Backend (api/config.py):
- resolve_model_provider(): check custom_providers for prefix match
BEFORE the config_base_url branch. Previously, providers with a
base_url set (e.g. deepseek) would catch all slash-delimited model
ids and return the config provider, preventing custom provider
routing.
- get_available_models(): include model aliases in response so the
frontend can resolve them on /model commands.
Frontend (static/commands.js):
- cmdModel(): resolve aliases by fetching /api/models before fuzzy
matching the dropdown.
- Add bare-model fallback when the alias resolves to a slash-delimited
provider/model id (e.g. "deepseek/deepseek-v4-flash").
- Add cross-provider fallback: when the model is from a custom provider
not in the active provider dropdown, call /api/session/update directly
with the provider/model id and provider override.
The get_available_models() function only handled dict-format models
(`{model_id: {}}`) for custom_providers entries, silently dropping
models specified as YAML lists (`[model1, model2]`) or list of dicts
(`[{id: ..., label: ...}]`).
This caused users who define their custom providers with list-format
model declarations to see zero or incomplete model entries in both
Settings → Preferences → Default Model dropdown and the chat
interface model picker.
The fix adds an `elif isinstance(_cp_models_dict, list)` branch with
support for three list sub-formats:
- Plain string list: `models: [m1, m2]`
- Dict list: `models: [{id: m1, label: ...}]`
- Mixed: `models: [m1, {id: m2}]`
Refs: hermes-agent issue where YAML list models were invisible
Replace the earlier frontend-reset approach with a backend side-channel
approach that preserves the queue (event, data) tuple shape.
Problem (Opus catch):
- Live SSE frames emitted by _sse() in api/streaming.py:2296 carried no
'id:' field. Only journal-replay frames (via _sse_with_id) emitted IDs.
- Frontend's _lastRunJournalSeq cursor stayed at 0 during live streaming.
- Mid-stream error → reconnect-to-replay arrived with after_seq=0.
- Server replayed every journaled event from seq 1.
- assistantText (closure-scoped) had accumulated all live tokens already
→ double-rendered output.
Fix:
- api/config.py: STREAM_LAST_EVENT_ID: dict = {} module-level dict.
- api/streaming.py put(): capture journal event_id, write to
STREAM_LAST_EVENT_ID[stream_id]. Keep queue tuple as (event, data).
- api/routes.py _handle_sse_stream: read STREAM_LAST_EVENT_ID[stream_id]
at emit time, use _sse_with_id when set.
- api/streaming.py finally block: pop STREAM_LAST_EVENT_ID for cleanup.
Why side-channel instead of 3-tuple:
- Earlier attempt (queue tuple → (event, data, event_id)) broke 4 existing
tests: test_cancel_interrupt, test_sprint42, test_sprint51,
test_issue1857_usage_overwrite. These all unpack 'event, data = q.get()'.
- Frontend-reset approach (reset assistantText before replay) broke 3
other tests: test_smooth_text_fade, test_streaming_markdown,
test_streaming_race_fix. _wireSSE must NOT reset accumulators because
legacy reconnect doesn't replay events; only journal-replay does.
Side-channel preserves both invariants:
- Queue contract stays (event, data) — legacy consumers unbroken.
- Frontend accumulators stay alive on _wireSSE — legacy reconnect unbroken.
- Live SSE emits 'id:' so the journal cursor advances correctly.
6 regression tests added in test_stage364_opus_live_sse_event_id.py.
1 existing test (test_run_journal_streaming_static.test_streaming_journals_sse_events_before_queue_delivery) updated to be tuple-shape-agnostic.
Test results:
- Full pytest: 5713 passed, 10 skipped, 1 xfailed, 2 xpassed, 0 failed
- Previously-failing 5 tests: ALL PASS
- 6 new regression tests: ALL PASS
fix(config): preserve nvidia/ prefix on NVIDIA NIM (closes#2177)
Self-built. nesquena APPROVED with extensive end-to-end trace including
cross-tool agent CLI verification and 12-shape behavioural harness.
Move the `_PORTAL_PROVIDERS` guard in `resolve_model_provider()` to run
BEFORE the `prefix == config_provider` strip branch. The guard was added
for NVIDIA (along with the Nous portal cases in #854 / #894) but was
placed after the strip, so it never fired when `config_provider == "nvidia"`
and the model id started with `nvidia/`.
For `model_id="nvidia/nemotron-3-super-120b-a12b"`,
`config_provider="nvidia"`:
- prefix = "nvidia", bare = "nemotron-3-super-120b-a12b"
- prefix == config_provider → True → strip branch returned bare name
- `_PORTAL_PROVIDERS` guard never reached
- bare "nemotron-3-super-120b-a12b" sent to NVIDIA NIM → HTTP 404
NIM requires the full namespaced path. The fix moves the portal guard
to run first, so all portal providers (Nous, OpenCode-Zen, OpenCode-Go,
NVIDIA NIM) always preserve the full `provider/model` id regardless of
whether the prefix happens to equal the provider name.
This also closes a latent symmetric bug for the Nous case if a
`nous/<model>` id ever existed in the catalog.
Test plan:
- New `tests/test_issue2177_nvidia_prefix_preservation.py` covers:
- nvidia/nemotron-... under nvidia (the reported case)
- cross-namespace qwen/ and meta/ under nvidia (regression pin)
- every static nvidia model in `_PROVIDER_MODELS` resolves to itself
- latent nous/<model> under nous (structural ordering pin)
- non-portal providers (anthropic) still strip — fix doesn't over-correct
- Existing portal-routing suites (test_nous_portal_routing.py,
test_issue895_894_nous_prefix.py) continue to pass.
- Full test suite: 5320 passed, 4 skipped, 3 xpassed.
Reported on Discord by @vishnu (Nathan forwarded as #2177).
When a provider's 'models' config contains dicts (e.g. {"id": "x", "label": "y"})
instead of plain strings, _apply_provider_prefix() crashes with:
AttributeError: 'dict' object has no attribute 'startswith'
This happens because the list comprehension at line 3505 passes the raw dict
as the model ID. The fix extracts 'id' and 'label' from dict entries while
keeping string entries as-is.
Fixes the /api/models and /api/onboarding/status 500 errors.
CI on Python 3.13 (clean editable install, no hermes_cli package) was still
failing the 3 lmstudio tests after the first fix attempt. Root cause: the
outer try/except in the lmstudio branch was catching ImportError from
`from hermes_cli.models import provider_model_ids`, hijacking the whole
branch and silently skipping the urlopen fallback.
Restructured into two independent tiers:
1. hermes_cli lookup in its own try/except — ImportError logs at DEBUG
and continues with lm_ids=[].
2. urlopen fallback runs unconditionally when lm_ids is empty, including
after hermes_cli import failure.
New regression test `test_lmstudio_fallback_works_when_hermes_cli_unavailable`
explicitly blocks hermes_cli via sys.meta_path and verifies the lmstudio
group still populates from the urlopen fallback. Without this test, the
CI-vs-local divergence (local env had hermes_cli installed, CI didn't)
would keep slipping through.
All 12 lmstudio-related tests pass, including the 3 #1527 tests that
broke on stage-337.
PR #1970 added a dedicated `elif pid == "lmstudio":` branch in
`get_available_models()` that fetches the live /v1/models list when the
hermes_cli helper doesn't have ids cached. The fallback path inside that
branch only looked at `cfg["providers"]["lmstudio"]["base_url"]`, missing
the historical config shape where the URL lives under `cfg["model"]`:
model:
provider: lmstudio
base_url: http://192.168.1.22:1234/v1 ← here, not under providers.lmstudio
providers:
lmstudio:
api_key: local-key
3 pre-existing tests in tests/test_issue1527_lmstudio_base_url_classification
broke on stage-337 because of this — they passed on master, failed after
the PR #1970 merge.
The simpler fix is to enhance the already-introduced `_get_provider_base_url()`
helper so it falls back to `cfg["model"]["base_url"]` when
`cfg["model"]["provider"] == provider_id`, then use the helper inside the
lmstudio branch instead of a direct lookup. This keeps the previous
behaviour (where the generic configured-provider branch handled lmstudio
via the model block) while preserving PR #1970's live-discovery additions.
Belt-and-suspenders: `_get_provider_base_url()` explicitly does NOT inherit
model.base_url for providers other than the active one — if a user's config
says `model.provider: anthropic` and they have `providers.openai` configured
without a base_url, openai must still resolve to None (use SDK default),
not to the anthropic proxy URL.
6 new regression tests in tests/test_pr1970_lmstudio_base_url_fallback.py
lock the two-location lookup, the precedence rule (explicit providers entry
wins over model fallback), trailing-slash stripping, and the negative case
(model.base_url MUST NOT leak to non-active providers).
All 51 tests in the existing model-resolver + custom-provider banks still
pass.
Caught by maintainer review on stage-337 (full pytest with the new network
isolation in place surfaced the regression that the fork-CI mock-server path
would have hidden).
Add xiaomi to _PROVIDER_DISPLAY, _PROVIDER_MODELS, and _PROVIDER_ALIASES
so the WebUI recognizes Xiaomi as a first-class provider.
Models included:
- mimo-v2.5-pro (MiMo V2.5 Pro)
- mimo-v2.5 (MiMo V2.5)
- mimo-v2-pro (MiMo V2 Pro)
- mimo-v2-omni (MiMo V2 Omni)
- mimo-v2-flash (MiMo V2 Flash)
Aliases: mimo, xiaomi-mimo -> xiaomi
The hermes-agent CLI already registers xiaomi as a provider
(hermes_cli/models.py, hermes_cli/auth.py) but the WebUI was missing
the corresponding entries, causing the model dropdown to fall back to
OpenRouter and the provider list to show 'Unsupported'.
- api/config.py: resolve merge conflict, keep both _custom_slug_rest_looks_like_host_port
and new _get_provider_base_url helper. Custom providers now return their configured
base_url in resolve_model_provider(). Add 'Configured' badge for explicitly configured
providers in the models dropdown. Detect LM Studio via LM_API_KEY+LM_BASE_URL env vars.
Fetch live loaded models from LM Studio with fallback to direct HTTP requests.
- api/providers.py: fetch live LM Studio model list via hermes_cli for the providers card.
- static/style.css: add purple 'Configured' badge style.
When multiple custom providers expose the same model ID (e.g. baidu,
huoshan, and liantong all offering glm-5.1), only the first provider's
entry was shown in the model dropdown.
Root cause (backend): used the bare model ID as the
dedup key, so the second and subsequent providers with the same model
were silently skipped.
Root cause (frontend): stripped the @provider: prefix before
comparing, so @custom:baidu:glm-5.1 and @custom:huoshan:glm-5.1 were
treated as duplicates.
Fix:
- Backend: change _seen_custom_ids key to '{slug}:{model_id}' so each
provider's models are tracked independently.
- Frontend: add _providerOf() helper and deduplicate on the composite
(normId, provider) key instead of normId alone. Bare model IDs
(without @provider: prefix) still deduplicate on normId for backward
compatibility.
model_with_provider_context can emit @custom:<host>:<port>:<model> when
model_provider is derived from an OpenAI base_url authority (e.g.
custom:10.8.0.1:8080). The colon-count heuristic meant for @custom:slug:model:free
mistook those extra colons for an over-split model ID and prepended the port
segment onto the bare model (8080:Qwen3-235B), breaking WebUI while CLI/curl
stayed correct.
Detect endpoint-style slugs (IPv4/localhost/hostname + numeric port) and skip
the peel in that case. Add regression tests for IPv4, dotted hostname,
localhost, and model_with_provider_context round-trip.
The goal evaluation hook was firing on every completed assistant turn
when a goal was active, even for unrelated messages like "what time is
it". This burned the goal budget, triggered continuation prompts that
interrupted unrelated conversations, and made /goal status numbers
misleading.
Add STREAM_GOAL_RELATED and PENDING_GOAL_CONTINUATION flags to gate
the evaluate_goal_after_turn() call in the streaming loop. Only streams
started from goal kickoff (/goal <text>) or goal continuation are
marked as goal-related. Normal user messages skip the hook entirely.
Conflict resolution: both #1928 (session jump buttons) and #1929 (endless
scroll) add their own settings/UI/i18n keys. Resolved by keeping both —
the features are independent opt-in toggles.
Two bugs in get_available_models() conspired to duplicate the active
provider's auto-detected models under a phantom 'Custom' group whenever
custom_providers was also declared in config.yaml:
1. custom:* PIDs not in _named_custom_groups (e.g. stale slugs left from
prior configs) fell through to the auto_detected_models fallback, copying
the active provider's whole catalog into a phantom Custom: <slug> group.
Fix: continue unconditionally for ANY custom:* PID — the named-group
branch is the only legitimate population path.
2. The bare 'custom' PID, with the active provider being concrete (e.g.
ai-gateway), hit 'elif auto_detected_models: copy.deepcopy(...)' and
built a duplicate Custom group of the active provider's models with
mismatched provider prefixes. Fix: when pid == 'custom' and the active
provider is non-custom, leave models_for_group empty.
The reporter also suggested a third fix gating resolve_model_provider() on
config_provider — that's intentionally NOT applied because it conflicts with
the long-standing model-specific-override semantics covered by
test_model_resolver.py::test_custom_provider_*_routes_to_named_custom_provider
(custom_providers entries explicitly override the active provider's routing
when the user opted-in). The reporter's symptom (duplicate UI group) lives
entirely in get_available_models()'s group construction and is fully fixed
by the two changes above.
Tests: 6 new regression tests (3 in #1881 file + reuse), 774 broader
tests still green (model/provider/custom/config domain).
Two in-stage fixes for v0.51.19 batch:
1) api/config.py — add resolve_alias=False param to
_resolve_configured_provider_id() and pass it from
resolve_model_provider(). The PR #1818 swap from
_resolve_provider_alias() to _resolve_configured_provider_id()
was correct for active-provider/badge surfaces but broke #1625's
local-server-provider literal-preservation contract: 'ollama' →
'custom' and 'lm-studio' → 'lmstudio' alias-collapse caused
_LOCAL_SERVER_PROVIDERS membership check to miss, breaking the
model-id full-path preservation for LM Studio/Ollama. The new
flag preserves the raw provider value when called from
resolve_model_provider, and named-custom-slug + base-url
fallback both still run unchanged.
2) tests/test_bootstrap_discover_agent.py — pin Path.home() in
_isolate_discover_agent_dir so the hard-coded
'Path.home() / .hermes / hermes-agent' / 'Path.home() /
hermes-agent' candidates in discover_agent_dir() can't pick up
the dev machine's real install. The original PR #1817 isolation
helper covered HERMES_HOME, HERMES_WEBUI_AGENT_DIR, and
REPO_ROOT but missed the Path.home() leak.
Both surfaced on full pytest pre-release gate, fixed in stage,
ship in v0.51.19. Tests: full suite green.
PR #1762 fixed the rsplit grammar collision for plain @openrouter:model:free
qualifiers, but skipped the fallback whenever the provider hint started with
'custom:' on the assumption that custom providers route directly. That left
'@custom:my-key:some-model:free' broken: rsplit yields
provider='custom:my-key:some-model', bare='free' → custom guard skips the
split-fallback → returns provider='custom:my-key:some-model', model='free'.
Detect the over-split structurally instead of using a known-suffix allowlist:
custom hints carry exactly one segment after 'custom:' (constructed at
api/config.py:1363 as 'custom:' + entry_name). So any rsplit result of
'custom:<a>:<b>' with bare model '<c>' has eaten one model segment — peel
it back with a second rsplit and prepend it to the bare model.
This is robust for :free / :beta / :thinking / :preview / any future
OpenRouter suffix without an allowlist to maintain.
Adds 5 regression tests covering the matrix (free/beta/thinking/preview/
slashed-model). All 7 existing #1744 tests still pass; #1228 tests
unaffected.
Co-authored-by: Cake <51058514+Sanjays2402@users.noreply.github.com>
The previous approach of prepending 'openrouter/' to the model ID in the
catalog was incorrect — it only masked the symptom while regressing the
config_provider=openrouter codepath.
The root cause is in resolve_model_provider(): rsplit(':', 1) on
'@openrouter:tencent/hy3-preview:free' yields provider='openrouter:tencent/hy3-preview'
and model='free', because the ':free' suffix collides with the @provider:model
grammar.
Fix: after rsplit, validate that the extracted provider hint is a known
provider (in _PROVIDER_MODELS, _PROVIDER_DISPLAY, or starts with 'custom:').
If not, fall back to split(':', 1) so trailing suffixes stay attached to
the model ID.
This fixes all current and future OR models with colon-suffixed tags
(:free, :beta, :thinking, :nitro, etc.) without catalog changes.
Also adds regression tests for the affected models and edge cases.
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
PR #1728's path/mtime-aware get_config() reload broke the common test
idiom monkeypatch.setattr(config, 'cfg', {...}). The cfg = _cfg_cache
alias bound at import time means the rebinding only changes the module
attribute; _cfg_cache stays unchanged, so _cfg_has_in_memory_overrides()
returned False and the path-aware reload silently overwrote the test's
override. test_issue1426_openrouter_* and test_issue1680_codex_* failed
in the full suite while passing standalone — exact polluter signature.
Fix:
- _cfg_has_in_memory_overrides() now also detects cfg-rebind via
cfg is not _cfg_cache.
- get_config() returns cfg (the override) when it differs from
_cfg_cache, so callers see the test's intended override.
- 4 new regression tests pin both prongs in
test_stage302_config_override_regression.py.
Defense-in-depth (prong 2 of test-isolation-flake-recipe):
- test_sprint3.py::test_skills_list and test_skills_list_has_required_fields
now skip on empty skills list rather than asserting > 0 / IndexError, so
future profile-switch / SKILLS_DIR repointing pollutions don't break
the build. The contract under test is 'API returns a non-empty list
when there are entries' — empty list signals a polluter elsewhere.
Pre-existing wall-clock flake fix (absorb-in-release):
- test_issue1144_session_time_sync.py::test_relative_time_uses_server_clock
now pins Date.now() to a fixed instant. Without pinning, when CI runs
near 08:00 UTC the projected server time crosses midnight and '5 minutes
ago' silently becomes '1d'. Same time-of-day-pin pattern as the sibling
test_session_bucket_uses_server_clock used.
Test count: 4580 → 4584 (+4 regression tests). 0 failures, stably green
across multiple runs.
Constituent PRs:
#1637 by @Michaelyklam — protect raw pre from glued-bold lift (closes#1451)
#1639 by @bergeouss — macOS auto-scroll race + custom:* provider list (closes#1360, #1619)
#1642 by @nesquena-hermes — YAML/JSON/diff code block newlines (closes#1618, #1463)
Opus advisor SHIP verdict on stage-295. One observation absorbed:
- api/config.py:2533 dead-code comment per Opus (defensive belt-and-braces
for #1619 fallback; load-bearing fix is in routes.py /api/models/live)
PR #1641 (Michaelyklam parallel-discovery duplicate of #1642) closed as
superseded; UI media adopted with co-author trailer.
4245 → 4255 tests passing (+10).
#1360 — On macOS WKWebView, trackpad momentum scrolling fires scroll
events that interleave with the _programmaticScroll setTimeout(0) guard.
A mid-momentum scroll event either gets swallowed (_programmaticScroll
still true) or falsely reports nearBottom (momentum hasn't settled),
keeping _scrollPinned=true and snapping the viewport back down.
Fix: rAF-debounce the scroll listener so the nearBottom check runs at
the next paint frame when the browser's scroll position has settled.
Added a hysteresis counter requiring 2 consecutive near-bottom samples
before re-pinning, preventing accidental re-pin during deceleration.
#1619 — When a custom:* provider (e.g. custom:relay via custom_providers)
has models that overlap with auto-detected models from base_url /v1/models,
the dedup logic at config.py:2263 skipped them all. The named custom
group ended up empty, and the continue at line 2334 silently discarded
the auto-detected models. Result: only the default model appeared.
Fix 1 (config.py): When custom:* named group has 0 models after dedup,
fall back to auto_detected_models_by_provider instead of dropping them.
Fix 2 (routes.py): Extended /api/models/live fallback to handle
custom:* slugs (not just bare "custom") for both custom_providers
config lookup and base_url live fetch.
Closes#1633. STATE_DIR/models_cache.json was persisted across server
restarts without any version stamp, so a Docker container update from
version A to B read the cache file written by version A — users saw
stale picker contents (missing models, phantom provider groups) for
up to 24 hours until either the TTL expired, an unrelated provider
edit triggered invalidate_models_cache(), or they manually deleted
the file.
Reporter Deor (Discord) updated to v0.50.292 — which contained fixes
for #1538, #1539, and #1568 — did a hard refresh and cleared site
data, and still saw byte-for-byte identical picker contents because
the server kept reading the v0.50.281 cache file off the host-mounted
state volume.
Fix:
* _save_models_cache_to_disk() stamps payloads with _webui_version
(resolved lazily from api.updates.WEBUI_VERSION via sys.modules
lookup to avoid the api.config <-> api.updates circular import)
and _schema_version = 2.
* New _is_loadable_disk_cache() validator checks both stamps in
addition to shape. Mismatch on either field rejects the load.
* _load_models_cache_from_disk() calls the new validator and
strips the disk-only metadata before returning, so the rest of
the code sees the same shape it always did.
* _is_valid_models_cache() kept loose (shape-only) so in-memory
cache writes that never touch disk don't fail validation.
Schema version is independent of the WebUI version stamp so future
cache-shape changes can invalidate older releases without relying
on a tag bump alone.
Early-init edge case (api.updates not yet loaded) skips the version
check rather than wedging the boot — at worst an unstamped file is
written once and rejected on the next call.
Updated existing tests/test_model_cache_metadata.py to use subset/
round-trip semantics rather than byte-for-byte equality, since the
disk payload now has additional stamps. The four response-shape
fields still round-trip verbatim; the load result is unchanged
(stamps stripped). 19 new regression tests.
4180 -> 4199 tests pass.
SHOULD-FIX: rate-limit _repair_stale_pending repair-firing telemetry. Switch
from unconditional logger.warning to age-keyed: WARNING when pending_age <
5min (the diagnostically valuable race window — actual leak-path candidates
that slipped past the grace guard) and DEBUG for the long-tail (orphaned
sidecars from prior process lifetimes). Prevents reconnect loops on stuck
sessions from flooding the log while preserving the diagnostic signal we
want for tuning _REPAIR_STALE_PENDING_GRACE_SECONDS empirically.
NIT: _LOCAL_SERVER_PROVIDERS expanded with lm-studio (hyphenated alias used
in some custom_providers configs and already recognized at api/config.py:2189
for SSRF host trust) and localai (LocalAI project). Test parametrize expanded
from 7 to 11 names, also covering pre-existing koboldcpp and textgen for
symmetry. +4 regression tests.
NIT (docs): CHANGELOG callout for the RFC1918 behavior change. Internal-
network OpenAI-compatible proxies now preserve the model prefix on private-IP
base_urls. Documented the migration path: configure as a custom_providers
entry to bypass the local-server detection.
NIT (deferred, optional): narrowing the heuristic to is_loopback only is
left as future work; the broader scope was an explicit goal in the bug
body and Opus flagged it as SHOULD-DISCUSS-but-not-block.
4184 -> 4188 passing. 0 regressions. ~10 LOC absorbed total.
Closes#1623 — Lower SSE app heartbeat from 30s to 5s at every long-lived
handler (main agent, terminal, gateway-watcher, approval-poller, clarify-poller).
Kernel TCP keepalive declares peer dead at 25s worst-case (10s KEEPIDLE +
5s KEEPINTVL * 3 KEEPCNT, added v0.50.289 #1581). 30s app heartbeat let the
kernel tear sockets down on flaky networks before the app sent its first
keepalive byte — drops at ~10s during long thinking phases. New named
constant _SSE_HEARTBEAT_INTERVAL_SECONDS=5; regression test pins the
inequality (app_heartbeat * 2 <= kernel_window) so future tuning can't
re-introduce the misalignment.
Closes#1624 — Add 30s grace period to _repair_stale_pending() trigger.
Without it, any narrow race between the streaming thread clearing
pending_user_message and STREAMS.pop(stream_id) produces a false-positive
'Previous turn did not complete.' marker on a turn that finished correctly
(reproducible after every command-approval turn). Defense-in-depth, not
the root-cause fix — the actual streaming-thread leak path is tracked
separately. Falsy pending_started_at (legacy sidecars) treated as
'old enough' so legitimate legacy-data recovery still works. Plus
logger.warning telemetry on every legitimate repair so the next batch of
user reports tells us whether the underlying race still fires.
Closes#1625 — Local model servers (LM Studio, Ollama, llama.cpp, vLLM,
TabbyAPI, koboldcpp, textgen-webui) now keep the full HuggingFace-style
model id (e.g. 'qwen/qwen3.6-27b' instead of stripped 'qwen3.6-27b'). New
_LOCAL_SERVER_PROVIDERS set + _base_url_points_at_local_server() loopback/
RFC1918 heuristic — either signal triggers no-strip. Backward compat
preserved for OpenAI-compatible proxies on public hosts (LiteLLM at
litellm.example.com still strips openai/gpt-5.4 -> gpt-5.4). Updated the
existing #230/#433 test to reflect that #1625 supersedes the strip-on-custom
rule for loopback hosts (see api/config.py and test_model_resolver.py
docstring update). Reported by @akarichan8231 in Discord on 2026-05-04.
42 regression tests across:
tests/test_issue1623_sse_heartbeat_alignment.py (3)
tests/test_issue1624_repair_stale_pending_grace.py (9)
tests/test_issue1625_local_server_model_id_preservation.py (30)
4142 -> 4184 passing. 0 regressions.
Reporter (Deor, Discord #report-bugs, May 03 2026 14:19 PT, relayed by
@AvidFuturist) saw the Settings → Default Model dropdown rendering the
OpenCode Go provider as TWO separate optgroups: "OpenCode Go" (the
canonical one with all 14 catalog models) and "Opencode_Go" (a phantom
group containing one self-referential entry).
Three structural causes, all in api/config.py:_build_available_models_uncached:
1. **Detection-path id leakage.** The detection block at line ~1980
reads cfg["providers"] keys verbatim. If the user's config has
``providers.opencode_go.api_key`` (underscore variant) AND another
path adds the canonical ``opencode-go`` (e.g. via active_provider),
both end up in detected_providers and the build loop creates two
distinct provider groups with the second labelled via the
``pid.title()`` fallback as ``"Opencode_Go"``.
2. **Injection-block rogue model.** The default-model injection block
at line ~2598 puts ANY ``model.default`` string into the picker as
a fake option. A stray ``model.default: opencode_go`` (provider id
mistakenly used as a model id) surfaces as a phantom model
labelled ``"Opencode GO"``.
3. **Empty-group bleed.** When a non-canonical provider id makes it
into detected_providers but has no entry in _PROVIDER_MODELS, the
build loop creates an optgroup with zero models — pure UI noise.
This PR addresses all three:
- **New `_canonicalise_provider_id()` helper** that folds underscores
to hyphens, lowercases, and applies alias resolution only when the
alias target is itself a canonical id in `_PROVIDER_DISPLAY`. The
last constraint avoids round-tripping ``x-ai`` (canonical) through
the alias table to ``xai`` (which the WebUI doesn't index by).
- **Detection-path canonicalisation.** The cfg["providers"] scan
applies the helper before adding to detected_providers. Same
treatment in the only_show_configured intersection so that mode
doesn't accidentally exclude the canonical id when configured_providers
only contains the underscore-variant key.
- **Post-collection dedup pass** that re-canonicalises every entry in
detected_providers — belt-and-braces against future regressions in
any of the ~25 ``detected_providers.add(...)`` callsites without
auditing each one. Idempotent for already-canonical ids.
- **Provider-id guard on the model.default injection block.** When
the injected value matches a known provider display name or alias
(after underscore/case normalisation), skip the injection and emit
a `logger.warning` instead. Real unknown model ids (newly released
models, custom endpoints) still get injected — only provider-shaped
values are rejected.
- **Empty-group filter at end of build.** Drop optgroups with zero
models. Custom: groups (`provider_id` starts with `custom:`) are
exempt — users may want an empty card visible as a reminder.
Tests
-----
`tests/test_issue1568_duplicate_provider_groups.py` (17 tests):
- TestCanonicaliseProviderId (8): unit tests pinning helper behaviour —
canonical preserved, underscore folded, case folded, aliases
resolved, x-ai not round-tripped, empty input, unknown ids
normalised, idempotence
- TestProviderGroupDedup (4): end-to-end picker behaviour —
underscored providers-key produces ONE group not two (Deor's case),
uppercase providers-key collapsed, aliased keys (z-ai → zai)
collapsed, happy path unchanged
- TestDefaultModelProviderIdGuard (3): provider id as model.default
doesn't inject phantom + WARNING logged; alias as model.default also
caught; legitimate unknown model IDs (forward-compat) still injected
- TestEmptyGroupFilter (2): empty optgroups dropped from picker;
custom: providers exempted from filter
Plus one structural test fix in
`tests/test_issue604_all_providers_model_picker.py:test_cfg_providers_only_adds_known`
— widened the regex window from 500 to 1500 chars so the new
documentation comment block doesn't push `_PROVIDER_MODELS` past the
substring slice. Pre-existing brittle window pattern, not a new issue.
Verification
------------
Live on port 8789 with Deor's exact reproduction config
(`providers.opencode_go.api_key` + `model.provider: opencode-go`):
/api/models groups: 1 (was 2)
Browser <select> optgroups: 1 (was 2)
Total options under "OpenCode Go": 14 (was 14 in real group + 0 in phantom group)
Five-scenario sweep all collapse to ONE provider group:
| Config shape | Pre-fix | Post-fix |
|---|---|---|
| Hyphenated provider + underscored providers-key (Deor's case) | 2 groups | 1 group ✅ |
| Hyphenated provider + UPPERCASE providers-key | 2 groups | 1 group ✅ |
| Aliased providers-key (z-ai resolved to zai) | 2 groups | 1 group ✅ |
| model.default = provider-id (orig #1568 scenario) | 15 models with phantom | 14 models, no phantom ✅ |
| Happy path (canonical-only) | 1 group | 1 group ✅ |
4070 pytest passed (was 4053 → 4070, +17 from this PR).
3 CI runs to follow on push.
QA harness 11/11 passed.
JS unaffected — pure backend fix.
Reporter: Deor (Discord #report-bugs, May 03 2026 14:19 PT)
Relayed by: @AvidFuturist
Two related dropdown bugs in one PR — same root shape (model-picker
endpoints disagreeing about which Nous Portal models exist) plus the
preemptive UX guard against the picker becoming unusable on large-tier
Nous accounts.
#1567 — Endpoint disagreement
=============================
Reporter (Deor, Discord, May 03 2026) saw Settings → Providers card
showing "Nous Portal — 396 models · OAuth" while the in-conversation
picker dropdown listed only the four hardcoded curated entries.
Two structural causes:
1. ``api/providers.py:get_providers`` iterates ALL OAuth providers
regardless of authentication state and unconditionally live-fetches
the catalog.
2. ``api/config.py:_build_available_models_uncached`` only iterates
providers in ``detected_providers``, gated on
``hermes_cli.models.list_available_providers().authenticated``.
That flag can disagree with ``get_auth_status(<id>).logged_in`` on
some hermes_cli versions.
When the disagreement happens for Nous, the picker silently falls
through to the curated 4-entry static list while the providers card
keeps showing the live catalog — exactly the asymmetry users report.
Plus: the Nous live-fetch branch in `_build_available_models_uncached`
fell back to the same curated 4-entry list when `provider_model_ids`
returned an empty list (transient failure / OAuth refresh in flight),
which doubles down on the disagreement instead of healing it.
UX cap (the design concern Nathan flagged on triage)
====================================================
Even with the disagreement fixed, dumping a 397-model catalog into a
flat dropdown is unusable. We trim the visible picker to a curated
~15-entry featured set when the catalog exceeds 25 models, and surface
the rest under a new ``extra_models`` field so:
- ``/model`` slash autocomplete (commands.js) covers the full catalog
- ``_dynamicModelLabels`` (ui.js) hydrates from both lists, so a model
selected from outside the featured slice still gets a proper label
- The optgroup label gets ``" (15 of 397)"`` appended so the user
understands the dropdown is intentionally trimmed, not broken
- The providers card surfaces ``models_total`` separately so the
header still reads "397 models · OAuth"
- A small "+N more" disclosure pill appears at the end of the rendered
pill list (only fires for non-OAuth providers — OAuth cards never
render pills) with a tooltip pointing at the slash command
Featured selection rules
------------------------
Deterministic; same algorithm runs in both `/api/models` and
`/api/models/live` so background enrichment doesn't undo the trim:
1. Always include the user's currently-selected model (sticky — no
orphan IDs in the dropdown after a refresh)
2. Always include every entry from the curated static
``_PROVIDER_MODELS["nous"]`` list whose id maps onto a live id
3. Top up to 15 by walking ``_NOUS_VENDOR_PRIORITY`` round-robin
(one model per vendor each pass) so no vendor monopolises the slots
Changes by file
===============
api/config.py
- New `_format_nous_label` neighbour: `_NOUS_FEATURED_THRESHOLD = 25`,
`_NOUS_FEATURED_TARGET = 15`, `_NOUS_VENDOR_PRIORITY` tuple,
`_build_nous_featured_set()` helper (~80 LOC)
- `_build_available_models_uncached` Nous branch:
- Apply featured-set cap with sticky-selection signal
- Return `extra_models` alongside `models` for the catalog tail
- Decorate optgroup label with truncation count
- Drop stale-4 fallback when authenticated but live-fetch empty
(omit the group entirely; truth lives in the providers card and
the next cache rebuild will heal it)
- Keep stale-4 fallback when hermes_cli is unavailable (test envs,
package mismatches) — that's a different failure mode
- Detection symmetry: explicit `get_auth_status("nous").logged_in`
check after the existing `list_available_providers()` loop, so the
picker matches the providers card on hermes_cli versions where the
two signals disagree
api/providers.py:get_providers
- Apply same featured-set cap so card body doesn't render 397 pills
- Add `models_total` field reporting full catalog size (used by
frontend for the "N models · OAuth" header text)
api/routes.py:_handle_live_models
- Apply same featured-set cap for `/api/models/live` so background
enrichment via `_fetchLiveModels()` doesn't undo the dropdown trim
- Use sticky-selection from `cfg["model"]["model"]` matching the main
endpoint's logic
static/ui.js:populateModelDropdown
- Hydrate `_dynamicModelLabels` from `g.extra_models` so a selection
outside the visible dropdown still renders with its proper label
static/commands.js:_loadSlashModelSubArgs
- Iterate `group.extra_models` so `/model` autocomplete covers the
full catalog (not just the trimmed featured slice)
static/panels.js:_buildProviderCard
- Header count uses `p.models_total` (full catalog size) instead of
`p.models.length` (trimmed slice)
- Render trailing "+N more" disclosure pill when `models.length <
models_total` with a tooltip pointing at the slash command
static/style.css
- New `.provider-card-model-tag-more` rule (italic, dashed border,
cursor:help, no select) — visually distinct from real model pills
Tests
=====
`tests/test_issue1567_nous_picker_capacity_and_symmetry.py` (20 tests):
- TestBuildNousFeaturedSet (8): unit tests on the helper —
small-catalog no-op, large-catalog cap to target, disjoint+complete
invariants, priority-vendor round-robin guarantee, sticky selection
with and without `@nous:` prefix, curated-flagship preservation,
empty-catalog handling, determinism
- TestApiModelsLargeCatalog (2): /api/models cap behavior end-to-end
on a synthetic 397-model catalog vs a 20-model catalog
- TestNousDetectionSymmetry (2): picker includes Nous when
`get_auth_status` agrees but `list_available_providers` disagrees;
picker omits Nous when both disagree
- TestNousLiveFetchEmpty (2): authenticated + empty-fetch omits group;
hermes_cli unavailable still falls back to static-4
- TestProvidersCardPickerSymmetry (1): both endpoints agree on
exactly the same featured-set IDs + total catalog count
- TestFrontendExtrasContract (4): static-source assertions pinning
the JS contract for `extra_models`, `models_total`, and the "+N more"
disclosure
Verified live on port 8789 (30-model catalog):
- /api/models Nous group: provider="Nous Portal (15 of 30)", 15 models,
15 extra_models
- /api/models/live?provider=nous: 15 entries (matches main path)
- /api/providers Nous card: models_total=30, models=15
- Browser dropdown after backfill: 15 options, 30 entries in
_dynamicModelLabels
- Sticky selection: Claude Opus 4.7 (the active model) in the featured
slice as expected
4073 pytest passed (was 4053 → 4073, +20 from this PR).
3 CI test runs (3.11/3.12/3.13) green.
QA harness 11/11 passed.
Reporter: Deor (Discord #report-bugs, May 03 2026 14:15 PT)
Relayed by: AvidFuturist
Augments @bergeouss's PR #1548 v2 with the structural fix the issue
actually requested. The original PR added 5 hardcoded entries to
_FALLBACK_MODELS which would rot fast as OpenRouter's free-tier roster
turns over monthly.
Adds proper live-fetch logic to the OpenRouter group population so the
free-tier list stays fresh without requiring a code release every time
a new free model lands.
api/config.py:2120 — replaces the static _FALLBACK_MODELS slice with:
1. Live curated catalog via hermes_cli.models.fetch_openrouter_models()
— applies the tool-support filter (Kilo-Org/kilocode#9068).
2. Free-tier live fetch — direct call to https://openrouter.ai/api/v1/models,
filtered to free-tier-only (pricing.prompt == 0 AND pricing.completion
== 0, OR :free suffix), bypasses the tool-support filter so newly-added
free variants appear even before OpenRouter annotates them with tools.
Capped at 30 entries to keep the picker usable.
3. Defense-in-depth fallback to _FALLBACK_MODELS (which retains
@bergeouss's hardcoded list for offline / test envs).
4. Deduplication via seen_ids — model in both surfaces appears once.
5 new tests + 1 fixed test in tests/test_minimax_provider.py (scoped the
provider='MiniMax' assertion to direct-MiniMax routes by filtering for
'minimax/' prefix and excluding ':free' since the OpenRouter free-tier
variant minimax/minimax-m2.5:free correctly carries provider='OpenRouter').
Co-authored-by: bergeouss <[email protected]>
Per review observation on PR #1544: the docstring claimed
'Gemini 3.1 Pro Preview' and 'Nemotron 3 Super 120B A12B' but the
helper reuses _format_ollama_label's 3-letter-token rule, which
uppercases 'PRO' (and the existing rule for tokens like 'a12b'
renders 'A12b' not 'A12B'). Update the examples to match actual
behavior — labels are unchanged, only the docstring.
Pure-comment change, no behavioral effect. Test counts unchanged
(4013 passed).
Closes#1538, #1539. Two related dropdown-staleness bugs reported by Deor
(Discord, May 03 2026).
#1538 — Nous Portal picker showed only 4 hardcoded models
=========================================================
The Settings → Default Model picker, the composer model dropdown, the
/model slash command, and the Settings → Providers card all showed only
four Nous models (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.4 Mini, Gemini
3.1 Pro Preview) because `_PROVIDER_MODELS["nous"]` had four hardcoded
entries and `_build_available_models_uncached()` fell through to the
generic `pid in _PROVIDER_MODELS` branch.
The actual Nous Portal catalog has 30 models live — Claude Opus 4.7, GPT-5.5,
Kimi K2.6, MiniMax M2.7, Gemini 3.1 Pro/Flash, several Xiaomi/Tencent/StepFun
entries, and more.
Fix:
- New `_format_nous_label()` helper in `api/config.py` — reuses the
`_format_ollama_label()` token rules, drops the vendor namespace, and
appends ` (via Nous)` so labels disambiguate from same-named direct-
provider entries (e.g. "Claude Opus 4.7" via direct Anthropic).
- New `elif pid == "nous":` branch in `_build_available_models_uncached()`
mirroring the Ollama Cloud pattern: live-fetch through
`hermes_cli.models.provider_model_ids("nous")`, prefix every id with
`@nous:` (matches the existing routing convention from PR-era #854 and
pinned in tests/test_nous_portal_routing.py), fall back to the curated
4-entry static list when hermes_cli is unavailable.
- Same fix applied to `api/providers.py:get_providers()` — that's the
separate code path that builds Settings → Providers card models, and
it had the identical bug shape.
#1539 — Removed provider lingered in dropdowns until restart
============================================================
After Settings → Providers → Remove, the provider's models still appeared
in every model dropdown until the page was reloaded. The server-side
TTL cache was correctly flushed (`set_provider_key()` calls
`invalidate_models_cache()` on both add and remove) but JS-side caches
were never dropped:
- `_slashModelCache` / `_slashModelCachePromise` (commands.js) — feeds
the `/model` slash-command suggestions.
- `_dynamicModelLabels` / `window._configuredModelBadges` (ui.js) —
populated by `populateModelDropdown()` on app boot and profile switch.
Pre-fix, `_removeProviderKey()` only called `loadProvidersPanel()`
which refreshed the providers card list but never asked any consumer
to re-fetch /api/models.
Fix:
- `static/commands.js`: new `_invalidateSlashModelCache()` helper that
nulls both cache slots, exposed on `window` (typeof-guarded so the
module remains importable in headless vm contexts — needed by the
existing tests/test_cli_only_slash_commands.py harness).
- `static/panels.js`: new `_refreshModelDropdownsAfterProviderChange()`
helper that calls the invalidator + `populateModelDropdown()`, wrapped
in try/catch so the providers panel update never breaks if a
downstream module hasn't loaded yet. Both `_saveProviderKey` and
`_removeProviderKey` invoke it (defense-in-depth: same staleness shape
applies to the add path too).
Tests
-----
- `tests/test_issue1538_nous_live_catalog.py` (12 tests): live-fetch
surfaces ≥20 entries, every id starts with `@nous:`, every label ends
with ` (via Nous)`, recent flagships (Opus 4.7, GPT-5.5, Kimi K2.6,
Gemini 3.1 Pro, MiniMax M2.7) reach the dropdown, static fallback
works when hermes_cli raises, label formatter unit tests (vendor
namespace stripping, variant rendering, MiniMax mixed-case), the
curated static list and its routing invariants are preserved.
- `tests/test_issue1539_provider_removal_dropdown_invalidation.py`
(11 tests): invalidator helper exists and clears both cache slots,
exposed on window with typeof guard, both save and remove paths
invoke the dropdown flush, helper calls both invalidator and
populateModelDropdown, helper is resilient to missing modules,
helper does not block panel refresh, server-side
`set_provider_key → invalidate_models_cache` invariant pinned.
Verified live on port 8789: `/api/models` Nous group returns 30
models (was 4); browser `document.getElementById('modelSelect')`
exposes 30 options under the "Nous Portal" group; the dropdown-flush
helper is callable from the browser and round-trip rebuild keeps the
dropdown at 30 options.
Test counts:
- Full pytest: 4013 passed, 2 skipped, 3 xpassed, 0 failures
(was 3990 → 4013, +23 from this PR).
- QA harness pytest: 20 passed.
- Browser API sanity: 11/11 passed.
- Agent Browser CDP: 21/23 passed (the 2 SSE liveness failures
reproduce on master and are unrelated to this PR).
Supersedes contributor PR #1511 (lost9999), which removed the label-suffix
logic in _deduplicate_model_ids() but left the underlying shared-reference
bug intact — IDs would still be silently corrupted across provider groups,
just with cleaner-looking labels.
## Bug shape
When multiple unconfigured providers (Ollama / HuggingFace / custom
endpoints / Google Gemini CLI / Xiaomi / etc.) all fell through to the
'else' branch in api/config.py:get_models_grouped() that ends with:
groups.append({..., "models": auto_detected_models})
every group ended up sharing the SAME list reference AND the SAME dicts
inside. When _deduplicate_model_ids() then mutated those dicts to add
@provider_id: prefixes and provider-name parentheticals, the changes were
applied to every group that referenced the same dict.
Visible symptom: user 'vishnu' reported the dropdown showing
'Deepseek V4 Flash (Xiaomi) (Ollama) (HuggingFace) (Google-Gemini-Cli)'
on every group. Hidden symptom (worse): the 'id' field collapsed to
'@xiaomi:deepseek-v4-flash' on every group too, so clicking the entry
under any group routed the request to Xiaomi.
## Fix
api/config.py:2078 — wrap auto_detected_models in copy.deepcopy() at the
groups.append site so each group gets its own independent dicts. The
existing _deduplicate_model_ids() logic is correct and unchanged; the
bug was in the assignment site, not the dedup function.
The single-parenthetical disambiguation in labels is retained because
the composer chip (composer-model-label) shows the model label without
the optgroup header context — 'Deepseek V4 Flash (Ollama)' is more
useful than ambiguous 'Deepseek V4 Flash' there.
## Tests
tests/test_issue1511_dedup_shared_reference.py — 3 new tests:
- test_groups_have_independent_model_lists: structural invariant pin
- test_unconfigured_providers_no_shared_dedup_bleed: end-to-end against
the corrected code path; verifies each group gets its own @provider_id:
prefix and exactly ONE provider parenthetical per disambiguated label
- test_shared_reference_pre_fix_demonstrates_corruption: documents the
broken state that motivated the fix
Full suite: 3925 → 3928 passing (+3 new, 0 regressions).
Co-authored-by: lost9999 <56498264+lost9999@users.noreply.github.com>
The _norm_model_id function was using split(':', 1)[1] which only removed
the first colon-separated segment, leaving provider names in the normalized
model ID. For example, '@custom:jingdong:GLM-5' became 'jingdong:glm.5'
instead of 'glm.5'.
This caused the default model injection check to fail, resulting in a
duplicate 'Default' group being added to the model list even when the
model already existed with a provider prefix.
Changes:
- Use split(':')[-1] to get the last segment after all colons
- Use split('/')[-1] consistently for slash-separated paths
- Replace local _norm lambda with _norm_model_id function call
Fixes duplicate Default group appearing in model dropdown when using
custom providers with @provider:model ID format.
