Fable's one recommended fast-follow: the streaming glue that computes
_explicitly_picked from the persisted signature (streaming.py:7905-7931) had no
direct coverage. Adds test_streaming_explicitly_picked_computation_5979 mirroring
that exact comparison across deliberate-pick (honored), stale-after-model-change
(ignored), stale-after-provider-change (ignored), and never-picked (unmarked).
Both advisors passed on the signature mechanism: Codex SAFE TO SHIP (session-update
hole verified closed: update->reload->cold openai/gpt-5.4 strips to gpt-5.4;
b3nw's deliberate pick preserves), Fable 'ship it' (clean data-model invalidation,
correct layering, no #433/#3872/#548 regression). Independent 10-scenario matrix
through the real resolver all correct.
Codex found a verified lifecycle hole: the bare boolean model_explicitly_picked
was not cleared on /api/session/update (api/routes.py:14438-14460 changes+persists
the model/provider without touching it), so after an update+reload the flag stayed
True and a cold #433 stale leftover (openai/gpt-5.4) was wrongly preserved instead
of stripped to gpt-5.4.
Fix (Codex's prescription): replace the unscoped boolean with a SIGNATURE bound to
the routing context.
- api/models.py: new model_explicit_pick_signature(model, provider) ->
'<model>\x1f<provider>'; Session.model_explicit_pick_signature persisted field
(replaces the boolean; in METADATA_FIELDS, restored via kwargs, survives restart).
- api/routes.py /api/chat/start: on a fresh explicit_model_pick, stamp the signature
of the resolved model+provider.
- api/streaming.py: treat the send as a deliberate pick ONLY when the persisted
signature equals the signature recomputed from the session's CURRENT persisted
model+provider. Any model/provider change (chat-start, session-update,
normalization, provider repair) yields a different signature and auto-invalidates
the stale pick — no per-mutation-site clearing needed.
So: deliberate pick on unchanged context -> preserve (b3nw); any switch away ->
signature mismatch -> unmarked -> legacy strip (#433 stays correct even after a
session-update+cold-restart). Adds signature match/invalidate/persist tests.
75 focused tests pass (resolver + sprint40 + the 2 streaming-mock suites).
Nathan's call: adopt Codex's complete mechanism instead of the unconditional
cold-preserve. The cold-catalog custom-proxy decision is now gated on whether the
user DELIBERATELY selected the model this session, closing the residual where a
stale #433 leftover would preserve+400 on a cold+no-disk send.
Mechanism:
- resolve_model_provider(model_id, *, explicitly_picked=False): in the custom-arm
cold branch (no warm provenance, not config-declared), explicitly_picked=True
preserves the vendor namespace verbatim (deliberate pick, proxy routes on it);
otherwise the legacy redundant-prefix strip runs (stale cross-provider leftover,
#433). Warm endpoint-advertised provenance still wins over the flag.
- Session.model_explicitly_picked: new persisted field (default False, restored
from metadata via **kwargs, listed in METADATA_FIELDS so it survives a restart
— b3nw's exact cold scenario).
- /api/chat/start stamps it: True on a fresh explicit_model_pick; cleared only
when the resolved model actually CHANGED without a pick (real switch away);
preserved across plain follow-up sends of the same model (the onchange marker
is one-shot, so per-send explicit_model_pick alone would drop the intent).
- streaming.py reads s.model_explicitly_picked and passes explicitly_picked= into
resolve_model_provider (inside the profile scope).
Independent end-to-end matrix (all pass): ben default cold, ben picked non-default
cold+warm, ben stale cold (strips), #433 warm+cold, #3872 bedrock, #548 zai-org.
67 resolver+sprint40 tests pass incl. new explicit-pick + stale-companion +
persistence round-trip.
Codex finding: the warm helper's fast no-op (if _models_cache_provenance is not
None: return) ignored WHICH profile the resident provenance belongs to. The
catalog globals are process-wide, so profile B's resident provenance could block
profile A from loading A's own valid disk cache — A would then resolve against
B's advertised ids (verified: profile A provenance resident + profile B bare-only
disk cache -> A's openai/gpt-5.4 wrongly preserved).
Fix: gate BOTH the fast no-op and the post-lock recheck on a current-profile
fingerprint match (_provenance_is_current: prov[1] == _models_cache_source_fingerprint()).
On mismatch the helper falls through, loads THIS profile's disk snapshot, and
republishes provenance stamped with the current fingerprint.
Tests: rewrote noop-when-warm to use the CURRENT runtime fingerprint (genuinely
hits the fast path) and added ignores-foreign-profile-resident (foreign fp -> falls
through -> loads this profile's disk snapshot). 66 resolver+sprint40 tests pass.
NOTE: Codex also re-raised the cold+no-disk #433 case (a stale openai/gpt-5.4 on a
bare-only relay preserves instead of strips when no catalog exists at all). Fable
weighed this exact residual and judged cold-preserve strictly better than master
(loud-fail-and-heal vs silent-recurring-truncation). The two advisors diverge;
surfacing to Nathan rather than auto-adopting the heavier persisted-explicit-pick
mechanism.
Codex gate found two real issues in the send-path warm:
1. Blocking lock wait (CORE): get_available_models(prefer_cache=True) still
acquires _available_models_cache_lock and can wait up to ~60s (unbounded in
synchronous rebuild mode) on an in-flight rebuild — unacceptable on the send
hot path. Rewrote warm_models_catalog_provenance_if_cold as a genuinely
non-blocking, disk-only publish: try the cache lock NON-BLOCKING (return
immediately if busy — a concurrent build will publish provenance itself),
read ONLY the on-disk cache via _load_models_cache_from_disk (no network, no
rebuild), publish snapshot+fingerprint + _sync_models_cache_provenance().
Verified: returns in ~3.5ms while the lock is held (was a 60s block).
2. Profile scope (CORE): the streaming worker is a separate thread that does NOT
inherit the HTTP handler's request-profile TLS. Both the warm and the resolve
read profile-keyed config (cache path + source fingerprint via
get_active_profile_name), so a cold send from a NAMED profile could resolve
against the DEFAULT profile's config and route to the wrong provider/base_url
(get_available_models reloads config on path mismatch). Wrapped warm + resolve
in profile_scope_for_detached_worker(_resolved_profile_name) so both see the
owning session's profile (no-op for default/root).
Tests: rewrote never-live-rebuilds (now asserts warm NEVER calls
get_available_models — reads disk directly) and added a non-blocking regression
(warm returns <2s while the cache lock is held). 65 resolver+sprint40 tests pass;
b3nw end-to-end + non-blocking + disk-warm all verified.
The earlier #5979 fix (PR #5980, exp-v0.52.53) resolved the WARM-catalog and
config-declared cases, but b3nw confirmed truncation persists on exp-v0.52.55/56.
Root cause (reproduced on master, confirmed by Codex + Fable):
b3nw's config is provider: custom:llm-proxy, model.default: x-ai/grok-4.5, no
models[] allowlist. His DEFAULT (x-ai/grok-4.5) preserves via config-declaration,
but a model he SELECTS in the picker that isn't the default (x-ai/grok-composer-2.5-fast)
truncates on every COLD send. The send path (api/streaming.py:7857) calls
resolve_model_provider without warming the catalog, and the #1855 chat/start fast
path deliberately skips the catalog when a session already has a persisted
model+provider -- so 'cold at send' is the designed state after any restart /
settings-save / cache-expiry, not a rare race. With provenance None and no config
declaration, the #5980 tri-state fell through to the LEGACY family strip
(if prefix in _PROVIDER_MODELS and _is_first_party_model: return bare), and
grok-composer-2.5-fast had graduated into the x-ai first-party catalog -> stripped.
Fix (two changes, both advisor-verified):
1. Flip the custom-arm cold default from family-strip to PRESERVE-verbatim
(api/config.py). For an explicit custom / custom:<slug> proxy the strip now
requires POSITIVE provenance (endpoint advertises only the bare id, #433). A
wrong strip destroys routing info the user can't recover; a wrong preserve
fails loudly and self-heals on warm. This also honours _endpoint_advertised_model_ids's
own documented 'None -> preserve verbatim' contract and removes the data-driven
flaw where a model graduating into the static catalog silently flips routing.
2. Network-free provenance warm on the send path
(warm_models_catalog_provenance_if_cold, called in streaming.py before the
resolve). prefer_cache=True republishes provenance from the durable disk cache
(no live probe, no _cfg_lock held on the worker thread -> no lock-order
deadlock), so #433's bare-only-strip stays exact whenever a disk cache exists.
Residual (accepted, strictly better than master): a #433-shaped wrong-preserve
now requires memory cold AND disk cache absent/invalid (first run, post-config-save
window, schema/version bump). It fails loudly and heals on the next /api/models or
session-visit rebuild. Under master b3nw's failure was the mirror image but silent,
recurring every send, and unfixable short of declaring every model in config.
Supersedes rodboev's #5996, which targeted the providers.<slug> arm and does not
cover b3nw's actual custom:llm-proxy config (confirmed by Codex).
Tests: rewrite the cold-fallback test (strip -> preserve), sharpen the
foreign-profile test so trusted-strip vs ignored-preserve genuinely diverge, add
b3nw's exact cold non-declared custom:slug case, and add three warm-helper tests
(publishes-from-disk, noop-when-warm, never-live-rebuilds). 64 resolver+sprint40
tests pass; b3nw end-to-end verified; #433 positive-provenance strip verified.
Thanks @b3nw for the CLI-vs-WebUI comparison that localized it.
CI's hermes-agent catalog version may not contain grok-4.5, so asserting the
live _PROVIDER_MODELS['x-ai'] already lists it failed on CI (shard 4, all py
versions). Reproduce the data-driven first-party trigger deterministically by
stubbing grok-4.5 into the x-ai catalog for the test (restored in finally),
regardless of the agent version CI runs against. The other #5979 tests preserve
x-ai/grok-4.5 via the advertised-full / cold-unknown-prefix paths, which don't
depend on x-ai being a known first-party prefix.
The round-2 fix acquired _available_models_cache_lock in the resolver to read
the snapshot+fingerprint consistently. Codex found that introduces a DEADLOCK:
config-save takes _cfg_lock -> _available_models_cache_lock while catalog-refresh
takes _available_models_cache_lock -> _cfg_lock (lock-order inversion), and it
also makes every send wait behind a catalog rebuild.
Fix: publish an immutable (snapshot, publisher_fingerprint) tuple atomically via
_sync_models_cache_provenance() at every cache publish/invalidate site (9 sites),
and have _endpoint_advertised_model_ids read that single global with ONE lock-free
load. Reading one tuple can't tear (GIL), acquires no lock (no ordering edge, no
deadlock), and never blocks behind a rebuild. A reader racing a publish sees the
previous consistent pair, never a torn snapshot/fingerprint.
Verified: two-thread probe — resolver completes in 0.0001s while another thread
holds _available_models_cache_lock (was the deadlock). Real catalog build publishes
the provenance tuple and closes the P0 end-to-end. Adds a deadlock-regression test.
Two more real issues from the second Codex gate:
1. Atomic provenance read: _endpoint_advertised_model_ids read the global
snapshot and its source fingerprint as two unlocked steps, while publishers
assign the new snapshot BEFORE computing the fingerprint (streaming.py
publish sites). In that window profile B's freshly-published catalog could
validate against profile A's still-current fingerprint and strip against the
wrong catalog. Fix: acquire _available_models_cache_lock (RLock) across the
snapshot read + fingerprint check + memo build, so the pair is always
consistent. The lock is held only for the two-assignment publish (the network
rebuild runs before it), so the per-send acquisition never blocks on a probe.
2. Tri-state cold fallback (was: unconditional cold->preserve): an unconditional
verbatim fallback newly broke the #433 relay cold path (openai/gpt-5.4 sent
to a relay that only accepts gpt-5.4; master stripped it). Now:
(1) config declares the full id verbatim (model.default / model.models /
custom_providers[].models) -> preserve (network-free, survives a cold
restart -> #5979 works cold via b3nw's model.default)
(2) endpoint catalog advertised it -> full=preserve / bare-only=strip
(3) genuinely unknown (cold + not declared) -> LEGACY family heuristic, so
the edge is never worse than pre-fix: first-party redundant prefix
still strips (#433 cold), intrinsic/unknown preserved (#3872/#548).
The #5979 active-user path can't reach (3): a selected id is either
config-declared or in the catalog the dropdown was built from.
New _model_id_declared_in_config() provides the config-provenance signal.
Tests updated: cold-fallback-to-legacy, config-declared-cold-preserve,
foreign-profile-catalog-ignored (proves untrusted snapshot is not used).
Codex regression gate found two real issues on the first cut:
1. extra_models bucket: a provider's advertised ids can be split across the
catalog's 'models' (visible) and 'extra_models' (picker overflow) buckets.
Reading only 'models' could miss a bare id that overflowed, leaving the
#433 stale prefix unstripped. _endpoint_advertised_model_ids now unions
both buckets.
2. Profile isolation (profiles are islands): _available_models_cache is a
process global, so a concurrently-active foreign profile could publish the
snapshot a different profile then reads for provenance — potentially
stripping an advertised full id against another profile's catalog. The
accessor now rejects the snapshot unless its published source fingerprint
(whose config_yaml axis is the profile-specific config path) matches the
current runtime fingerprint; any mismatch fails safe to verbatim preserve.
Adds tests: extra_models-counts-as-advertised (#433 via overflow bucket) and
foreign-profile-catalog-preserves-verbatim (isolation fail-safe). Test helpers
now stamp the source fingerprint exactly as the real publish sites do.
Resolve the model id for custom/proxy providers by ENDPOINT PROVENANCE
instead of a catalog-family guess. The old gate stripped a vendor prefix
whenever the bare id was first-party of the prefix's home vendor, which
could not distinguish a namespace the proxy actually routes on
(x-ai/grok-4.5 -> keep) from a stale cross-provider leftover on a relay
that only serves bare ids (openai/gpt-5.4 -> strip). A model graduating
into the first-party catalog (agent commit adding grok-4.5) silently
flipped a working custom-proxy id from preserved to stripped -> HTTP 400.
New _endpoint_advertised_model_ids() reads the already-published catalog
snapshot (never builds/probes/touches disk; O(1) per send via a
snapshot-identity memo) and answers what the endpoint advertised:
- full vendor/model advertised -> preserve (#5979, #3872, #548)
- only the bare id advertised -> strip the redundant prefix (#433)
- cold/unbuilt catalog -> preserve verbatim (fail-safe toward
the id the user actively selected)
The non-custom first-party-proxy strip arm is unchanged.
Rewrites the two pinned #433 tests to model provenance (bare-only
advertised) and adds #5979 coverage: advertised-full preserve (bare +
named custom), cold-catalog preserve-verbatim, and both-advertised
prefers-full.
Thanks @b3nw for the report and precise root-cause analysis.
Codex 5th CORE: the ownership set only covered static _PROVIDER_MODELS, not the
active provider's own config.yaml providers.<active>.models allowlist. So an
active provider defined purely via providers: (no static catalog) could be
hijacked by another slug's overlapping entry earlier in config order. Fix folds
the active provider's own providers: allowlist (list + dict shapes, copilot
excluded) into _provider_models_set — completing the ownership model so the
active provider owns its declared models from EITHER the static catalog OR its
own providers: entry. Regression test (fail-without-fix verified). 40 model_resolver
tests green.
Co-authored-by: akay64 <akay64@users.noreply.github.com>
Codex 4th CORE (root latent bug in the shared canonicalizer): _canonicalise_provider_id
only accepted an alias resolution if the target was in _PROVIDER_DISPLAY. But some
canonical targets (e.g. 'gemini') are indexed in _PROVIDER_MODELS and NOT
_PROVIDER_DISPLAY, so 'google-gemini'->'gemini' was rejected and returned raw
'google-gemini' — leaving the ownership guard empty and letting providers.openai.models
hijack an active Gemini-owned model. Fix mirrors the direct-hit check (line 1445):
accept alias targets in EITHER _PROVIDER_DISPLAY OR _PROVIDER_MODELS. This is the
root fix for the whole alias-canonicalization class (gemini/google-gemini/
google-ai-studio etc.), not just the reported site. Blast-radius verified: 81
canonicalise/provider-alias tests across 5 files still green; x-ai still stays x-ai,
opencode_go still folds. Gemini regression test added (fail-without-fix verified).
Co-authored-by: akay64 <akay64@users.noreply.github.com>
Codex 3rd CORE (pre-existing latent bug the providers: scan exposes): the
_provider_models_set ownership lookup used the RAW config_provider (e.g. 'z-ai')
but _PROVIDER_MODELS is keyed by canonical slug ('zai'), so an aliased active
provider got an empty ownership set -> _skip_custom_providers guard-2 silently
failed -> another providers.<slug>.models entry could hijack an active-owned
model (verified: active z-ai + glm-5 + providers.openai.models.glm-5 -> openai).
Fix: canonicalise config_provider before the _PROVIDER_MODELS lookup (and reuse
the canonical slug for the scan's ownership guard). Regression test with an
aliased active provider (verified fail-without-fix).
This is the root-cause fix for the alias class — the ownership guard now holds
for every aliased provider (zai/kilo/etc.), not just the reported site.
Co-authored-by: akay64 <akay64@users.noreply.github.com>
Codex 2nd CORE: the new providers: scan didn't honor the _skip_custom_providers
active/default ownership guard (config.py:2535) that the custom_providers scan
uses. So with an active provider (e.g. ai-gateway) + model.default=gpt-5 AND a
providers.openai.models.gpt-5 entry, the scan hijacked gpt-5 to openai instead of
the active ai-gateway path. Fix: when _skip_custom_providers is true, the
providers: scan only considers the active provider's own canonical slug (skip all
other slugs). Regression tests: active-provider-owns-default not hijacked (fails
without fix); active provider's OWN providers: entry still matches.
Co-authored-by: akay64 <akay64@users.noreply.github.com>
Gate-cert (nesquena-hermes) + Codex found 1 CORE: akay64's new providers: scan in
resolve_model_provider() iterated ALL providers.<slug>.models INCLUDING
providers.copilot.models — but that key is a documented per-model SETTINGS map
(reasoning_effort/limits), NOT a routing allowlist (config.py exception). So a
Copilot per-model settings entry hijacked that model's routing to Copilot
(Codex reproduced gpt-5 -> copilot). Fix (Codex-exact): skip copilot in the scan.
Added the regression tests the PR shipped 0 of: user-defined allowlist routes;
active-openai + providers.copilot.models (dict + list shapes) stays openai
(both verified to fail without the exclusion).
Co-authored-by: akay64 <akay64@users.noreply.github.com>
* fix: keep custom provider when model id has known-provider prefix (#4210)
Siblings of #3872 (v0.51.349, custom provider with a configured base_url
preserves intrinsic routing prefixes). The base_url unset variant of
the same bug was missed: with provider=custom or custom:<slug> and no
base_url set, a model id like 'x-ai/grok-2' or 'google/gemma-2' was
hijacked to provider='openrouter', base_url=None. The backend then
failed to initialize with 'RuntimeError: No LLM provider configured'
because the user had no OpenRouter credentials configured.
resolve_model_provider() now applies the same _is_custom exemption the
config_base_url branch already carries to the cross-provider redirect
at the end of the slash branch. The custom provider keeps the request
and the full model id; real first-party providers pointed at an
OpenAI-compatible proxy (e.g. provider=openai + base_url=litellm) are
unaffected because their strip path is in the base_url branch and the
'prefix == config_provider' branch above the change.
Verified:
- test_model_resolver.py 32 passed (30 existing + 2 new RED tests)
- test_custom_provider_bare_model_reasoning.py 34 passed
- Pre-existing F401 (api/config.py:23,24,28) are not introduced by
this change; same 3 errors exist on upstream/master.
Refs #4210
* chore(release): reposition #4210 CHANGELOG entry under v0.51.426 (OM)
---------
Co-authored-by: b3nw <b3nw@users.noreply.github.com>
Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
A bare 'custom' provider with a remote base_url is a vendor-routing proxy
(LiteLLM, Bedrock gateway, etc.). A slashed model id like bedrock/opus-4-6 is
intrinsic — the proxy routes on the full string. The prior heuristic stripped
ANY known-provider prefix (bedrock is also a provider name), truncating it to
opus-4-6 and causing 403 'model not allowed'.
Fix: in the custom-base_url branch, strip the prefix only when (a) the configured
provider is a real first-party provider pointed at an OpenAI-compatible proxy
(provider=openai + proxy base_url — the #433 path), OR (b) for a bare 'custom'
provider, the bare id is genuinely a first-party model of that prefix's catalog
(openai/gpt-5.4 -> gpt-5.4, since gpt-5.4 IS an OpenAI model). An intrinsic
routing prefix whose bare id is NOT first-party of that namespace is preserved
(bedrock/opus-4-6 stays whole). Unknown prefixes (zai-org/...) preserved as before.
Both prior behaviors work in tandem with the fix:
- #433 sprint40: bare custom + openai/gpt-5.4 -> gpt-5.4 (redundant, strip)
- #433 model_resolver: provider=openai + google/gemma -> gemma (proxy strip)
- #548: custom + zai-org/GLM-5.1 -> preserved (unknown prefix)
- #1625: custom + loopback base_url -> preserved (local server)
- #3872: custom + bedrock/opus-4-6 -> preserved (intrinsic vendor prefix)
Adds regression tests covering all of the above.
Co-authored-by: nesquena-hermes <nesquena-hermes@users.noreply.github.com>
Co-authored-by: haolf000 <haolf000@users.noreply.github.com>
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.
When multiple providers expose the same bare model ID (e.g. two custom
providers both listing gpt-5.4), the model picker cannot distinguish
them — both rows appear active and clicking the other provider's copy
is a no-op.
Fix:
- Add _deduplicate_model_ids() post-process in api/config.py that
detects duplicate bare model IDs across groups and prefixes
collisions with @provider_id: so each entry is globally unique
- Update norm() regex in static/ui.js to strip @provider: prefixes
for fuzzy matching, so existing sessions with bare model IDs still
restore correctly
- First occurrence stays bare for backward compatibility with sessions
that already store the bare model name
- Update test_model_resolver to be dedup-aware
Closes#1228
Root cause of persistent CI failure on `test_custom_endpoint_uses_model_config_api_key_for_model_discovery`:
The helper `_available_models_with_full_cfg` and the failing test itself both do:
```python
try:
_cfg._cfg_mtime = stat().st_mtime
except Exception:
pass # ← leaves _cfg_mtime at stale value from a prior test
```
On CI (no `~/.hermes/config.yaml`), `stat()` raises `OSError`. `get_available_models()` then sees `_current_mtime=0.0 != _cfg_mtime=<stale from prior test>`, calls `reload_config()`, overwrites the test's in-memory `cfg`, `api_key` disappears, `urlopen` is never called, `captured['auth']` raises `KeyError`.
**Fix:** The `except` clause now sets `_cfg._cfg_mtime = 0.0` (matching what `get_available_models()` will see when `stat()` also fails), so the reload guard becomes a no-op and the test's `cfg` mutation survives. Same pattern already used correctly in `test_custom_provider_display_name.py` and `test_ttl_cache.py`.
Verified with `HERMES_CONFIG_PATH=/tmp/nonexistent` to reproduce the CI no-config condition locally. 1747/1747 full suite.
Fixes the CI failure introduced by #817: test_model_resolver::test_custom_endpoint_uses_model_config_api_key_for_model_discovery was failing with KeyError: 'auth' due to the 60s TTL cache in get_available_models() being populated by test_byok_model_dropdown.py tests that ran earlier. Added autouse _isolate_models_cache fixture to 5 test files. Full suite 1747/1747, QA harness green.
* fix: custom provider with slash model name no longer rerouted to OpenRouter (#255)
When base_url is configured in config.yaml, resolve_model_provider() now
trusts the configured provider/base_url entirely and skips the slash-based
OpenRouter heuristic. Fixes google/gemma-4-26b-a4b with provider:custom
being silently routed to OpenRouter, resulting in 401 errors.
Fixes#230
* test: mobile layout regression suite — 14 tests for every QA run (#254)
Adds tests/test_mobile_layout.py with 14 static regression tests that run
on every QA pass to catch mobile layout breakage before it reaches prod.
Covers: breakpoints at 900px/640px, right panel slide-over CSS, mobile
overlay, bottom nav, files button, profile dropdown z-index, chip overflow,
workspace close, 100dvh, 44px touch targets, 16px font-size on textarea.
* feat: /skills slash command lists and filters available Hermes skills (#257)
Adds /skills [query] command to commands.js. Fetches from /api/skills,
groups by category (alphabetically sorted), displays as a formatted
assistant message. Optional query filters by name, description, or category.
i18n keys added for en, de, zh, zh-Hant. 1 regression test added.
Fixes#248
* feat: shared app dialogs replace native confirm()/prompt() calls (#251)
Adds showConfirmDialog() and showPromptDialog() helpers to ui.js, backed
by a themed #appDialogOverlay. Replaces all 11 native browser confirm/prompt
call sites across panels.js, sessions.js, ui.js, workspace.js.
Supports: danger mode, keyboard focus trap (Tab/Escape/Enter), focus restore,
ARIA roles, mobile-responsive stacked buttons at 640px. i18n for en/de/zh/zh-Hant.
5 new tests in test_sprint33.py verify markup, CSS, helpers, and absence of
native dialog calls.
Extracted from PR #242.
* fix: Android Chrome mobile — workspace panel close + profile dropdown (#256)
Fix#247: toggleMobileFiles() now shows/hides the mobile overlay when
toggling the right workspace panel. New closeMobileFiles() helper closes
the panel with correct overlay state tracking. Overlay onclick calls both
closeMobileSidebar() and closeMobileFiles(). Mobile-only close button (x)
added to workspace panel header.
Fix#246: profile dropdown uses position:fixed;top:56px;right:8px at
max-width:900px, escaping the overflow-x:auto stacking context that was
clipping it on Android Chrome.
Fix applied during review: closeMobileSidebar() now checks if the right
panel is still open before hiding the overlay, preventing the overlay from
disappearing when only the sidebar is closed.
Fixes#247Fixes#246
* feat: session ⋯ action dropdown replaces per-row buttons (#252)
Replaces the 5 per-row hover action buttons (pin/move/archive/duplicate/trash)
with a single ⋯ trigger that opens a positioned dropdown menu. Menu has full
keyboard (Escape), click-outside, scroll, and resize-reposition handling.
Position:fixed prevents sidebar clipping.
5 actions: Pin/Unpin, Move to project, Archive/Unarchive, Duplicate, Delete
(danger style). Each with icon and descriptive subtitle.
Updated test_sprint16.py: test_sessions_js_uses_action_menu_not_per_row_buttons
asserts the new trigger and menu functions exist, old per-row classes are gone.
Extracted from PR #242.
* docs: v0.47.0 release notes, bump version, update test counts (645)
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: decode HTML entities before markdown processing + zh/zh-Hant translations (#239)
Adds decode() helper in renderMd() to fix double-escaping of HTML entities
from LLM output (e.g. <code> becoming &lt;code&gt; instead
of rendering). XSS-safe: decode runs before esc(), only 5 entity patterns.
Also adds 40+ missing zh (Simplified Chinese) translation keys and a new
zh-Hant (Traditional Chinese) locale with 163 keys.
Fix applied: removed duplicate settings_label_notifications key in both
zh and zh-Hant locales.
Fixes#240
* fix: restore custom model list discovery with config api key (#238)
get_available_models() now reads api_key from config.yaml before env vars:
1. model.api_key
2. providers.<active>.api_key / providers.custom.api_key
3. env var fallbacks (HERMES_API_KEY, OPENAI_API_KEY, etc.)
Also adds OpenAI/Python User-Agent header and a regression test covering
authenticated /v1/models discovery.
Fixes users with LM Studio / Ollama custom endpoints configured in
config.yaml whose model picker silently collapsed to the default model.
* feat: Docker UID/GID matching to avoid root-owned .hermes files (#237)
Adds docker_init.bash with hermeswebuitoo/hermeswebui user pattern so
container files match the host user UID/GID. Prevents .hermes volume
mounts from being owned by root when using a non-root host user.
Configure via WANTED_UID and WANTED_GID env vars (default 1000/1000).
Readme updated with setup instructions.
Fix applied: removed duplicate WANTED_GID=1000 line in docker-compose.yml
that was overriding the ${GID:-1000} variable expansion.
* security: redact credentials from API responses and fix credential file permissions (#243)
Adds response-layer credential redaction to three endpoints:
- GET /api/session — messages[], tool_calls[], and title
- GET /api/session/export — download also redacted
- SSE done event — session payload in stream
- GET /api/memory — MEMORY.md and USER.md content
Adds api/startup.py with fix_credential_permissions() at server startup.
Adds 13 tests in tests/test_security_redaction.py.
Merged with #237 container detection changes in server.py.
* fix: cancel button now interrupts agent and cleans up UI state (#244)
Wires agent.interrupt() into cancel_stream() so the backend actually
stops tool execution when the user clicks Cancel, rather than only
stopping the SSE stream while the agent keeps running.
Changes:
- api/config.py: adds AGENT_INSTANCES dict (stream_id -> AIAgent)
- api/streaming.py: stores agent in AGENT_INSTANCES after creation,
checks CANCEL_FLAGS immediately after store (race condition fix),
calls agent.interrupt() in cancel_stream(), cleans up in finally block
- static/boot.js: removes stale setStatus(cancelling) call
- static/messages.js: setBusy(false)/setStatus('') unconditionally on cancel
Race condition fix: after storing agent in AGENT_INSTANCES, immediately
checks if CANCEL_FLAGS[stream_id] is already set (cancel arrived during
agent init) and interrupts before starting. Check is inside the same
STREAMS_LOCK acquisition, making it atomic.
New test file: tests/test_cancel_interrupt.py with 6 unit tests.
* docs: v0.46.0 release notes, bump version, update test counts
---------
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
* fix: do not build phantom "Custom" group when active provider is set
When model.provider is a real provider (e.g. openai-codex) and model.base_url
is configured, hermes_cli reports 'custom' as an authenticated provider. The
WebUI model picker was building a separate "Custom" group for it and parking
the configured default_model there instead of under the active provider's
group — diverging from the TUI which correctly shows the model under its
configured provider.
Two fixes in api/config.py get_available_models():
1. Discard 'custom' from detected_providers when active_provider is set and
isn't 'custom' itself. The base_url belongs to the active provider.
2. Replace the substring-based default-model injection check with an exact
match against _PROVIDER_DISPLAY. The old check `active_provider.lower() in
g.get('provider', '').lower()` silently failed for hyphenated IDs like
'openai-codex' vs display name 'OpenAI Codex' (hyphen vs. space),
falling through to groups[0] and landing the model in the alphabetical
first group instead.
Adds two regression tests in tests/test_model_resolver.py covering both
conditions.
* fix: do not build phantom Custom group when active provider is set
Two bugs in get_available_models():
1. Phantom Custom group: hermes_cli reports 'custom' as authenticated
whenever model.base_url is set. With provider=openai-codex + base_url,
detected_providers contained both 'openai-codex' and 'custom', producing
a duplicate group. Fixed by discarding 'custom' from detected_providers
when the active provider is any real named provider.
2. Hyphen/space mismatch in default_model injection: the substring check
'openai-codex' in 'openai codex' is False (hyphen vs space), causing the
default model to fall through to groups[0] (alphabetically first provider)
instead of the active provider group. Fixed by using _PROVIDER_DISPLAY
for exact display-name comparison.
Also fixes test helper _available_models_with_full_cfg to clear model env
vars during the call, preventing real hermes profile env from leaking into
the test assertions.
---------
Co-authored-by: mbac <marco.baciarello@gmail.com>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>
The previous fix (#142) prefixed non-default provider models with
'provider/model' which then hit the cross-provider guard and routed
to OpenRouter — worse than before for users without an OpenRouter key.
New approach: non-default provider models use '@provider:model' format
(e.g. @minimax:MiniMax-M2.7). resolve_model_provider() parses this
hint and returns (bare_model, provider, None). streaming.py and
routes.py then pass the resolved provider to
resolve_runtime_provider(requested=provider) which gets the correct
per-provider API key and base_url from hermes-agent.
This uses the agent's own credential resolution instead of reinventing
routing logic in the webui.
Fixes#138
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: prefix non-default provider model IDs for correct routing
When multiple providers are configured, models from non-default providers
(e.g. MiniMax when Anthropic is default) were sent as bare names without
provider context. resolve_model_provider() couldn't determine the target
provider and routed them to the default provider's API, which failed.
Fix: get_available_models() now prefixes model IDs with the provider name
(e.g. minimax/MiniMax-M2.7) for providers that are NOT the active config
provider. The default provider's models keep bare names for direct API
routing. This matches the existing pattern for OpenRouter models.
Added 2 tests to test_model_resolver.py for cross-provider routing.
Closes#138
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: model prefix — null-guard, case normalization, mutation safety, tests
Four fixes on top of original PR:
- active_provider=None guard: without a confirmed provider all models were
being prefixed. Only prefix when active_provider is set.
- Case normalisation: compare pid against active_provider.lower() so
config.yaml entries like 'Anthropic' match pid 'anthropic'.
- Mutation safety: default branch used raw reference to _PROVIDER_MODELS[pid];
the default_model injector later calls list.insert() on that reference,
permanently mutating the shared constant. Both branches now use a copy.
- Already-prefixed model IDs pass through as-is (no double-prefix).
Added 3 tests for get_available_models() prefix behaviour:
- Non-default provider models are prefixed
- Active provider's own entries remain bare
- No double-prefix when active_provider is absent
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When config has provider=openrouter and model=openrouter/free,
resolve_model_provider() stripped the 'openrouter/' prefix because
prefix == config_provider. This sent 'free' to OpenRouter's API,
which returned 404 (model not found).
OpenRouter always needs the full provider/model path (e.g.
openrouter/free, anthropic/claude-sonnet-4.6). The prefix-stripping
logic is only correct for direct-API providers.
Fix: skip prefix stripping entirely when config_provider is 'openrouter'.
Return the full model_id with provider='openrouter'.
Added 7 unit tests for resolve_model_provider() covering:
- openrouter/free keeps full path (the bug)
- openrouter cross-provider models keep full path
- direct API providers still strip prefix correctly
- cross-provider routing to openrouter
- bare model names use config provider
- empty model returns defaults
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
nesquena-hermes