Ollama sizes a model's real context window by free VRAM at load time,
often far below the GGUF trained max that /api/show reports, and its
OpenAI-compatible endpoint has no options passthrough — per-request
num_ctx and keep_alive are silently dropped, so the window cannot be
controlled from the client. The compressor was being sized to the
trained max (e.g. 262K for a model actually running at 32K).
- Add query_ollama_loaded_context() reading the effective window from
/api/ps (60s cache, never persisted — transient load state).
- Reconcile after each successful response via
sync_ollama_loaded_context(): resize the compressor to the loaded
window and warn when it is below the tool-use minimum. Selection
surfaces keep showing the trained max; explicit model.context_length
still wins. No-op for non-Ollama providers.
- Refresh model.ollama_keep_alive through the native API (rate-limited
/api/generate ping) since /v1 drops it.
- Remove the inert num_ctx/keep_alive request-body plumbing; warn that
model.ollama_num_ctx has no effect and point at OLLAMA_CONTEXT_LENGTH
/ Modelfile num_ctx. Keep detection for the pre-flight window check.
- Disable thinking on Ollama thinking models via reasoning_effort
'none' — the only switch its /v1 handler parses (think is dropped).
Bundle Fireworks AI as a first-class BYOK provider across the CLI, web/TUI,
and desktop onboarding.
- New model-provider plugin with attribution headers (HTTP-Referer / X-Title)
so Fireworks can attribute Hermes traffic; PAYG-safe default aux + fallback
models (accounts/fireworks/models/...), IDs tracking fw-ai/fireconnect.
- Registered in CANONICAL_PROVIDERS so it appears in the CLI/web/TUI pickers.
- Alias wiring (fireworks-ai, fw) into both CLI resolvers.
- First-class wiring: OPTIONAL_ENV_VARS, HERMES_OVERLAYS (FIREWORKS_BASE_URL
override), doctor env hints. Live catalog + model_metadata are auto-derived.
- doctor: treat Fireworks' native slash-form IDs (accounts/fireworks/...) as
valid, not aggregator vendor prefixes, so it no longer tells Fireworks users
to switch to openrouter or drop the prefix.
- picker: plugin providers with no static curated list now lead with their
profile fallback_models, so the default is an agentic chat model instead of
whatever the live catalog returns first (Fireworks listed an image model,
flux-*, ahead of its chat models).
- Desktop onboarding: Fireworks as a RECOMMENDED hero card with the official
Fireworks logomark and a brand-purple badge, routing to the BYOK key form;
i18n in en/ja/zh/zh-hant.
- Tests: profile contract, first-class wiring (both resolvers, overlay, config,
doctor incl. the slash-form regression, aux headers, credentials), discovery
spot-check, and a live smoke test driven through the Hermes runtime.
Fire Pass (fpk_) support is coming soon; the future wiring is kept as a
commented-out scaffold in the plugin.
Promote bare "ollama" from a custom-endpoint alias to a real provider and
build the desktop UX around it. A local server's connection kind is
reachability rather than a credential, so every credential-shaped gate
(provider registry, picker filters, settings surfaces) gets an explicit
path for it.
Backend:
- Provider overlay + registry entry (127.0.0.1:11434/v1 default, keyless
with a local-only placeholder, base-url normalization for /api and /v1
forms). Existing provider=custom configs are untouched.
- GET /api/local-servers/detect fingerprints well-known local ports plus a
local configured base_url; response shape leaves room for a future
installed/running/managed distinction.
- /api/ollama/* management endpoints: installed+running+recommended models,
registry pull as a poll-able background job streaming native NDJSON
progress, delete, and load (warm-up / keep_alive pinning). Pull and
delete bust the picker's model-id cache.
- Model picker payload: per-model capabilities widened to tools/vision/
context_length; local Ollama rows enriched from the server's native
/api/show (authoritative for on-disk tags, where models.dev is sparse),
backfilled off the request path by a background thread.
- The explicit-only picker filter keeps ollama rows: the row only exists
when the server answered a probe, which is as explicit as a pasted key.
- Reasoning safety: /api/show thinking capability gates all reasoning
fields (Ollama 400s reasoning_effort on non-thinking models), and
OpenAI-only effort levels map to the nearest accepted level
(xhigh->max, minimal->low).
- model.ollama_keep_alive config: sent per-request as extra_body.keep_alive.
- Latency discipline for a local server that may be down: a 300ms TCP
pre-check with a short negative cache guards every native-API read; the
status endpoint probes fresh so a just-started server is noticed
immediately; localhost is rewritten to 127.0.0.1 (Windows resolves
localhost to ::1 first and Ollama binds IPv4 loopback — each request
otherwise pays a ~2s failed IPv6 connect, including chat inference).
Desktop:
- Providers -> Accounts: a "Local servers" card mirroring the OAuth card
language ("Running · N models" / a start-the-server hint), expanding to
model management: installed models with size/quant/VRAM, delete, warm-up,
curated pull recommendations with a progress bar, free-form pull, and a
KV-cache advisory when a loaded model runs well under its trained window.
Polls while down so it flips to Running by itself.
- Model picker: "No tools" badge (explicit tools:false only — absence means
unknown) with demotion, plus context window / parameter size / quant per
row; refetches once after open so backfilled metadata appears in place.
- Onboarding: detected-server row with a model select, replacing blind
first-model assignment for detected servers.
- Composer status stack: "Loading <model> into memory" row during cold
starts, confirmed against /api/ps so ordinary slow generations stay quiet.
Chat inference stays on the OpenAI-compatible /v1 endpoint; the native
/api surface is used read-only for metadata plus explicit management
actions. Lifecycle management (starting or installing Ollama) is not
included.
Port from Kilo-Org/kilocode#11555: GLM-5.2 exposes a native
reasoning_effort knob with two enabled levels (high / max) on its
OpenAI-compatible endpoints. Previously the zai profile (direct Z.AI
/api/paas/v4) used the base ProviderProfile and emitted nothing, and the
OpenCode Go profile only handled Kimi K2 / DeepSeek — so a user's effort
preference for GLM-5.2 was silently dropped on both routes.
- zai: ZaiProfile maps effort onto high/max (xhigh/max -> max, lower -> high)
- opencode-go: same mapping for GLM-5.2, alongside existing Kimi/DeepSeek
- alias spellings recognized (glm-5.2 / glm-5-2 / glm-5p2, vendor-prefixed)
- disabled / no effort leaves the server default untouched
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>
A Z.ai desktop user reported thinking reverting to medium after one turn,
burning ~200% of a week's credits in 4 days despite reasoning_effort: false
in config.yaml. Four compounding bugs:
- _session_info reported reasoning_effort "" for disabled reasoning,
indistinguishable from unset — the desktop adopted it after the first
turn, wiping its sticky "thinking off" pick so every later chat
reverted to the default effort.
- config.set key=reasoning always wrote agent.reasoning_effort to global
config.yaml, so every desktop model-menu selection (preset.effort ??
'medium') clobbered the user's configured value. Now session-scoped
like the messaging gateway's /reasoning, landing on
create_reasoning_override so lazily-built sessions keep it too.
- YAML `reasoning_effort: false`/`off`/`no` (boolean False) was coerced
to "" by every loader's `str(x or "")`, silently re-enabling thinking.
parse_reasoning_effort now treats False/"false"/"disabled" as
{"enabled": False}; loaders (tui gateway, gateway, cli, cron,
delegate) pass the raw value through. The desktop config reader also
crashed on the boolean (false.trim()), aborting voice/STT settings.
- The zai provider profile never sent thinking on the wire, and GLM-4.5+
defaults to thinking ON server-side — so disabling reasoning was a
silent no-op on direct Z.ai, the actual token burner. The profile now
emits extra_body.thinking {"type": "enabled"|"disabled"} for
thinking-capable GLM models, mirroring the DeepSeek profile.
Also: /new (session reset) now carries reasoning_config across the
rebuild like model_override; config.get reasoning prefers the session's
live value and maps a config False to "none"; Settings shows "Off"
instead of a blank select for hand-written false.
Map Hermes xhigh→max to unlock DeepSeek V4's 'Max thinking' tier
through Ollama Cloud's OpenAI-compatible /v1/chat/completions endpoint.
low/medium/high pass through unchanged; disabled/none suppress
reasoning entirely.
Empirically confirmed: reasoning_effort:max produces ~2.5× more
thinking tokens than high on deepseek-v4-pro:cloud (1576 vs 642).
Follow up PR #46609's api.minimax.io reasoning report by moving the behavior out of the broad run_agent host gate and into the MiniMax provider profile. Only MiniMax-M3 on the documented OpenAI-compatible /v1 route gets reasoning_split/thinking/reasoning_effort; Anthropic-format MiniMax and non-M3 models keep their existing wire shapes.
Co-authored-by: goku94123 <gooku94123@gmail.com>
The standalone Kimi/Moonshot profile (api.moonshot.ai/v1) sent both
extra_body.thinking AND a top-level reasoning_effort. With no reasoning
config it even defaulted to thinking:enabled + reasoning_effort:medium,
pairing them on every default call. Moonshot treats these as mutually
exclusive (cannot specify both 'thinking' and 'reasoning_effort').
Align with the kimi-k2 handling already shipped for the opencode-go relay:
send effort when a recognized low|medium|high is requested, otherwise fall
back to the extra_body.thinking toggle. Disabled sends thinking:disabled
only. Never both.
Reported by Cars29 (NOUS Discord). DeepSeek was deliberately left untouched:
its native endpoint accepts both (verified by the live guardrail in
test_deepseek_v4_thinking_live.py), so the report's DeepSeek claim does not
hold there.
Tests: tests/plugins/model_providers/test_kimi_profile.py pins the xor
contract across all config shapes.
Three Copilot inline review comments on #37664, two worth landing
in a polish pass before merge:
1. auxiliary_client.py:270 — Copilot suggested keeping the
minimax-* entries in _API_KEY_PROVIDER_AUX_MODELS_FALLBACK as
a safety net for environments where the profile-based
resolution can't import or run plugin discovery. **Declined.**
The deepseek precedent (commit 773a0faca) explicitly removed
deepseek from the same dict for the same reason — the profile
layer is the source of truth and the dict is a legacy
pre-profiles-system fallback. We do not want to fragment the
codebase by provider: either the profile layer is authoritative
or the dict is. The minimax PR picks profile (matching deepseek)
and the dict stays cleaned up. The risk Copilot raises is
real but theoretical — plugin discovery runs at import time of
the providers module, which is the first thing any modern
Hermes entrypoint imports.
2. tests/agent/test_minimax_provider.py:162 — Copilot flagged
that the test class relies on _get_aux_model_for_provider()
resolving via provider profiles but doesn't explicitly trigger
plugin discovery. **Fixed.** Added 'import model_tools # noqa:
F401' at the top of both test_minimax_aux_is_standard and
test_minimax_aux_not_highspeed. The fixtures in the parallel
test_minimax_profile.py already did this; the legacy test in
test_minimax_provider.py was order-dependent and would silently
break if anyone reorganised the test ordering. Pinned the
dependency explicitly so the test is order-independent.
3. tests/plugins/model_providers/test_minimax_profile.py:46 —
Copilot flagged that the docstring referenced a hard-coded
line number 'hermes_cli/models.py:298' that would go stale.
**Fixed.** Replaced with the symbol reference
'hermes_cli.models._PROVIDER_MODELS[\'minimax\']' which is
stable under file edits and grep-friendly. The new docstring
also reads more naturally — readers don't have to look up
'what's at line 298' to follow the reasoning.
All 221 minimax-related tests still pass.
The minimax / minimax-cn / minimax-oauth profiles still advertised
M2.7 (and M2.7-highspeed for OAuth) as their default_aux_model,
predating the M3 release (2026-06-01). The user-facing
_PROVIDER_MODELS['minimax'] catalog top entry is M3, and the
recommended config for a Token-Plan install now sets
model.default: MiniMax-M3, so the aux default was the only
remaining drift.
Updates:
* minimax default_aux_model: M2.7 -> M3
* minimax-cn default_aux_model: M2.7 -> M3
* minimax-oauth default_aux_model: M2.7-highspeed -> M2.7
(M3 is not on the OAuth / Coding Plan tier per
platform docs as of this PR; the highspeed
variant was the 2x-cost regression from #4082
that PR #6082 collapsed to plain M2.7 for
minimax / minimax-cn but missed OAuth)
* agent/auxiliary_client.py: drop the three legacy
_API_KEY_PROVIDER_AUX_MODELS_FALLBACK entries for the minimax
family. _get_aux_model_for_provider() reads from
ProviderProfile.default_aux_model first (line 250) and only
falls back to the dict when the profile has no aux model or
the profile import fails. With the profile now set, the dict
entries are dead code and a drift hazard. Mirrors the deepseek
cleanup in 773a0faca.
* tests/agent/test_minimax_provider.py: update the existing
TestMinimaxAuxModel assertions from MiniMax-M2.7 to MiniMax-M3
(the intent — 'standard, not highspeed' — is unchanged; the
pin value is).
* tests/plugins/model_providers/test_minimax_profile.py: new
file mirroring tests/plugins/model_providers/test_deepseek_profile.py.
Pins each of the three profiles' default_aux_model and
asserts _get_aux_model_for_provider() returns it. A second
class guards against the highspeed regression coming back.
Refs:
- Closes#36196 in spirit (M3 support — the catalog half of
that issue is #36212; this PR covers the profile half)
- Related: #4082 (M2.7-highspeed 2x-cost), #6082 (previous
M2.7-highspeed -> M2.7 fix that missed OAuth + the
auxiliary_client.py fallback dict)
- Pattern: 773a0faca (same profile-layer fix for deepseek)
The Kimi K2 branch added in the prior commit only emitted extra_body.thinking
and dropped reasoning_effort entirely. KimiProfile (api.moonshot.ai/v1) sends
both fields, and OpenCode Go proxies to the same Moonshot backend. Mirror that
shape on the Go path so /reasoning effort actually reaches Kimi.
- low/medium/high pass through verbatim
- xhigh/max clamp to high (Moonshot's max supported value)
- minimal / unknown effort → omit reasoning_effort, keep thinking on
- disabled / no config → unchanged
- DeepSeek branch unchanged
Closes#26924 (and supersedes #26926) in spirit.
DeepSeek was missing `default_aux_model` on its `ProviderProfile`, so
`_get_aux_model_for_provider("deepseek")` returned an empty string and
the compression / vision / session-search paths emitted
"No auxiliary LLM provider configured -- context compression will
drop middle turns without a summary."
on every DeepSeek session, even when the user had perfectly working
DeepSeek credentials.
Fix lands at the profile layer rather than the legacy
`_API_KEY_PROVIDER_AUX_MODELS_FALLBACK` dict the original PR targeted.
Every modern provider (gemini, zai, minimax, anthropic, kimi-coding,
stepfun, ollama-cloud, gmi, novita, kilocode, ai-gateway, opencode-zen)
sets `default_aux_model` on its `ProviderProfile`; the fallback dict
only exists for providers that predate the profiles system.
Tests added under `tests/plugins/model_providers/test_deepseek_profile.py`:
- `test_profile_advertises_deepseek_chat` -- pins the profile attribute
- `test_consumer_api_returns_deepseek_chat` -- pins the consumer API behavior
- `test_consumer_api_returns_non_empty` -- regression guard for the
symptom in the issue
Original diagnosis and aux-model choice from @kriscolab in PR #26926;
moved one layer up.
Co-authored-by: kriscolab <71590782+kriscolab@users.noreply.github.com>
The cherry-picked PR #15251 from @tw2818 correctly identified the
DeepSeek 400 root cause but placed the fix in the legacy fallback path
of `build_kwargs`, which DeepSeek never reaches — DeepSeek has a
registered ProviderProfile and goes through `_build_kwargs_from_profile`
instead. The legacy-path block was therefore dead code.
This commit pivots the fix to where it actually fires:
- New `DeepSeekProfile` in `plugins/model-providers/deepseek/__init__.py`
overrides `build_api_kwargs_extras` to emit DeepSeek's expected wire
format (mirrors `KimiProfile`):
{"reasoning_effort": "<low|medium|high|max>",
"extra_body": {"thinking": {"type": "enabled" | "disabled"}}}
- Model gating: only `deepseek-v4-*` and `deepseek-reasoner` emit
thinking control. `deepseek-chat` (V3) is untouched — current behavior.
- Effort mapping: low/medium/high passthrough, xhigh/max → max, unset →
omitted (DeepSeek server applies its own default).
- Revert the legacy-path additions from PR #15251 — they were dead code,
and the `_copy_reasoning_content_for_api` strip block specifically
would have nullified the existing reasoning_content padding machinery
(`_needs_deepseek_tool_reasoning` → space-pad on replay) that the
active provider already relies on for replay correctness.
- Unit tests pin the wire-shape contract and the model gating rules
(26 tests, all passing). Existing transport + provider profile suites
(321 tests) continue to pass.
- AUTHOR_MAP: map twebefy@gmail.com → tw2818 for release notes credit.
Closes#15700, #17212, #17825.
Co-authored-by: tw2818 <twebefy@gmail.com>
