefix heuristic
• @xai-oauth: hint stripping before lookup
• False overrides a prior known prefix heuristic
• get_reasoning_status() hydrates from config defaults

Files

• api/config.py: +114/-39 (new _models_dev_reasoning_efforts() helper, resolver call-chain update, no-query path hydration)
• tests/test_models_dev_reasoning.py: NEW, 6 tests
• CHANGELOG.md: v0.51.150 / Release DV entry

Non-blocking follow-ups (Opus suggestions)

1. Add a model.reasoning_efforts_force_supported: true config knob so a user can override an authoritative-False signal when models.dev cache is stale. File as sprint-candidate.
2. Add a direct test for the ImportError branch (mock sys.modules["agent.models_dev"] = None) — the existing "metadata returns None" tests cover the same downstream behavior via a different upstream path.

Both can land in a subsequent PR.