fix(moa): add tool_call_batch_cadence for periodic advisor refresh (#63393)

[webtecnica]

Summary

Adds tool_call_batch_cadence — a new per-preset setting that controls how many tool-call iterations pass between reference advisor refreshes. Default 0 = run every iteration (current behavior, fully backward compatible).

Changes

hermes_cli/moa_config.py

• _default_preset(): add "tool_call_batch_cadence": 0
• _normalize_preset(): coerce via existing _coerce_int(raw.get("tool_call_batch_cadence"), 0)
• normalize_moa_config(): flatten into active view

agent/moa_loop.py

• MoAChatCompletions.__init__: add _tool_call_batch_count counter
• create(): increment counter on each call; when cadence > 0 and the counter doesn't align, skip the reference fan-out entirely — the aggregator runs alone with its own context

How it works

presets:
  default:
    tool_call_batch_cadence: 3   # run advisors every 3rd tool iteration

With cadence: 3:

• Iterations 1, 4, 7, ... → advisors refreshed, aggregator gets fresh guidance
• Iterations 2, 3, 5, 6, ... → no advisor re-run, aggregator acts alone

Values:

• 0 (default): every iteration — no behavioral change
• 1: every iteration (same as 0)
• N > 1: run every Nth iteration

Testing

All 76 existing MoA tests pass:

22 passed (test_moa_config.py)
39 passed (test_moa_loop_mode.py, test_moa_switch_api_mode.py, test_moa_aggregator_*)
15 passed (test_moa_streaming.py, test_moa_reference_emit.py, test_moa_trace_streamed_capture.py)

Closes #63393

[teknium1]

Thanks for addressing the real latency gap between per_iteration and user_turn.

Problems

• The off-cadence branch clears reference_models. On current main, the reference-model labels are part of the cache key (agent/moa_loop.py:986), so this misses the prior cache and yields empty outputs; guidance is attached only for nonempty outputs (agent/moa_loop.py:1062-1077). The aggregator therefore loses prior advisor guidance rather than reusing it, contrary to #63393's requested behavior.
• The setting is not preserved by Dashboard saves: MoaPresetPayload and the persistence reconstruction omit it (hermes_cli/web_server.py:1001-1009, hermes_cli/web_server.py:5713-5720), and the Dashboard response type omits it too (web/src/lib/api.ts:2306-2323).
• The two-file diff adds no regression coverage for cadence timing, prior-guidance reuse, or configuration round trips.

Suggested changes

• Build the cadence atop the existing reference cache so skipped iterations reuse cached outputs but do not rerun or re-account advisors.
• Wire the setting through Dashboard payload/types/editor and document its interaction with fanout.
• Add runtime and config/API tests, including a parallel tool-call batch.

Automated hermes-sweeper review.

[teknium1]

Clearing reference_models also changes the current cache key, whose third element is the reference-model labels (agent/moa_loop.py:986 on main). That prevents reuse of the prior advisor outputs and leaves the aggregator with no reference guidance on skipped iterations. Preserve cached outputs while suppressing only the new advisor fan-out.

[teknium1]

Closing with credit — the cadence feature shipped via cluster PR #70284 (commit f7b90e6) as a redesigned implementation with you credited (Co-authored-by). The redesign folds the cadence into the existing fanout enum (fanout: every_n:N) and reuses cached advisor guidance on off-cadence iterations instead of leaving the aggregator with no advice, with the counter scoped per user turn. Your issue #63393 (multiplicative advisor latency) was the real problem and is now fixed. Thanks!