When OpenRouter routes to an endpoint that does not support tool/function
calling, it returns HTTP 404 with the message 'No endpoints found that
support tool use. Try disabling "browser_back".'
The raw error body does not contain 'model not found' or any other
_MODEL_NOT_FOUND_PATTERNS entry, so it falls through to FailoverReason.unknown
with retryable=True. The retry loop wastes 3-5 attempts on the same
deterministic rejection, then surfaces a confusing generic error instead of
automatically failing over to a fallback model or provider.
Adding the OpenRouter phrase to _MODEL_NOT_FOUND_PATTERNS classifies it as
model_not_found (retryable=False, should_fallback=True), which triggers the
client-error fast-fallback path in conversation_loop.py: the agent switches
to a configured fallback model/provider before the user sees the error.
Existing buffered guidance in conversation_loop.py (the 'support tool use'
hint at line ~2967) remains intact and surfaces only if every fallback
exhausts.
_try_anthropic() hard-failed (return None, None) when the anthropic
credential pool was present but had no selectable entry — e.g. the pooled
OAuth token expired and its refresh_token had gone stale, so
_select_pool_entry("anthropic") returned (True, None). This wedged every
auxiliary task routed to Anthropic (goal judge surfaced "no auxiliary
client configured") even when a perfectly valid ANTHROPIC_TOKEN /
credentials-file token was available. The main session stayed healthy
because it resolves the env token directly.
The openrouter path (_try_openrouter) and codex path already fall through
to their standalone credential on (True, None); anthropic was the only
provider that hard-failed. Make _try_anthropic fall through to
resolve_anthropic_token() on that branch so the three paths are symmetric:
a temporarily dead pool entry must not block auxiliary tasks when a valid
standalone credential exists.
Adds a regression test covering: (1) pool present + no entry + valid env
token -> client built from the env token, (2) pool present + no entry + no
resolvable token -> clean (None, None), (3) base_url defaults correctly
when falling through with pool_present=True.
Strict providers (DeepSeek) reject a payload where the same tool_call_id
appears more than once with HTTP 400 'Duplicate value for tool_call_id'.
The issue was filed as an 'orphaned tool message' compression bug, but the
pasted error is a DUPLICATE tool_call_id — orphans are already handled on
main; duplicates were not. Reproduced live on main: both shapes leaked
through repair_message_sequence and sanitize_api_messages.
Two chokepoints, two shapes:
- repair_message_sequence: consume the id from known_tool_ids on first
match so a SECOND tool result reusing it falls into the drop branch
(duplicate tool-result shape). This is @Robinlovelace's kernel from
#55436 (applied manually — that PR was ~800 commits stale and bundled
an unrelated duplicate-DB-write change for #860, which is dropped here).
- sanitize_api_messages (final pre-API pass): add a dedup pass covering
BOTH (a) duplicate tool_calls sharing an id WITHIN one assistant message
(the message[6] shape) and (b) later tool result messages reusing an
already-seen id. #55436 covered neither of these at this chokepoint.
Tests: duplicate-tool-result dedup at both functions, duplicate-assistant-
tool_call-id collapse, and a negative control proving distinct ids are
never dropped (no over-dedup).
Credit: @Robinlovelace (#55436) for the repair_message_sequence dedup kernel.
Closes#58327.
Replace the inline dict-copy + _db_persisted pop in
_ensure_compressed_has_user_turn with the canonical
_fresh_compaction_message_copy helper (the same primitive the compressor's
own protected-head/tail assembly uses), so the persistence-marker strip
stays consistent across all compaction copy sites (#57491). Expand the
docstring to record the alternation-safety and end-placement rationale.
22c5048d9 restored Anthropic-style cache_control for two of MoA's three
call paths: the acting aggregator (MoAChatCompletions.create, the
persistent `provider: moa` model) and the advisor fan-out (_run_reference).
aggregate_moa_context() -- the /moa <prompt> one-shot command's synthesis
call -- is the third, independent call path and was never covered: its
call_llm(task="moa_aggregator", ...) sent a single undecorated user message
containing the full joined reference output, re-billing the entire input on
every invocation even when the resolved aggregator slot is a cache-honoring
route (Claude on OpenRouter/native Anthropic, MiniMax, Qwen/DashScope).
- Generalize _maybe_apply_advisor_cache_control to
_maybe_apply_moa_cache_control (it never had advisor-specific logic --
same policy function, same breakpoint layout as the main loop, judged
purely on the passed-in runtime) and reuse it in aggregate_moa_context
the same way _run_reference already does.
- Compute _slot_runtime(aggregator) once and reuse it for both the
decoration call and the call_llm kwargs, instead of calling it twice.
Mutation-verified: reverting the moa_loop.py change makes the new
regression test fail by asserting a plain string aggregator-message
content where the cache-honoring case expects native cache_control
content blocks.
Both wired against features the iron-proxy author (@mslipper) confirmed on
PR #30179 — and both verified present in the pinned v0.39.0 source.
Header-auth providers (match_headers):
- New _HEADER_AUTH_PROVIDERS: Anthropic native (x-api-key), Azure OpenAI
(api-key on *.openai.azure.com / *.cognitiveservices / *.services.ai),
Gemini (x-goog-api-key + ?key= query param via match_query).
- TokenMapping grows match_headers + alias_env_names; per-provider header
sets flow into the secrets rules; mappings.json roundtrips them
(legacy files load with the Authorization default).
- GEMINI_API_KEY / GOOGLE_API_KEY collapse into ONE mapping (two
require-rules on the same host would reject each other); the sandbox
gets the token under both names, and the proxy child env mirrors the
alias into the canonical name when only the alias is set.
- Docker backend injects alias env names alongside canonical ones.
- The fail-closed tier is now empty, so fail_on_uncovered_providers and
discover_blocked_providers are deleted (dead toggle otherwise);
_NON_BEARER_PROVIDERS shrinks to genuinely-unswappable signature auth
(AWS SigV4, GCP service-account OAuth) — warn-only, as before.
Management API (hot reload):
- Generated proxy.yaml enables the v0.39 management listener: loopback
only at tunnel_port+2, bearer key from HERMES_IRON_PROXY_MGMT_KEY.
- Key minted at setup (management.token, 0600); start_proxy injects it
(v0.39 refuses to start when api_key_env is empty).
- hermes egress reload -> POST /v1/reload: re-reads proxy.yaml and
atomically swaps the pipeline; 422 leaves the running ruleset
untouched; actionable errors for not-running / pre-management config /
key mismatch. Secrets changes still require restart (daemon env is
read at spawn) — the CLI says so.
Validation: 218/218 unit+CLI+docker tests; 3/3 gated live E2E against the
real v0.39.0 binary (Authorization swap, x-api-key swap, live reload with
token rotation on the same pid). Docs updated.
Two review findings on the #57922 salvage:
1. Stale inline comment at the login-exchange site still claimed the token
endpoint uses the claude-code/ UA prefix and 404s claude-cli/ — now
contradicts the axios/ fix. Repointed it at _OAUTH_TOKEN_USER_AGENT.
2. The inherited Path.home test isolation on the three TestRefreshOauthToken
tests only stubbed the ~/.claude *file* source, not the macOS Keychain.
_refresh_oauth_token re-reads read_claude_code_credentials() (keychain
first) in its adopt-already-refreshed branch, so on any macOS dev/CI runner
with real Claude Code creds the branch short-circuits and the 3 tests fail.
Stub read_claude_code_credentials -> None so the tests are hermetic.
(The remaining TestResolveAnthropicToken/TestResolveWithRefresh/TestRunOauthSetupToken
failures on macOS are the same pre-existing keychain-leak class on origin/main,
unrelated to this OAuth-UA fix, and pass in CI — left out of scope.)
hermes auth add anthropic fails 100% at token exchange with HTTP 429 while
Claude Code /login succeeds through the same client_id/redirect/scope. The
discriminator is the User-Agent on the /v1/oauth/token request.
Verified live against platform.claude.com (throwaway code, nothing burned):
claude-code/2.1.200 (external, cli) -> 429 rate_limit (Hermes, blocked)
Mozilla/5.0 -> 429 rate_limit
axios/1.7.9 -> 400 invalid_grant (reached validation)
node / empty / SDK-style UAs -> 400 invalid_grant
Anthropic now rate-limits token-endpoint requests whose UA starts with
claude-code/ (the anti-abuse net for Max-sub-as-API-key). This is the same
prefix-block shape that #48534 first hit on claude-cli/, then #56263 dodged by
switching to claude-code/ — which held ~2 weeks and is now blocked too. Bumping
_CLAUDE_CODE_VERSION_FALLBACK cannot help; the gate is prefix-based.
Fix: shared _OAUTH_TOKEN_USER_AGENT (axios/) on the token endpoint only — the
two refresh POSTs (refresh_anthropic_oauth_pure) and the login exchange POST
(run_hermes_oauth_login_pure). The real Claude Code CLI exchanges the auth code
with a bare axios client, NOT its claude-code/ inference UA.
The INFERENCE client (build_anthropic_kwargs, /v1/messages) is deliberately left
on claude-code/ + x-app: cli — that fingerprint is required there and is NOT
throttled on the messages API. Two endpoints, opposite UA requirements.
Also isolate two _refresh_oauth_token tests from live ~/.claude creds and update
the UA regression tests to assert the split (token endpoint uses a
non-claude-code UA while inference keeps claude-code/).
Verified E2E: Hermes' own login path now returns 400 (past the 429 wall)
instead of 429, using the real _OAUTH_TOKEN_USER_AGENT constant against the live
platform.claude.com token endpoint.
Salvaged from #57922 (authorize-host + scope changes dropped as non-load-bearing;
they only add a redirect hop back to claude.ai and the UA fix alone clears 429).
repair_message_sequence Pass 1 registered only tc.get("id") when building
the set of known assistant tool_call ids, then matched tool results against
it by tool_call_id. In the Codex Responses format an assistant tool_call
carries both id (fc_...) and a distinct call_id (call_...); a tool result's
tool_call_id may be keyed on either depending on which builder produced it.
Registering only id made a valid tool result whose tool_call_id matched
call_id look orphaned, so the pass dropped it and left the assistant
tool_call unanswered -- producing HTTP 400 on strict providers (DeepSeek,
Kimi): 'Messages with role tool must be a response to a preceding message
with tool_calls'. Long-running sessions that persisted such a sequence were
permanently broken, re-sending the orphan every turn.
Register both id and call_id for each assistant tool_call so a result
matching either key is recognized, consistent with
AIAgent._get_tool_call_id_static and the compressor's _sanitize_tool_pairs.
Apply the same call_id||id precedence to the corrupted-args sanitizer's
existing-result scan / stub insertion, which had the identical mismatch.
Adds 3 regression tests covering the codex id!=call_id case (match on
call_id, match on only call_id, match on id when both present).
PR #57601's original branch added a top-level reasoning_effort emit to the
LEGACY build_kwargs path (agent/transports/chat_completions.py), but
provider=custom resolves to CustomProfile (plugins/model-providers/custom/),
so chat_completion_helpers takes the profile path and returns early — the
added branch was unreachable dead code for every custom endpoint.
Move the fix to its real site, CustomProfile.build_api_kwargs_extras(), and
follow the DeepSeek/Zai profile precedent:
- disabled -> extra_body.think = False (unchanged)
- enabled + effort -> TOP-LEVEL reasoning_effort (the OpenAI-compatible
format GLM-5.2/ARK expect), passed through verbatim
incl. max/xhigh
- enabled + no effort -> omit, so the endpoint's server default applies
(avoids silently forcing 'medium' as the original
branch did)
Deliberately does NOT force think=True on enable — that flag is Ollama-only
and risks a 400 on GLM/vLLM endpoints that don't recognize it; thinking is
already server-default-on for these backends.
Verified end-to-end through the real profile dispatch (temp HERMES_HOME):
custom+high -> reasoning_effort=high; custom+max -> reasoning_effort=max;
custom+none -> think=False; custom+unset -> nothing; num_ctx composes.
Adds tests/plugins/model_providers/test_custom_profile.py (13 cases).
Addresses the custom-provider half of #55276.
Co-authored-by: huanshan5195 <huanshan5195@users.noreply.github.com>
- Add 'max' to VALID_REASONING_EFFORTS (GLM-5.2 native parameter)
- Emit top-level reasoning_effort string for custom providers
- Stop hardcoding 'medium' in legacy extra_body.reasoning, use actual effort
Custom providers (e.g. GLM-5.2 on Volcengine ARK) silently dropped
reasoning_effort — the value never reached the upstream API. Kimi,
TokenHub, and LM Studio all had dedicated branches for this, but
custom providers had none.
Self-review (hermes-pr-review Phase 2) flagged the mid-turn pre-API
compaction for reuse/duplication (W1/W2); fixing that surfaced a
regression against a deliberate existing feature, now also fixed.
The block now mirrors the turn-prologue preflight's guard chain exactly
(agent/turn_context.py) instead of a hand-rolled pressure limit:
1. should_defer_preflight_to_real_usage(rough) — defer when the rough
estimate is known-noisy vs a recent real provider prompt that fit
under threshold (schema overhead / post-compaction over-count, #36718).
2. get_active_compression_failure_cooldown() — skip during a same-session
compression-failure cooldown.
3. should_compress(rough) — reuses the canonical threshold_tokens (output
room already reserved by _compute_threshold_tokens) plus its summary-LLM
cooldown + anti-thrash guards (#11529).
Dropped the seven inline _reserve/_output_pressure locals (W2: they
re-derived _compute_threshold_tokens and omitted its 85% degenerate-window
fallback). compression_attempts stays as the hard per-turn backstop.
Without guard (1) the block fired a compaction the preflight deliberately
defers, breaking test_413_compression::test_preflight_defers_when_recent_
real_usage_fit (ValueError from the mocked _compress_context). Verified:
test_413_compression 26/26, codex 82/82, and anti-thrash engaged
(_ineffective_compression_count=2) still suppresses the block (0 calls).
The salvaged max-output/pressure fix set conversation_history=None after
the new pre-API compaction. That is only correct for legacy session-
rotation. Under the default in-place compaction (compression.in_place:
True), archive_and_compact inserts the compacted rows into the session DB
directly without stamping them with the intrinsic persisted-marker, so a
subsequent flush with conversation_history=None re-appends them — doubling
the active context and retriggering compression (the early-persist
duplicate-row trap).
Use conversation_history_after_compression(agent, messages), matching the
two existing compaction sites (post-response should_compress and the
turn-prologue preflight), which returns None for rotation and
list(messages) for in-place so the compacted dicts are skipped by identity.
Adds a regression test with a real SessionDB + real archive_and_compact
that asserts the compacted summary row is persisted exactly once (fails
with 2 copies on the None variant).
Follow-up to the salvaged #57845 fix. _can_carry_marker used
any(isinstance(part, dict)) but _apply_cache_marker only marks the LAST
content part, so a list whose last element is a non-dict passed the carrier
gate yet received no marker — wasting one of the four breakpoints. Tighten
the predicate to require content[-1] to be a dict (mirroring the apply
logic) and add a regression test. Flagged by a 3-agent review.
Follow-up to the per-provider guards. Three improvements from review:
1. Extract agent.file_safety.raise_if_read_blocked() as a single shared
chokepoint and route the OpenAI, OpenRouter, and (newly) xAI image
providers through it, replacing the 3x-duplicated inline try/except.
Fixes the whole bug class: xai/_xai_image_field read a model-supplied
local path via open() with no guard — the same vulnerability the PR
fixed for OpenAI/OpenRouter, in a sibling provider it missed.
2. Strengthen the regression tests from pass-on-any-ValueError to true
security invariants: spy open()/read_bytes() and assert the blocked
credential is NEVER read; add negative controls (legit local image
still loads; remote/data: URIs pass through unguarded) so a
block-everything regression can't pass.
3. Guard is best-effort by design (defense-in-depth, not a security
boundary) — documented on the shared helper.
- agent/file_safety.py: raise_if_read_blocked()
- plugins/image_gen/{openai,openrouter,xai}: route through helper
- tests: no-read spies + negative controls across all three providers
Two caching holes made MoA re-bill essentially its entire input stream:
1. AGGREGATOR: anthropic_prompt_cache_policy() judged the agent's own
model/provider — on the MoA path those are the virtual preset name and
'moa', which match no caching branch, so _use_prompt_caching was False
and the acting aggregator (Claude on OpenRouter) ran with ZERO
cache_control breakpoints. Measured on identical opus-4.8 sessions:
85% cache share solo vs 2% via MoA — ~30M re-billed input tokens on one
132-task benchmark run. Fix: when provider == 'moa', resolve the policy
from the preset's real aggregator slot (provider/model/base_url/api_mode
via resolve_runtime_provider).
2. ADVISORS: _run_reference never applied cache_control at all, and
Anthropic caching is opt-in per request — Claude advisors served 0
cache reads across 1,227 benchmark calls (11.5M re-billed input tokens)
even though the advisory view is append-only across iterations (stable
prefix; the synthetic end marker is last so it never pollutes it). Fix:
_maybe_apply_advisor_cache_control() reuses the SAME policy function and
SAME system_and_3 layout as the main loop, judged on the advisor slot's
own resolved runtime — advisor requests are now decorated exactly like
an acting agent on that provider. Auto-caching routes (OpenAI-family)
are left untouched by policy.
Live-verified on the wire (per-iteration opus+gpt5.5 preset, 4 fan-outs):
claude advisor fan-out 2-3 cache_write=2161/2344, fan-out 4
cache_read=2206 / fresh_in=2; aggregator session cache share 84%/77%
(vs 2%/0% before). Sub-1024-token prompts correctly stay uncached
(Anthropic minimum).
A custom:<name> main provider resolves at runtime to the bare provider id
"custom". In the vision auto-detect chain, the main-provider branch called
resolve_provider_client("custom", ...) WITHOUT explicit_base_url/api_key,
so it returned (None, None) ("no endpoint credentials found") and the whole
chain fell through to OpenRouter/Nous. A user on a custom endpoint with no
aggregator configured then got "No LLM provider configured for task=vision
provider=auto" on every image, even though their main model fully supports
vision.
Recover the live endpoint that set_runtime_main() records each turn
(_RUNTIME_MAIN_BASE_URL/_API_KEY/_API_MODE) and forward it to Step 1, with
a fallback to _resolve_custom_runtime() for non-gateway callers. Mirrors the
existing explicit-base_url branch directly above.
Adds TestResolveVisionCustomProvider covering custom, custom:<name>, and the
no-runtime fallback path.
When model.provider is set to custom:<name>, _supports_vision_override()
previously tried only the runtime provider key ('custom') and the raw
config value ('custom:my-proxy'). It did not try the stripped name
('my-proxy'), which is the actual key under providers: in config.yaml.
This caused native image routing to fall back to text mode even when the
user explicitly declared supports_vision: true on the named provider's
model entry.
Fixes#39963
The advisory view appends a synthetic user marker when it ends on an
assistant turn (Anthropic end-on-user rule) — i.e. on every tool iteration
after the first. The user_turn prefix hash treated that marker as the last
user message, so the hashed prefix included the grown mid-turn context and
the signature changed every iteration: advisors re-ran per iteration,
silently defeating the once-per-turn cadence (live smoke test: 2 fan-outs
for a 2-iteration task; expected 1). Hoist the marker to a module constant
and skip it when locating the last REAL user message. Verified: iteration-2
signature now equals iteration-1 (cache HIT); a new real user message still
re-triggers the fan-out.
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.
Follow-up to the per-site strips from the review gate. The two copy-site
strips are correct but positional — a copy site added after the assembly
loops would re-leak _db_persisted into the child-session flush. Add a single
terminal sweep (_strip_persistence_markers) run once on the fully-assembled
compressed list so the invariant 'no compacted message leaves compress()
carrying a persistence marker' is structural, not dependent on copy-site order.
- agent/context_compressor.py: _strip_persistence_markers() called before
compress() returns; helper docstring notes the sweep is the authoritative guard
- tests/agent/test_context_compressor.py: structural regression — neuter the
per-site helper to a leaking copy, assert the terminal sweep still strips
- tests/run_agent/test_compression_persistence.py: pin the fixture assumption
behind the exact-equality row-count assertion
Shallow messages[i].copy() during context compression propagated the
_db_persisted marker from cached gateway incremental flushes into the
post-rotation compressed list. _flush_messages_to_session_db then skipped
every row when writing to the new child session, so gateway restarts
lost the compacted transcript (severe amnesia).
Strip the marker in _fresh_compaction_message_copy() and add regression
tests for rotation flush + compressor assembly.
Fixes#57491
Self-review (ruff+ty lint diff = 0 net-new; 2-agent deep review) surfaced one
Warning + comment-accuracy nits; no Critical:
- W1: the local-probe TTL cache memoized None (probe failure) for 30s, so a
probe that failed during a startup race would suppress a legit retry once
the server came up. Cache only positive results — still fully bounds the
hot-path probe rate (reachable servers cache their value) while an
unreachable one re-probes on the next call. Add a regression test asserting
a None result is NOT cached (retry re-probes); mutation-verified.
- Tighten the platform-guard comment: gateway/TUI/cron already construct with
quiet_mode=True (gated by `not agent.quiet_mode`), so the guard's active job
is CLI dedup vs show_banner, not "filling the gateway/TUI gap" as originally
worded.
Verified not-issues (per review): positive-value 30s cache does not break the
reconcile-after-restart freshness contract (restart = fresh process, empty
cache); cache key is collision-safe; platform guard is correct in both
directions (no runtime path leaves platform None on a non-CLI surface).
Tests: 149 passed. ruff clean; ty 0 net-new vs base.
Salvage review of #56431 surfaced one Critical + two Warning issues; fix
them on top of the contributor's cherry-picked commits:
1. Critical — duplicate non-agentic warning on the interactive CLI. The new
agent_init warning fires on every platform, but cli.py show_banner()
already warns on CLI (richer output + /model hint), so a CLI user saw the
warning twice per startup. Guard the agent_init emit to skip platform=="cli"
— it now fills exactly the gateway/TUI gap the PR intended, no duplication.
2. Warning — vLLM error-parse regex under-matched. The patterns required a
literal space before the number, so "max_model_len: 32768", "=32768",
"(32768)", and "... is 32768" all returned None. Broaden both patterns to
accept :/=/(/ 'is' delimiters. Add a parametrized test over all delimiter
variants.
3. Warning — per-call live probe latency on local endpoints. The new
reconcile-on-hit + pre-defaults step-7 probe made every local resolution
fire a synchronous network probe (banner + /model switch + compressor
update_model each within one startup). Add a 30s in-process TTL cache
keyed by (model, base_url) around _query_local_context_length so back-to-
back resolutions reuse one round-trip; not persisted to disk, so the
reconcile freshness contract (re-probe after restart) is preserved. Add an
autouse fixture clearing the cache between tests + TTL coverage.
Tests: 148 passed (was 138). ruff clean.
Reconcile stale local disk cache against live vLLM/Ollama max_model_len
probes, probe local servers before the llama hardcoded default, parse
vLLM max_model_len overflow errors, and surface the non-agentic Hermes 3/4
warning at agent init on gateway/TUI.
Sub-64K live probes are returned for startup rejection but are not
persisted to the context cache — preserving the 64K minimum-context
contract instead of normalizing undersized windows as valid config.
(cherry picked from commit c3a02db4fd)
normalize_usage only read output_tokens_details.reasoning_tokens (the
Responses API shape). Chat Completions providers — OpenAI, OpenRouter,
DeepSeek, and every OpenAI-compatible proxy — report it under
completion_tokens_details.reasoning_tokens, so reasoning_tokens was 0 for
every chat_completions reasoning model: hidden thinking was invisible in
session accounting, MoA traces, and the eval's per-task token columns.
Measured impact (HermesBench MoA run on deepseek-v4-flash, 4,828 advisor
calls): reasoning_tokens showed 0 everywhere while individual calls burned
up to 21.5K hidden thinking tokens to emit ~500 visible tokens. Verified
live against OpenRouter: deepseek-v4-flash returns
completion_tokens_details.reasoning_tokens=61 for a 74-completion-token
call; the field was simply never read.
Responses-shape reads are unchanged; the new read only fires when the
Responses shape yielded nothing.
The terminal-refresh quarantine filtered in-memory entries on
source == "device_code" but built removed_ids from the deleted
"loopback_pkce" source name, so the revoked device-code entry was
never pruned from the persisted pool in auth.json. Also restores the
_print_loopback_ssh_hint test suite scoped to Spotify (the helper's
remaining caller) instead of deleting it wholesale.
Replace the loopback/PKCE-callback server and manual-paste fallback with
the RFC 8628 device-code flow as the only xAI Grok OAuth login path. The
flow works in headless/SSH/container sessions with no 127.0.0.1 listener,
shrinking the local attack surface.
- Poll the token endpoint with server-provided interval, honoring
slow_down and expires_in; store tokens with auth_mode
oauth_device_code.
- Adaptive proactive refresh skew for short-lived device-code JWTs;
rotated tokens sync back to auth.json, the global root store, and the
credential pool (no refresh-token replay).
- Clear source suppression on successful re-login (CLI + dashboard) and
drop the duplicate dashboard pool entry so exactly one seeded
device_code entry exists.
- Use the shared device_code source name for consistency with the
nous/codex device-code providers.
- Desktop: remove the loopback OAuth flow states and dead type variants;
pkce providers' sign-in URL selection is unchanged.
- Docs (EN + zh-Hans) rewritten for device-code login; drop the deleted
--manual-paste flag from documented commands.
Named providers / custom_providers entries in config.yaml now accept an
extra_headers dict scoped to that endpoint — for reverse proxies, API
gateways, and custom auth schemes (e.g. Cloudflare Access service tokens).
- hermes_cli/config.py: normalize extra_headers on provider entries
(_normalize_custom_provider_entry + providers-dict translation), add
get_custom_provider_extra_headers /
apply_custom_provider_extra_headers_to_client_kwargs helpers keyed on
base_url (case/trailing-slash insensitive, no substring bypass —
mirrors the TLS helpers)
- hermes_cli/runtime_provider.py: surface extra_headers in the resolved
runtime for named custom providers (providers dict, legacy
custom_providers list, and the credential-pool path)
- run_agent.py / agent/agent_init.py: merge per-provider extra_headers
onto the OpenAI client default_headers at construction and on every
_apply_client_headers_for_base_url re-application (credential swaps,
rebuilds), most-specific level wins; OpenAI-wire only (native
Anthropic/Bedrock scoped out)
- agent/auxiliary_client.py: accept model.extra_headers as an alias of
model.default_headers for the global variant
- cli-config.yaml.example: documented commented example
- Header values are treated as secrets and never logged
Salvaged from PR #3526 by @jneeee, reimplemented against current main.
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
The helper docstring described the typical ~15-25k gateway payload but
read as if that were the trigger range; the floor actually engages above
10k tokens. Clarify the prose to match the gate.
Lower the openai-codex stale-timeout floor from 25k to 10k estimated
tokens so Telegram/gateway sessions (~20k tools+instructions) are not
aborted at the generic 90s cutoff while Codex is still prefilling.
Three CLI reliability fixes:
1. Interrupt reliability: chat() only re-queued the user's interrupt
message when the turn result carried interrupted=True. When the agent
thread raced past its last interrupt check (or finished) before the
interrupt landed, the message was silently dropped — and the stale
_interrupt_requested flag left on the agent instantly aborted the
NEXT turn. Un-acknowledged interrupt messages are now re-queued as
the next turn and the stale flag is cleared (only when the agent
thread actually exited). The clarify-race path also parks the message
in _pending_input instead of dropping it.
2. Slow exit (5+ min): stdlib ThreadPoolExecutor workers are non-daemon
and joined unconditionally by concurrent.futures' atexit hook — even
after shutdown(wait=False). One wedged tool worker (abandoned after
interrupt/timeout) held the process open forever. Promoted
async_delegation's daemon executor to a shared tools/daemon_pool
module and adopted it in tool_executor (concurrent tool batches),
memory_manager (background sync), delegate_tool (child timeout wrapper
+ batch fan-out), and skills_hub (source fan-out). Added a 30s exit
watchdog (HERMES_EXIT_WATCHDOG_S) armed at _run_cleanup start as a
backstop for wedged cleanup steps.
3. Exit jank: after prompt_toolkit tears down the input/status bars the
terminal sat silent for the whole cleanup window, looking hung. Print
'Shutting down… (finalizing session)' immediately at exit start.
E2E: live PTY interrupt of a foreground 'sleep 120' terminal tool now
aborts in ~1s and the typed message runs as the next turn; wedged-worker
+ wedged-cleanup subprocess exits in 5.8s (watchdog) instead of hanging.
Follow-up on the salvaged #56392 guard. The cherry-picked change matched
custom:<name> pool entries against the primary by raw base_url string
equality, which (a) can't disambiguate two named custom providers sharing
one gateway base_url and (b) left a latent bare-"custom" entry bypass.
Route the match through get_custom_provider_pool_key(rt[base_url]) compared
against the entry's custom:<name> key, mirroring the sibling guard in
recover_with_credential_pool. Use CUSTOM_POOL_PREFIX instead of the literal.
Add regression tests for the custom same-endpoint (swap) and cross-endpoint
(skip) branches, plus the plain-provider fallback-pool case from #56885.
Two related hardening fixes for auxiliary calls (which include MoA reference
advisors — a pinned-model path where provider fallback is not a meaningful
recovery):
1. Transient-transport retries: the same-provider retry on a connection reset /
timeout / 5xx / 408 was a single attempt, then fallback. For a pinned aux
call a second blip silently loses the call (root of the run2 double-advisor
'Connection error' collapse — a genuine upstream blip). Now retries N times
with exponential backoff, N = auxiliary.transient_retries (default 2 -> 3
total attempts, clamped [0,6]). Compression-on-timeout fast-fail carve-out
preserved.
2. Per-model client-cache isolation: _client_cache_key excluded the model, so
two concurrent auxiliary calls to the same provider/base_url/key but
different models (e.g. an opus + gpt-5.5 MoA fan-out) shared one cache entry
and could race each other's client lifecycle. Model now participates in the
key -> distinct clients, no cross-call races. Same-model reuse unchanged.
- agent/auxiliary_client.py: _transient_retry_count() + backoff loop; model in
_client_cache_key and both call sites.
- hermes_cli/config.py: auxiliary.transient_retries default (2).
- tests: new retry/isolation tests; updated 2 stale-expectation tests to the
corrected behavior (per-model resolve; N-retry escalation).
Backoff base is overridable (_TRANSIENT_RETRY_BACKOFF_BASE) so tests don't sleep.
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
