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
On the MoA path agent.model/provider are the virtual preset name (e.g.
"closed") and "moa", which have no pricing entry. estimate_usage_cost()
returned None for the aggregator turn, so the `if amount_usd is not None`
guard skipped it and the session's estimated_cost_usd reflected only the
advisor fan-out — a ~50% undercount when the aggregator does the full acting
loop (verified: $0.91 advisor-only vs $1.96 true, aggregator = 54%).
MoAChatCompletions.create() now stashes the resolved aggregator slot as
last_aggregator_slot (exposed via MoAClient); conversation_loop reads it to
price the aggregator turn at its real model/provider. cost_source flips from
'none' to 'provider_models_api'.
Adds Vertex AI as a first-class provider for Gemini models via Vertex's
OpenAI-compatible endpoint. Vertex authenticates with short-lived OAuth2
access tokens (service-account JSON or ADC), not a static API key — the
missing piece behind the recurring requests (#13484, #12639, #56259).
- agent/vertex_adapter.py: OAuth2 token minting + refresh-on-expiry
(5-min margin), ADC->service-account fallback, global vs regional
endpoint URLs. Config precedence: env var > config.yaml > default.
- plugins/model-providers/vertex/: provider profile (auth_type=vertex),
reuses Gemini's extra_body.google.thinking_config translation.
- runtime_provider: vertex short-circuit BEFORE the credential pool so a
credentials-file path is never mistaken for a static API key; mints a
fresh token + computes base_url per resolve.
- run_agent + conversation_loop: _try_refresh_vertex_client_credentials()
re-mints the token and rebuilds the client on a mid-session 401, so a
long-lived gateway agent survives token expiry (~1h).
- auxiliary_client: vertex auth_type branch for side-LLM tasks.
- config.yaml: vertex.project_id / vertex.region (non-secret, bridged to
env); credential path stays in .env (VERTEX_CREDENTIALS_PATH).
- setup wizard + model picker: dedicated _model_flow_vertex; curated
google/gemini-* model list; --provider choices.
- pricing/metadata: Vertex prices off the gemini docs snapshot; endpoint
host auto-maps to the vertex provider (no probe spam).
- lazy_deps + pyproject [vertex] extra: google-auth, opt-in only.
- docs: guides/google-vertex.md + providers page; tests for adapter +
runtime resolution.
Salvages and modernizes #8427 by @slawt onto current main: rewired from
the legacy PROVIDER_REGISTRY path to the provider-profile architecture,
moved non-secret config out of .env into config.yaml, and added the
per-turn 401 token-refresh the original lacked.
MoA full-turn traces (moa.save_traces) recorded the aggregator's acting
output only on the non-streaming path, where it's captured inline at
call time. On the streaming path — which every hermes chat --query run
and every live gateway/CLI turn takes — the aggregator's raw token
stream is handed to the live consumer, so the trace left output=null and
only pointed at the session-db assistant row. An offline audit of a
benchmark run (HermesBench drives --query) then couldn't see what the
aggregator produced without hand-joining to state.db.
Capture the resolved streamed acting text at trace-flush time (the agent
already holds it in _current_streamed_assistant_text) and fold it into
the trace, so the record is self-contained in both modes. New
output_location value inline_from_stream marks a streamed turn whose text
was captured this way; a genuinely empty acting turn (pure tool call)
still points at the session db, matching state.db exactly.
Touches only the trace side-channel — no change to the acting path,
message history, role alternation, or prompt cache.
- agent/moa_loop.py: consume_and_save_trace(..., aggregator_output_fallback)
on both the facade and the MoAClient wrapper; prefer inline capture,
fall back to the resolved streamed text.
- agent/moa_trace.py: embed the fallback; add inline_from_stream location.
- agent/conversation_loop.py: pass _current_streamed_assistant_text at flush.
- tests: 5 cases across streaming / non-streaming / empty-fallback / no-double-write.
When text compression can't reduce a 413 request further, evict base64
image parts from tool messages and retry once instead of dead-ending
with 'Payload too large and cannot compress further.'
A 413 is a request-body byte-size limit, not a token limit. browser_vision
screenshots (2-5MB base64 each) keep the HTTP body oversized even after
aggressive summarization. The strip pass passes remember_model=False so a
413 does not poison _no_list_tool_content_models — that set is for providers
that reject list-type tool content, a distinct failure mode.
Cherry-picked from #47397 by Tranquil-Flow; placed onto main's current
token-aware 413 recovery else branch.
verify_on_stop / pre_verify append a synthetic assistant "done" plus a
synthetic user nudge to keep the agent going one more turn before it can
claim completion. Both were flagged (_verification_stop_synthetic on the
nudge only), but the flags were never registered in
_EPHEMERAL_SCAFFOLDING_FLAGS, so the central _is_ephemeral_scaffolding()
filter that guards both persistence sinks (SQLite flush + JSON snapshot)
let them through. The resumed transcript then inherited loop-only
scaffolding, invalidating the prompt-prefix cache on later turns.
- add _verification_stop_synthetic and _pre_verify_synthetic to
_EPHEMERAL_SCAFFOLDING_FLAGS (the single chokepoint both sinks use)
- flag the blocked attempt assistant message too, not just the nudge, so
the whole synthetic pair drops together and persistence does not keep a
premature done with the nudge stripped (assistant to assistant adjacency)
The API-payload leak claimed in the report is already handled: the
chat_completions transport strips every underscore-prefixed message key
before the wire, so the marker never reaches strict providers.
Reported by patppham.
AIAgent.run_conversation() promises a dict with final_response, but 16
terminal-failure branches returned dicts that either omitted the key or
set it to None. Callers that index result['final_response'] directly
(run_agent.py chat() + the __main__ printer) turn a real provider/context
failure into an opaque KeyError instead of surfacing the actionable error.
Every offending branch already carried usable 'error' text, so this
mirrors that text into final_response for all 16 sites (8 that omitted the
key, 8 that returned None). Adds an AST regression test that fails if any
run_conversation() dict return omits final_response or sets it to a literal
None, and tightens the invalid-response test to assert final_response == error.
When an Anthropic Claude Pro/Max OAuth subscription hits the "out of extra
usage" 400 (now classified as billing), surface actionable guidance pointing
at claude.ai/settings/usage and the cycle-reset option instead of the generic
"add credits with that provider" line — which does not apply to a
subscription. Folds in the UX from #40073 (@harsh-matchmyflight) without the
extra FailoverReason enum; the billing reclass already provides the recovery
behavior.
Adds moa.save_traces (default off). When on, every MoA turn that runs the
reference fan-out appends one JSON line to
<hermes_home>/moa-traces/<session_id>.jsonl capturing the TRUE FULL turn:
each reference model's exact input messages (system advisory prompt + full
advisory view, not the truncated display preview) + full output + usage +
per-advisor cost, and the aggregator's exact input (including the injected
reference-context guidance block) + output. Lets MoA runs be audited and
improved offline — what every model saw, said, and cost.
- agent/moa_trace.py: config-gated JSONL writer, profile-aware path via
get_hermes_home(), best-effort (never breaks a turn), moa.trace_dir override.
- agent/moa_loop.py: _RefAccounting now carries full input/output/model/
provider/temperature; create() stashes the full turn on a cache MISS
(once per turn, never on the cache-HIT repeat iterations); non-streaming
aggregator output captured inline, streaming marked + pointed at the
session assistant message. consume_and_save_trace(session_id) flushes it.
- agent/conversation_loop.py: flushes the trace with the live session_id
right after MoA usage consumption. No-op for non-MoA clients.
- hermes_cli/config.py: moa.save_traces + moa.trace_dir defaults.
Traces are a side channel — NOT the messages table, never in replay, safe
to delete. Off by default; only overhead when off is one config read on a
MoA cache-MISS turn.
Tests: full-trace-when-enabled (per-ref input+output+cost, aggregator
input-with-guidance + output), nothing-when-disabled. Live E2E through
run_conversation confirmed the loop wiring writes the file.
MoA ran the reference models before the aggregator but returned only the
aggregator's usage to the loop — _run_reference discarded each advisor
response's .usage entirely. Session accounting (state.db, /insights, cost)
therefore undercounted every MoA turn by the whole reference fan-out, which
is usually the bulk of the spend and scales with advisor count.
- _run_reference normalizes each advisor's usage with ITS OWN resolved
provider/api_mode and prices it at ITS OWN model rate (correct cache-read/
cache-write split), returning a _RefAccounting(usage, cost).
- create() sums advisor usage + cost once per turn (cache MISS only, so a
repeat tool-iteration reusing cached advice does not double-charge) and
exposes it via MoAClient.consume_reference_usage().
- conversation_loop folds advisor tokens into the reported/persisted token
counts and adds advisor cost (priced per-advisor) on top of the
aggregator cost, in both the in-memory session totals and the state.db
per-call delta. Aggregator cost is still priced on aggregator-only usage
so advisor tokens are never repriced at the aggregator rate.
- CanonicalUsage gains __add__ for per-bucket summing.
Tests: advisor usage/cost capture, per-turn sum + consume-clears +
cache-hit no-double-charge, CanonicalUsage.__add__.
MoA sessions could not stream: the gateway streaming toggle was a no-op for
provider "moa", so users saw nothing until the entire response finished — minutes
of silence on long turns. The aggregator's reply was always fetched whole.
Root cause was twofold:
1. conversation_loop hard-disabled streaming for provider in {"copilot-acp",
"moa"} (MoA grouped with the ACP client, whose facade isn't a stream).
2. MoAChatCompletions.create() fetched the aggregator response whole via
call_llm(), which had no streaming mode.
For provider "moa", _create_request_openai_client() returns the MoAClient facade
itself, so the existing streaming consumer already calls
MoAChatCompletions.create(stream=True). We reuse that battle-tested consumer
(text-delta delivery, tool_call reassembly, stale-stream detection, non-streaming
fallback) instead of adding a parallel streaming path.
Changes:
- call_llm() gains stream/stream_options. When streaming it returns the raw SDK
stream iterator directly, bypassing _validate_llm_response and the
temperature/max_tokens/payment fallback chain (which assume a complete
response). The caller owns reassembly and fallback.
- MoAChatCompletions.create() runs the references first (unchanged), then when
stream=True returns the aggregator's raw stream, forwarding stream_options and
the consumer's per-request read timeout. stream=False is byte-identical to
before (no stream/stream_options/timeout forwarded).
- conversation_loop streams MoA only when a display/TTS consumer is present;
quiet/subagent/health-check paths keep the complete-response path.
Tests: tests/run_agent/test_moa_streaming.py — create() stream/non-stream
branches, stream_options + timeout forwarding, call_llm raw-stream return vs
validated non-stream. Existing MoA tests unchanged (20 passed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
OpenRouter returns 429 in two shapes: an account-level throttle on the
user's key, and an upstream-provider throttle (DeepSeek/Anthropic/etc.
rate-limiting OpenRouter's aggregate traffic). The classifier treated
both identically and rotated/exhausted OPENROUTER_API_KEY on every 429 —
burning the key for ~24min and silently disabling auxiliary features
(compression, summarization, vision) on an upstream throttle where the
key was healthy.
Add a FailoverReason.upstream_rate_limit classified from OpenRouter's
unambiguous wrapper message "Provider returned error" (the same signal
the metadata-raw parser already trusts). Recovery skips credential
rotation and defers to the fallback chain to switch models instead.
Co-authored-by: Hermes Agent <127238744+teknium1@users.noreply.github.com>
An over-cap model.max_tokens produces a provider 400 that mentions
max_tokens, which trips _CONTEXT_OVERFLOW_PATTERNS and is classified as
context_overflow. On providers whose wording isn't recognized by
parse_available_output_tokens_from_error() (e.g. DashScope/Qwen:
"Range of max_tokens should be [1, 65536]") the smart-retry is skipped
and the error falls into the compression fallback, which re-sends the
same oversized max_tokens, fails identically, and loops until
"cannot compress further" on a tiny conversation (#55546).
Root-cause fix for the whole class, not just DashScope:
- parse_available_output_tokens_from_error(): recognize the DashScope
"Range of max_tokens should be [1, N]" form and return N (smart-retry
then caps output and retries WITHOUT compressing).
- new is_output_cap_error(): broader yes/no gate for output-cap 400s.
In the loop, when the error is output-cap-shaped but unparseable, fail
fast with an actionable message (lower model.max_tokens) instead of
routing into compression. Mirrors the existing GPT-5 max_tokens guard.
Real input overflows and GPT-5 unsupported-param 400s are unchanged.
Add a `pre_verify` user/plugin/shell hook fired once per turn when the agent
edited code and is about to finish, after the existing verify-on-stop guard. A
hook can keep the agent going one more turn (run a check, defer it, tidy the
diff) by returning {"action":"continue","message":...} (the Claude-Code Stop
shape {"decision":"block","reason":...} is accepted too). Hooks receive coding,
attempt, final_response, and sorted changed_paths so they can self-scope and
self-throttle; the path is bounded by agent.max_verify_nudges and preserves
message-role alternation.
Hermes still ships its default coding guidance (agent.verify_guidance, on by
default), but it now rides the evidence-based verify-on-stop missing-evidence
nudge instead of a separate default pre_verify continuation, so it costs no
extra model turn of its own. Guidance reuses the shared utils.is_truthy_value
parser rather than a local copy.
When a provider's output-layer safety filter (MiniMax "output new_sensitive
(1027)", Azure content_filter, etc.) kills a streaming response after deltas
were already sent, interruptible_streaming_api_call swallows the raw error
into a finish_reason=length partial-stream stub. The conversation loop then
burned 3 continuation retries against the SAME primary — re-hitting the
content-deterministic filter every time — and gave up with "Response remained
truncated after 3 continuation attempts", never consulting fallback_providers.
Builds on @595650661's classifier change (cherry-picked) so error_classifier
recognizes the filter; then:
- chat_completion_helpers: run the swallowed error through error_classifier at
the stub-creation point and stamp _content_filter_terminated on the stub
(single source of truth — no parallel pattern list).
- conversation_loop: read the tag and activate the fallback chain BEFORE
burning any continuation retries; roll partial content back to the last
clean turn and re-issue against the new provider (restart_with_rebuilt_messages).
Plain network stalls are unaffected (only content_policy_blocked is tagged).
Credits #32479 (@sweetcornna) and #33845 (@Tranquil-Flow) which fixed the
same issue via the stub-tag and loop-escalation approaches respectively.
Live E2E confirmed: before, _try_activate_fallback called 0x; after, fallback
fires on the first stub and the fallback provider completes the turn.
Salvage of NousResearch/hermes-agent#41498 (0-CYBERDYNE-SYSTEMS-0).
- Leave response_previewed false on partial_stream_recovery so gateway
fallback delivery can send the recovered fragment plus explanation.
- Always append the turn-completion explainer for partial_stream_recovery,
not only for empty or very short fragments (#34452 gap).
- Launch the detached /restart helper before drain, idempotently, with a
bounded wait of restart_drain_timeout + 5s.
* fix(agent): config-driven intent-ack continuation for all api_modes (#27881)
The agent could end a turn after only stating intent ('I will run a health
check...') without executing the announced tool call, forcing the user to
re-prompt. A continuation guard that catches this and nudges the model to
proceed already existed but was hard-gated to the codex_responses api_mode,
so Gemini/Claude/OpenRouter turns never benefited.
- New agent.intent_ack_continuation config (default 'auto' = codex-only,
byte-stable for existing conversations). 'true'/model-list opts every
api_mode in; 'false' disables. Mirrors agent.tool_use_enforcement's shape.
- looks_like_codex_intermediate_ack gains require_workspace (default True).
The opted-in path drops the codebase/filesystem requirement so general
autonomous workflows (server ops, deploys, API calls) are caught, not just
coding tasks. Future-ack + action-verb + short-content + no-prior-tool
guards still apply; the 2-nudge-per-turn cap is unchanged.
- Resolution centralized in intent_ack_continuation_mode (off/codex_only/all).
* docs(infographic): intent-ack continuation (#27881)
The Anthropic SDK clients were built without max_retries, so the SDK
default (max_retries=2) retried 429/5xx with its own backoff that ignores
Retry-After — double-retrying inside hermes's outer loop and burning
request slots against a bucket that won't refill for minutes. Set
max_retries=0 on all Anthropic/AnthropicBedrock client constructions so
the outer conversation loop (which already honors Retry-After) owns retry.
Also raise the Retry-After cap in the conversation loop from 120s to 600s.
Anthropic Tier 1 input-token buckets reset in ~171s, so the 120s cap made
hermes retry before the reset window and re-trip the limit.
Refs #26293
A persistent upstream 401 on a single-entry OAuth pool (common for Claude
Max subscribers) made the credential-pool recovery spin forever:
try_refresh_current() re-mints a fresh token and reports success on every
401, so recover_with_credential_pool returned True and the retry loop
continue'd without ever incrementing retry_count or reaching the
auth-failover block. The configured fallback_model never activated and the
agent appeared to hang.
Cap consecutive successful same-entry refreshes (keyed by provider +
pool-entry id) at 2; once exceeded, treat the credential as unrecoverable
and return not-recovered so the loop falls through to
_try_activate_fallback. The 429/billing paths already rotate-or-fall-through
correctly (mark_exhausted_and_rotate returns None on a single entry), so
only the auth-refresh branch needed the cap.
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
Eager fallback previously fired only on rate_limit/billing. A stale-
detector-killed hung stream classifies as FailoverReason.timeout
(retryable=True) and the retry loop re-hit the same dead primary until
the budget exhausted -- 3 x ~180-300s stale kills compounding into a
15+ min silent hang while the configured fallback chain sat idle.
Extend the existing eager-fallback gate to also cover timeout and
overloaded, but only after one real retry (retry_count >= 2) so genuine
transient hiccups still recover on the primary. Reuses the same
pool-recovery guard and state-reset as the rate_limit branch -- no new
config flag, no change to the rate-limit intent.
Salvaged from PR #50228 by @linyubin. Closes#22277.
Co-authored-by: Hermes Agent <127238744+teknium1@users.noreply.github.com>
The agent's image-rejection fallback strips images and retries text-only when
a provider rejects image content, which is what lets the gateway drain its
queued messages. The fallback only fires on a hardcoded phrase list, and the
OpenRouter wording — HTTP 404 'No endpoints found that support image input' —
was missing. For OpenRouter-routed non-vision models the fallback never fired,
the retry loop re-sent the same rejected request until exhaustion, and every
subsequent message (including plain text) stayed queued behind the stuck turn.
Add the phrase to _IMAGE_REJECTION_PHRASES (the 404 already passes the 4xx
gate). Add a positive test and a guard test so the sibling OpenRouter
'no endpoints ... data policy / guardrail' 404s do NOT get their images
stripped.
Fixes#21160. Reported by @liu14goal14-ux; PR #21198 by @ygd58.
MoA was calling reference and aggregator models through a bare
call_llm(provider=slot["provider"], model=slot["model"]) with a forced
temperature and a forced max_tokens (the preset's hardcoded 4096). That left
base_url/api_key/api_mode unresolved — so the auxiliary auto-detector guessed
the API surface instead of using the provider's real runtime, and the 4096 cap
truncated long aggregator syntheses.
A MoA slot is just a model selection and must be called the same way any model
is called elsewhere. Each slot is now resolved through resolve_runtime_provider
(the canonical provider→api_mode/base_url/api_key resolver the CLI, gateway, and
delegate_task all use) via a new _slot_runtime() helper, and the resolved
endpoint is passed into call_llm. So a reference/aggregator gets its provider's
actual API surface — MiniMax → anthropic_messages, GPT-5/o-series →
max_completion_tokens, custom endpoints → their base_url — identical to how that
model is handled as the acting model.
MoA also no longer imposes its own output cap: max_tokens defaults to None
(omitted → the model's real maximum) for references and is passed through from
the caller for the aggregator. The preset's hardcoded 4096 is gone. The
max_tokens preset config field is left in place (config/web/desktop unchanged);
it is simply no longer applied as a forced cap.
Tests: slots route through resolve_runtime_provider with resolved base_url/
api_key; resolution errors fall back to bare provider/model; neither call
carries an output cap even when the preset config still contains max_tokens.
The verify-on-stop guard (#52296) printed '↻ Verification required before
finishing' to the terminal on every internal nudge turn, adding noise to
CLI/gateway sessions whenever code was edited without fresh passing checks.
Demote the user-facing status emit to a logger.debug breadcrumb — the loop
still nudges the model to verify before finishing, just silently.
Stopping a turn while the model is streaming (stop/esc to redirect) raised
InterruptedError, set final_response to the throwaway "waiting for model
response" sentinel, and persisted messages WITHOUT the assistant text that
was already streamed to the screen. The next turn then had no record of the
half-finished reply, so the model appeared to "forget" what it just said.
Recover the on-screen text from _current_streamed_assistant_text in the
InterruptedError branch and append it as the assistant turn (and surface it
as final_response). The metadata sentinel is kept only when nothing was
streamed yet, preserving the ACP/client suppression behavior.
Completes the partial-stream recovery from 397eae5d9 (which wired the same
_current_streamed_assistant_text salvage into the connection-failure twin
but missed the user-interrupt path). The lossy handler dates to c98ee9852.
Two-part fix:
Part 1 (classifier override at agent/error_classifier.py:720-738):
A transport disconnect on a reasoning model — even on a large session —
now routes to FailoverReason.timeout instead of context_overflow. Without
this, large-session reasoning-model disconnects route to the compression
branch and silently delete conversation history on a phantom
context-length error. The override is strictly targeted: non-reasoning
models (gpt-4o, claude-3-5-sonnet, llama-3.3-70b, etc.) still route to
context_overflow on large sessions — the existing intentional behavior
for chat models whose proxy doesn't idle-kill during prefill/generation.
Part 2 (new agent/thinking_timeout_guidance.py + integration at
agent/conversation_loop.py:3488-3567):
New is_thinking_timeout() and build_thinking_timeout_guidance() helpers.
When a known reasoning model (NVIDIA Nemotron 3 Ultra, OpenAI o1/o3,
Anthropic Opus 4.x thinking, DeepSeek R1, Qwen QwQ, xAI Grok reasoning)
hits a transport-kill on a small session (classifier says timeout
directly) or after Part 1 routes correctly (large session), the user
now sees reasoning-specific guidance with three actionable workarounds
in priority order:
1. Set providers.<provider>.models.<model>.stale_timeout_seconds: 900
in ~/.hermes/config.yaml (Hermes's built-in floor is already 600s
for known reasoning models; raise further if upstream is even
tighter).
2. Lower reasoning_budget or set reasoning_effort: medium on this
model if the provider supports it.
3. Use a smaller / faster reasoning model if the task doesn't
require deep thinking.
The new guidance takes precedence via if/elif over the existing
_is_stream_drop block, so a reasoning-model user with a transport-kill
message sees actionable advice instead of the misleading "try
execute_code with Python's open() for large files" advice (which is
correct for the unrelated large-file-write stream-drop case but
actively wrong for the thinking-timeout case).
Verified:
- 478 tests passing across 9 directly-relevant files (49 new + 429
existing, zero regressions).
- Ruff lint clean on all 4 modified/new files.
- Negative test: 6 parametrized regression guards confirm non-reasoning
models still route to context_overflow on large sessions; 4
parametrized gates confirm non-timeout classifier reasons never
trigger the guidance; 5 parametrized cases confirm non-transport
messages never trigger it.
- Regression guard: new guidance message does NOT contain
"execute_code" or "open()" — the misleading advice is fully
replaced, not appended alongside.
- Cross-vendor dual review via agy -p:
- Gemini 3.5 Flash (Medium) — passed: true, zero blockers, one
SHOULD-FIX (vprint block duplication — fixed by extracting
detection into a helper module).
- GPT-OSS 120B (Medium) — passed: true, zero blockers, two nits
(test placement — adopted at tests/agent/test_thinking_timeout_guidance.py;
primary-model capture — accepted as non-issue per Flash's nit).
Dependency note for maintainers:
This PR includes agent/reasoning_timeouts.py (the reasoning-model
allowlist module from PR #52238) because the Layer 1 override is
load-bearing on get_reasoning_stale_timeout_floor(). After PR #52238
lands on main, this PR's duplicate agent/reasoning_timeouts.py should
be rebased away. Either PR can land first; the other rebase is
mechanical.
Fixes#52271.
* 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.
#48879 closed the tool-call sequence on interrupt inside finalize_turn so a
/stop after a tool no longer persists a `tool` tail that the next user message
turns into a `tool -> user` role-alternation violation (which strict providers
like Gemini/Claude react to by hallucinating a continuation and ignoring prior
context — what users see as "lost context after stop").
But the retry-wait, error-handling, and post-error retry-wait interrupt aborts
in conversation_loop return early and never reach finalize_turn, so they still
persisted and returned a raw `tool` tail. Interrupting during provider
backoff/rate-limiting (common under heavy work) hit exactly this path.
Extract the close into a shared close_interrupted_tool_sequence helper and apply
it at every interrupt abort (finalize_turn + the three early returns) so the
whole bug class is fixed, not just the one site.
Make verification closure the default coding behavior after landed file edits while keeping bounded retries and config/env switches for users who need to disable it.
Compression can materially reduce request size (tool-result pruning,
in-place summarization) without reducing message count. The two
compression-success checks in conversation_loop.py (413 handler and
context-overflow handler) only compared len(messages) to detect
success, missing token-only compression.
Now re-estimates tokens after compress_context() returns and treats
any >=5% reduction as a successful compression pass. Error logs
also use the post-compression token count instead of the stale
pre-compression estimate.
Fixes: #39550
When the active provider returns a 401/403 that survives its per-provider
credential-refresh attempt (revoked OAuth, blocked/expired key, or an
account pinned to a dead/staging inference endpoint), the conversation
loop now escalates to the configured fallback chain instead of dead-ending.
Before: the generic failover dispatch fired only for {rate_limit, billing};
auth/auth_permanent fell through to 'switch providers manually' advice and
never called _try_activate_fallback(). A user whose primary credential was
broken kept thrashing on the same dead credential every turn — the main
agent appeared 'stuck in fallback mode' while never actually failing over.
This also affected auxiliary tasks (compression, vision, title-gen), since
auto-resolved aux follows the main provider.
After: a persistent auth failure with a configured fallback chain switches
to the next provider (mirroring the rate-limit/billing failover path),
guarded one-shot per attempt by TurnRetryState.auth_failover_attempted.
When no fallback is configured the behavior is unchanged — it falls through
to the existing terminal handling and provider-specific troubleshooting
guidance.
Tests: test_auth_provider_failover.py — 401/403 classify as auth, the
gating condition fires only with a chain present + guard unset, the guard
blocks repeats, and non-auth (500) errors do not trigger auth failover.
The session-stable system prompt embeds Model:/Provider: identity lines,
but mid-turn failover (try_activate_fallback) swaps the runtime without
touching them, so a fallback model misreports itself as the primary when
asked "what model are you?".
rewrite_prompt_model_identity() rewrites the last occurrence of each line
on _cached_system_prompt when a fallback activates (and back on restore,
byte-identical so the primary's prefix cache still hits). The rewrite is
never persisted to the session DB. _sync_failover_system_message() patches
the in-flight api_messages[0] at all 8 failover sites so the current turn
ships the corrected identity. Cache-safe: the fallback's prefix cache is
cold on a model switch anyway.
Co-authored-by: Hermes Agent <noreply@nousresearch.com>
The non-retryable abort path now computes _nonretryable_summary once and
reuses it at the emit sites and the returned error field. The
content-policy-blocked return branch still recomputed the identical
value into a separate _summary local, half-honoring the 'summarize once'
intent. _summarize_api_error is a pure staticmethod and api_error is
never reassigned in this block, so _summary was provably byte-identical
to _nonretryable_summary. Reuse the hoisted value and drop the redundant
call. Behavior-preserving.
When a non-retryable client error aborts the turn (e.g. a Codex/Cloudflare
HTTP 403 "managed challenge" page), the conversation loop returned the
failure dict with `error: str(api_error)` — the entire ~60KB HTML page.
Downstream consumers deliver that field verbatim: a cron job dumped a
Cloudflare challenge page to Discord, where it was split into ~31 messages.
The sibling "max retries exhausted" path already collapses such bodies via
`_summarize_api_error` (which extracts the <title> / status from HTML error
pages). This makes the non-retryable path consistent: compute the summary
once and use it for both the status emit and the returned `error`.
Weak open models (mimo, nemotron-class) that see tool-call XML/JSON sitting in
file contents or tool output get primed and emit their own structured tool
calls mimicking the payload — usually with an empty/whitespace name. Those
calls can't be fuzzy-repaired toward a real tool, so the dispatch loop returns
an error and the model retries. Before this fix, every empty-name error dumped
the full tool catalog back to the model, which fed the priming loop more names
to mimic and inflated context 3-4x across the retry budget.
A blank/whitespace-only tool name now gets a terse anti-priming error that
tells the model in-context tool-call syntax is DATA, with no catalog dump. A
genuinely-wrong-but-nonempty name (a real typo) still gets the full catalog so
the model can self-correct.
Not a sandbox/auth boundary issue: Hermes never parses tool-call text from
content into executable calls (structured tool_calls only; the lone text->call
parser is the Copilot ACP transport and it also rejects empty names). The
reporter's own debug dump confirms the injection never executed.
Behavior-contract test added: empty-name -> terse error, no catalog; nonempty
unknown -> catalog preserved. Exercised end-to-end via run_conversation against
an in-process mock provider.
Consolidates these related Amy fork patches:
- 429830f39 feat(gateway): inject message timestamps into user messages for LLM context
- 3c3d6fac0 fix: handle both ISO string and epoch float timestamps in history replay
- 2874f7725 feat: human-friendly timestamp format with weekday and timezone name
- 3735f4c8b fix: render gateway message timestamps once
A live config.set model switch already moved the next API call to the new model,
but the conversation could still restore an old sessions.system_prompt snapshot
whose Model/Provider lines named the previous runtime. That made "what model are
you?" answer from stale metadata even while inference ran on the new model.
After a live switch we now refresh the stored system prompt and append a real
system-history pivot (not a fake user turn) so the transcript itself records the
new model/provider. Restore also rejects already-stale prompt snapshots when
their Model/Provider lines disagree with the runtime, so existing bad sessions
self-heal.
The HTTP-200 refusal handler (finish_reason=content_filter) and the
exception-path handler (a provider moderation error classified as
content_policy_blocked) independently built the same terminal turn result —
the same {final_response, messages, api_calls, completed:False, failed:True,
error:'content_policy_blocked: ...'} dict — and ended their user-facing
message with the same 'Try rephrasing... hermes fallback add' trailer, copied
verbatim. The two copies could drift.
Funnel both through a shared _content_policy_blocked_result() builder and a
shared _CONTENT_POLICY_RECOVERY_HINT constant. Also collapse the HTTP-200
path's two near-identical with/without-explanation templates into one (compute
the detail fragment once) and pass reason=FailoverReason.content_policy_blocked
.value to the error hook instead of a hand-written string literal, matching the
sibling hook call.
Behavior-preserving: the provider/refusal lead-in wording stays distinct (a
provider safety filter vs the model declining are genuinely different signals),
the with-text and exception messages are byte-identical to before, and the
no-explanation case only gains a paragraph break for consistency. Surfaced by
the simplify-code reuse/quality reviewers.
The efficiency reviewer's 'redundant normalize_response' flag was deliberately
NOT applied: that branch is cold (refusal-only) and pure-CPU, and reusing the
sibling-branch normalized locals would risk a NameError on the codex_responses
path (which sets finish_reason without normalizing) — re-normalizing is the
robust choice.
A Claude refusal (HTTP 200, stop_reason="refusal", empty content) was
laundered into a generic retry loop and surfaced as a misleading
"rate limited / invalid response" or "no content after retries" error,
burning paid attempts reproducing a deterministic refusal.
This hit two distinct paths:
- Direct Anthropic (anthropic_messages): validate_response rejected the
empty-content refusal *before* normalize_response mapped refusal ->
content_filter, so it fell into the invalid-response retry loop.
- Nous Portal / OpenAI-compatible (chat_completions): the portal surfaces
a Claude refusal via message.refusal with empty content, which sailed
past validation and died in the empty-response retry loop.
Fix (one unified content_filter dispatch for all backends):
- AnthropicTransport.validate_response: accept empty content when
stop_reason == "refusal" so it flows to normalize_response.
- ChatCompletionsTransport.normalize_response: promote message.refusal to
content + a content_filter finish reason.
- conversation_loop: handle finish_reason == "content_filter" - fire the
api_request_error hook (content_policy_blocked), try a configured
fallback once, else return a clear terminal refusal message. Never retry
a deterministic refusal.
Supersedes #43084, which fixed only the direct-Anthropic path and could
not reach the chat_completions/portal path.
Tests: transport-level (validate_response refusal, message.refusal
promotion) + end-to-end loop (refusal surfaced, exactly one API call).
(cherry picked from commit 01f546f92cb1629ec1427be270dbd7c504e962ad)
Parse provider-reported image pixel ceilings so many-image Anthropic requests can recover by shrinking Retina screenshots below the stricter limit instead of retrying the same rejected payload.
Follow-up to the #44837 clamp: a min() clamp only fixes cursor overshoot
past the new end of the list. When repair_message_sequence drops/merges
messages at indexes below the cursor, the clamp leaves the cursor pointing
past unflushed rows and the turn-end flush silently skips them.
Extract repair_message_sequence_with_cursor(): snapshot the flushed prefix
by object identity before repair, then recompute the cursor as the count
of surviving flushed messages. Falls back to the clamp when no snapshot is
available. Keeps the safety guard in _flush_messages_to_session_db.
Adds targeted tests for overshoot, before-cursor compaction, no-repair,
bare-agent, and the flush guard.
The genuine-rate-limit branch set retry_count = max_retries before
continue, intending the top-of-loop Nous guard to handle fallback or
bail cleanly. But the loop condition is retry_count < max_retries, so
the guard never ran: no fallback activation, no clean rate-limit
message — just the generic retry-exhaustion error.
Set retry_count = max(0, max_retries - 1) so the loop body runs exactly
once more and the guard sees the breaker state recorded moments earlier.
Extracted from the #44061 bugfix rollup by @AIalliAI.
The thinking-signature recovery in agent/conversation_loop.py popped
reasoning_details from messages, then continued to retry. That had two
defects.
First, the strip never reached the wire payload. api_messages is built
once at the start of the turn by shallow-copying every entry in messages
(line 919 area). Each api_messages entry has its own reference to the
same reasoning_details list. When build_api_kwargs runs on every retry
iteration of the inner while-loop, it consumes api_messages, not
messages. Popping reasoning_details from messages left api_messages
untouched, so the retry's request still carried the same thinking
blocks Anthropic had just rejected. The classifier latched
thinking_sig_retry_attempted = True after the first attempt, and the
loop terminated with max_retries_exhausted on the same 400.
Second, the pop mutated the canonical message list. messages is the
same list _persist_session writes to state.db and the session
transcript, so a single recovery permanently wiped every signed
thinking block from the stored conversation. Subsequent turns reloaded
the stripped state, hit the same 400 ('invalid signature' or 'cannot
be modified', see #24107), and the agent stopped responding entirely.
Cascading compaction-ended sessions then chained off the corrupted
parent and the affected chat could not produce a response on any
future turn.
Move the strip onto api_messages, which is the API-call-time list
rebuilt into kwargs on every retry. messages is no longer touched, so
disk I/O stays clean and the recovery actually reaches the wire.
Observed against the native Anthropic Messages API on claude-opus-4-7
and claude-opus-4-8 with the interleaved-thinking-2025-05-14 beta on
hermes-agent 0.12.0 and 0.14.0. PR #24107 narrows the trigger; this
change makes the recovery do what it always claimed to do, and
prevents the destructive aftermath.
Tests cover the api_messages strip in isolation: pop on a shallow copy
does not affect the source, the canonical messages list survives the
strip, idempotency on a duplicate firing path, and a no-op when no
reasoning_details exist on the messages.
Related: #24107, #26959, #17861.
