feat(curator): support reasoning effort override

[kortylokai-web]

What does this PR do?

Adds a curator-specific reasoning_effort override for the LLM consolidation/review fork, using Hermes' normalized reasoning configuration instead of provider-specific request-body hacks.

This lets users keep ordinary chat at one reasoning level while giving curator consolidation a different level, or explicitly disable curator reasoning with none.

Behavior

Reasoning precedence for the curator review fork is:

1. valid, non-empty auxiliary.curator.reasoning_effort
2. deprecated curator.auxiliary.reasoning_effort when the canonical key is absent or empty
3. the selected curator model's agent.reasoning_overrides entry
4. valid agent.reasoning_effort
5. provider or transport default

Additional semantics:

• none and YAML false explicitly disable reasoning and stop fallback.
• Invalid canonical values warn once, skip deprecated legacy config, and fall back through model-aware agent defaults.
• Legacy deprecation notices are emitted once even when provider resolution rewrites the selected model.
• The final provider-resolved model is used for model-aware fallback.
• Reasoning still reaches the fork when provider resolution falls back or fails.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests
• ♻️ Refactor

Changes Made

• agent/curator.py
  • Resolves canonical, legacy, model-specific, and global reasoning settings with explicit precedence.
  • Carries normalized reasoning through _ReviewRuntimeBinding alongside the existing provider, model, credential, and request overrides.
  • Re-resolves model-aware fallback only when provider resolution changes the selected model, without duplicating warnings.
  • Passes reasoning_config into the forked AIAgent without disturbing credential pools, ACP settings, or output caps.
• hermes_constants.py
  • Adds a backward-compatible keyword-only switch for suppressing duplicate invalid-value warnings during the second resolution pass.
• Configuration
  • Uses the auxiliary.curator.reasoning_effort schema default already present on current main.
  • Keeps config version 33; the default arrives through load_config() deep merge without migration or user-YAML rewrites.
• Tests
  • Covers canonical and legacy precedence, selected-model overrides, provider-rewritten models, global fallback, none, YAML false, invalid values, exact-once logging, runtime binding, constructor plumbing, provider-resolution failure, schema defaults, and no-write config loading.
• Docs and example
  • Documents inheritance, model-aware overrides, explicit disable behavior, consolidation requirements, and deprecated fallback in English and zh-Hans.

How to Test

Candidate head: 1a2276ae649298771b5a0c8aa3aa9609b77d341b

scripts/run_tests.sh -j 8 \
  tests/agent/test_curator.py \
  tests/agent/test_curator_classification.py \
  tests/hermes_cli/test_aux_config.py \
  tests/hermes_cli/test_config.py \
  tests/cron/test_reasoning_config_per_model.py \
  tests/gateway/test_reasoning_config_per_model.py \
  tests/tui_gateway/test_reasoning_config_per_model.py \
  tests/run_agent/test_fallback_reasoning_override.py -q

Result:

144 passed, 0 failed (8 workers)

Additional verification:

uv run ruff check .: passed
git diff --check upstream/main..HEAD: passed
Spec-compliance review: passed
Code-quality review: passed
Official upstream/main at final rebase: f88ed6c71768cdc7ea3bfa8cf62d16654792fd2a

The full Python suite was run with the exact locked CI dependency set and an 8-worker cap:

23,750 passed, 13 failed; one unrelated file passed on retry

All 13 failures reproduced exactly on pristine f88ed6c7 in the same environment (72 passed, 13 failed across the seven failing files). They are pre-existing host-state/timing failures outside this PR's files.

Scope

The refreshed PR is one commit on upstream/main (f88ed6c7), touching 8 feature files (+469/-14). It ports the original intent onto the current curator runtime architecture without replaying obsolete migration machinery or restoring tests pruned from current main.

Checklist

Code

• Read the contributing guidance
• Conventional commit message
• Feature-only scope
• Tests added for behavior, model-aware fallback, exact-once logging, and no-write config loading
• Full repository suite run with at most 8 workers; all failures reproduced on pristine upstream

Documentation & Housekeeping

• Relevant documentation updated
• cli-config.yaml.example updated
• Cross-platform impact checked
• No tool schema changes required

Screenshots / Logs

Not applicable; this changes curator/config behavior and documentation, not UI.

[teknium1]

Thanks for routing the setting through Hermes' normalized reasoning_config mechanism. Current main's curator fork has no curator-specific reasoning resolver or AIAgent plumbing (agent/curator.py:1758-1803, agent/curator.py:1875-1934), so this remains a valid addition.

Problems

• hermes_cli/config.py:3304 bumps _config_version for a default-only key. AGENTS.md:584-590 says new keys in an existing section are supplied by the deep merge and do not require a version bump. As written, migrate_config() persists a version-only rewrite (hermes_cli/config.py:6197-6200) even though load_config() already merges the new default (hermes_cli/config.py:6953-6967).

Suggested changes

• Retain version 33 and adjust the lean-config test to verify the default through load_config() without invoking a version migration.

Automated hermes-sweeper review.

[teknium1]

Please do not bump _config_version for this default-only field. load_config() deep-merges new keys from DEFAULT_CONFIG, while the version bump makes migrate_config() persist a version-only rewrite. AGENTS.md:584-590 reserves bumps for active config transformations.

[kortylokai-web]

Addressed in c71716a68: rebased onto current upstream/main, retained config version 33, removed the redundant hermes_cli/config.py diff, and replaced the migration-based test with a load_config()-only no-write contract. Refreshed curator/config suite: 339 passed.