Flips the default fan-out cadence from per_iteration (advisors re-run on
every tool iteration, multiplying advisor spend by tool-loop depth) to
user_turn (advisors run once on the first message of each user turn; the
acting aggregator works the rest of the tool loop with that turn's
advice). Until per-mode benchmarks justify a costlier default, MoA
defaults to the cheapest, lowest-impact cadence (#67199).
One default for everyone — no split legacy/new-preset semantics; presets
that want per-step advising set fanout: per_iteration explicitly. All
three modes (user_turn / per_iteration / every_n:N) remain selectable;
every_n:1 still collapses to per_iteration (semantic identity), while
unparseable values now fall to user_turn (the default).
Docs updated with a default-change note; the per-iteration rerun test
pins its mode explicitly.
Co-authored-by: skyer-flyyy <188930297+skyer-flyyy@users.noreply.github.com>
Follow-ups for salvaged #53784:
- reference_timeout now defaults to None = no per-preset override, so the
reference fan-out inherits auxiliary.moa_reference.timeout (900s default)
via call_llm's own per-task timeout resolution. The PR's 30.0s default
would have cut off long-thinking advisors mid-response, and its 300s max
cap capped legitimate explicit values — both removed. Explicit per-preset
values are still honored as-is.
- _is_failed_reference also treats '[skipped: …]' recursion-guard notes as
internal sentinels, keeping them out of both aggregator prompts.
- Dashboard/desktop TS types updated to number | null; web_server validator
accepts null/empty as 'inherit'.
Extends the fanout enum with 'every_n:<N>' (N >= 2): advisors run on the
first iteration of each user turn and every Nth tool iteration after it;
off-cadence iterations REUSE the cached guidance from the last on-cadence
run via the same cache mechanism the user_turn fanout uses, so the
aggregator still gets advice on every step. The cadence counter is scoped
per user turn (resets on a new user message) and only advances when the
advisory state actually changes, so streaming retries never consume a
cadence slot. Mapping form {mode: every_n, n: N} normalizes to the
canonical string. Unknown/degenerate values fall back to per_iteration.
Addresses issue #63393 (advisor fan-out multiplies turn latency/cost by
the tool-iteration count). Redesigned from PR #63448: the submitted shape
skipped references entirely on off-cadence iterations (aggregator ran
advice-less); this version keeps the last advice in play, credited for
the idea and cadence framing.
Config-gated, default-off (default fanout remains per_iteration).
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
When reference_models is stored as a JSON string (e.g. from hermes moa
configure or hand-edited config.yaml), _normalize_preset silently
falls back to hardcoded defaults because the string fails both
isinstance(x, list) and isinstance(x, dict) checks.
Add json.loads() parsing before the type checks so both formats work.
MoA reference_max_tokens is preset-level — one cap for all reference
models. When mixing a verbose model with a terse one, a single cap is
either too tight for the terse model or too loose for the verbose one.
Now each reference slot can optionally carry its own max_tokens:
reference_models:
- provider: openrouter
model: deepseek/deepseek-v4-pro
max_tokens: *** # per-slot cap, overrides preset-level
- provider: openai-codex
model: gpt-5.5
# no max_tokens → falls back to preset-level reference_max_tokens
_clean_slot (moa_config.py) preserves an optional max_tokens field on
the slot dict, coerced via _coerce_int_or_none. _run_reference
(moa_loop.py) reads slot-level max_tokens first, falling back to the
preset-level cap passed by the caller. Slots without the field are
unaffected — backward compatible.
Type hints on slot-handling functions updated from dict[str, str] to
dict[str, Any] to reflect the now-heterogeneous slot shape.
Keep invalid persisted preset names fail-closed, list the valid configured choices, and classify the local lookup failure as deterministic so it reaches Desktop immediately.
_clean_reasoning_effort kept its own whitelist that stopped at 'max',
silently dropping 'ultra' from MoA slot configs. Route it through
hermes_constants.parse_reasoning_effort — the same one-source-of-truth
fix the salvaged commit applies to the gateway — so future effort
levels can't drift here either. Docs updated to list ultra.
Follow-up to salvaged PR #64012.
Follow-up hardening on top of #64158 (@DavidMetcalfe):
Backend (the root-cause fix):
- hermes_cli/moa_config.py: add validate_moa_payload() — strict write-time
counterpart to the deliberately tolerant normalize_moa_config(). Flags
half-filled slots, empty reference lists, recursive moa slots, naming the
exact preset/slot.
- hermes_cli/web_server.py: PUT /api/model/moa validates before normalizing
and returns 422 with the specific problems instead of silently swapping the
user's preset for hardcoded defaults (#64156). Also declares
fanout / reference_max_tokens / reasoning_effort on the Pydantic payload so
client round-trips no longer erase hand-set values.
Desktop:
- Replace sanitize-then-send with hold-while-incomplete: the debounced
autosave is deferred (not repaired) while any slot is half-filled, and
flushes once the model pick completes the edit. Mid-edit UI state is never
repainted by a save response (generation guard covers held edits too).
- updateMoaSlot only clears the model when the provider actually changed.
- Explicit preset ops (set default / add / delete) cancel the pending
autosave and invalidate in-flight responses so the two writers can't race.
- Stable row keys (preset+index) so mid-edit rows don't remount; cleared
model shows the 'Model' placeholder instead of vanishing.
Both TS clients' MoaConfigResponse types now declare the round-tripped
fields (fanout, reference_max_tokens, reasoning_effort).
Tests: 12 new backend unit tests (validate_moa_payload contract incl.
validate/normalize agreement), 3 new web_server endpoint tests (422 on
half-filled ref/aggregator, fanout round-trip), 3 new desktop vitest cases
(autosave held while half-filled, flush on completion, same-provider
reselect no-op). E2E validated against a live TestClient with isolated
HERMES_HOME: bug sequence now 422s with config untouched.
Fixes#64156
New preset key 'fanout': 'per_iteration' (default, unchanged behavior)
re-runs the reference fan-out whenever the advisory view changes — every
tool iteration. 'user_turn' runs the advisors ONCE per user turn and lets
the aggregator act alone for the rest of the tool loop — the original MoA
shape (upfront multi-model synthesis, then a single acting model), and the
obvious lever on MoA's wall/cost multiplier (advisor generation dominates
per-turn latency).
Implementation reuses the existing turn-scoped reference cache: in
user_turn mode the cache signature hashes only the prefix up to the LAST
user message, so mid-turn advisory-view growth doesn't change the key and
iteration 2+ is a cache HIT (advice reused, zero advisor spend, no
re-trace). A new user message changes the prefix and re-triggers the
fan-out. Unknown fanout values normalize to per_iteration.
A single-model Hermes agent never sends temperature; the provider default
applies. MoA hardcoded reference_temperature=0.6 / aggregator_temperature=0.4,
and the coercion float(preset.get(key, 0.6) or 0.6) made unset IMPOSSIBLE to
express: absent, null, empty, and even an explicit 0 all collapsed to the
baked-in default. Every MoA advisor and aggregator therefore ran at 0.6/0.4
while the same model running solo used the provider default — silently
skewing solo-vs-MoA comparisons and overriding provider-tuned defaults.
- moa_config normalization: temperatures coerce to None when absent/blank/
invalid (new _coerce_float_or_none); explicit values incl. 0 honored.
- moa_loop: _preset_temperature() resolves preset values; None flows to
call_llm, which already omits the parameter when None (same contract as
max_tokens). Aggregator still inherits the acting agent's own configured
temperature when the preset doesn't pin one.
- conversation_loop (context-mode MoA): same resolution, no more hardcoded
0.6/0.4 at the call site.
- DEFAULT_CONFIG preset + web_server payload models + docs updated: unset
is the default, pinning stays available.
MoA per-turn latency is dominated by advisor GENERATION: turn wall time
correlates ~0.88 with output tokens and ~-0.03 with input tokens (measured over
52 turns). Each turn waits for the slowest advisor to finish writing, and
advisors were uncapped — writing multi-thousand-token essays the aggregator
only needs the gist of.
Add an opt-in per-preset reference_max_tokens knob (mirrors reference_temperature)
that caps ADVISOR output only; the acting aggregator is never capped. Default
None = uncapped, so existing presets are byte-for-byte unchanged (no regression).
Wired through both MoA execution paths (MoAChatCompletions.create and
aggregate_moa_context).
E2E: same task, closed preset uncapped vs reference_max_tokens=600 -> 59s to 33s
(~44% faster), final answer identical/correct.
- hermes_cli/moa_config.py: _coerce_int_or_none helper + reference_max_tokens
in _normalize_preset/_default_preset/flattened view
- agent/moa_loop.py: read preset.reference_max_tokens, pass to reference fan-out
- agent/conversation_loop.py: pass reference_max_tokens on the per-turn path
- tests + docs
exact_moa_preset_name matched any bare model name equal to a preset key,
regardless of the preset's enabled flag. On the no-explicit-provider switch
path (PATH B in model_switch.py), a plain /model switch whose name collided
with a preset key (e.g. "default") silently pivoted the session onto the MoA
virtual provider — even when the user had set enabled: false to opt out
(issue #55187). The LLM driving a routine model switch could land on a broken
moa provider with empty default_preset / unconfigured aggregator credentials.
Gate the implicit bare-name match on the per-preset enabled flag. Explicit
selection via --provider moa / the model picker uses PATH A and does not go
through exact_moa_preset_name, so a disabled preset stays reachable when the
user explicitly asks for it.
/moa no longer does a sticky model switch. It now always runs a single
prompt through the default MoA preset and restores the prior model
afterward; the whole argument is the prompt (no preset-name matching).
To switch to a MoA preset for the session, select it from the model
picker, where presets already surface under a virtual Mixture of Agents
provider on every model-selection surface.
Also fixes#53444: the TUI one-shot only set session[model_override],
which the already-built cached agent ignored, so MoA silently never ran
and the turn used the original model. The TUI now does a real in-place
agent.switch_model() via _apply_model_switch() when a live agent exists
(with a proper restore after the turn), and falls back to a model_override
for lazy/unbuilt sessions.
Removes the redundant sticky-switch branch from the CLI, gateway, and TUI
/moa handlers; updates the command description, usage string, and docs.
A MoA preset whose reference or aggregator slot points at the moa virtual
provider creates a recursive MoA tree. The runtime guards in moa_loop.py only
surface this mid-turn (references silently skipped, aggregator raises). Reject
it at the config chokepoint (_clean_slot) so it can never be saved, and hide it
from the desktop/dashboard slot pickers so it isn't offered as a dead choice.
_normalize_preset uses bare float() and int() to coerce
reference_temperature, aggregator_temperature, and max_tokens from
config.yaml. When a user hand-edits a non-numeric value (e.g.
max_tokens: "8k" or reference_temperature: "hot"), the coercion raises
ValueError. Since normalize_moa_config runs on every model-selection
and MoA turn (via resolve_moa_preset), the crash is unrecoverable and
blocks all MoA usage until the config is manually fixed.
Replace the bare casts with _coerce_float / _coerce_int helpers that
fall back to the default on TypeError/ValueError instead of raising.
* feat(moa): expose MoA presets as selectable virtual models
Reconstructed onto current main (PR #46081's base had diverged with no common
ancestor, marking the PR dirty so CI never dispatched). MoA is now a virtual
provider: each named preset is a selectable model under provider 'moa', and the
preset's aggregator is the acting model that answers and calls tools.
Reference models fan out in parallel via a bounded ThreadPoolExecutor (the same
batch pattern delegate_task uses) — all references dispatched at once, collected
when every one finishes, then handed to the aggregator. Output order is
preserved, failures and the MoA-recursion guard stay isolated per reference.
- Removed the old mixture_of_agents model tool and moa toolset.
- Added moa as a virtual provider in the provider/model inventory.
- /moa is shortcut behavior over model selection (default preset / named preset
/ one-shot prompt).
- Dashboard + Desktop manage named presets; presets appear in model pickers.
- Parallel reference fan-out in agent/moa_loop.py with regression test.
* fix(moa): thread moa_config through _run_agent to _run_agent_inner
The reconstructed gateway MoA wiring declared moa_config on _run_agent (the
profile-scoping wrapper) and used it inside _run_agent_inner, but the wrapper
never forwarded it — _run_agent_inner had no such parameter, so the runtime hit
NameError: name 'moa_config' is not defined on the compression-failure session
sync path. Add moa_config to _run_agent_inner's signature and forward it from
both wrapper call sites (multiplex and non-multiplex). Caught by
tests/gateway/test_compression_failure_session_sync.py on CI shard test(4).
* fix(moa): classify moa as a virtual provider in the catalog
The moa virtual provider has no PROVIDER_REGISTRY/ProviderProfile entry, so
provider_catalog() fell through to the default auth_type="api_key" with no
env vars — tripping two catalog invariants:
- test_provider_catalog: api_key providers must expose a credential env var
- test_provider_parity: every hermes-model provider must be desktop-configurable
moa already declares auth_type="virtual" in HERMES_OVERLAYS; consult that
overlay as an auth_type fallback so the catalog reports moa as virtual (no real
credential, no network endpoint). Exempt virtual providers from the desktop
parity union check the same way 'custom' is exempt — derived from the catalog,
not a hardcoded slug, so future virtual providers are covered too.
