Root cause of the silent-crash-after-hours cluster (#4765/#2233/#4633): the
global in-memory SESSIONS LRU evicted with a blind popitem(last=False), which
could drop an actively streaming or not-yet-persisted session (data loss) and
was capped only via an env var. On long-running installs the effective result
was unbounded RAM growth until segfault.
- Add _session_is_evictable(): a session is evictable ONLY when it is not
streaming (no active_stream_id), has no in-flight turn (no pending_user_message
/ pending_started_at), and its full state is proven on disk (sidecar
message_count >= in-memory count; metadata-only stubs and zero-message shells
are trivially safe).
- Add _evict_sessions_over_cap(): replaces all 8 blind popitem loops across
models.py, routes.py, streaming.py. Walks the LRU oldest-first and removes only
provably-safe entries; never acquires LOCK/stream locks itself (caller holds
LOCK) so no lock-ordering deadlock. May briefly exceed the cap rather than ever
evict an active/unsaved session.
- Make the cap configurable via config.yaml webui.sessions_cache_max
(get_sessions_cache_max()); precedence config.yaml -> HERMES_WEBUI_SESSIONS_MAX
(legacy) -> DEFAULT_SESSIONS_CACHE_MAX=300. No new HERMES_* env var. Invalid or
<1 values fall back so a typo can never disable the bound.
- Evicted sessions lazily reload from their JSON sidecar via the existing
get_session() accessor; no call sites changed. _index.json sidebar behavior
unchanged (the sidebar reads the index, not SESSIONS).
- Add tests/test_issue4765_sessions_lru_eviction.py (8 tests): eviction past
cap, active/streaming never evicted, unsaved/stale-tail never evicted, lazy
reload with identical content, and no-data-loss under heavy churn.
- README: document the config.yaml key + safety semantics.
Fixes#4765.
Structural hardening per user request, following the 3rd round of the
same bypass class on PR #5313's reasoning_efforts feature:
Round 1: plain-ordering regression (provider-config short-circuit ran
before the nested-route deny).
Round 2: a provider-qualified hint (@custom:<slug>:vertex/gemini-...)
left a slug fragment that the deny's prefix-match missed.
Both were legitimate, narrowly-targeted fixes, but the pattern —
_nested_route_reasoning_denied() requiring the model string to START
WITH the route prefix — meant every future wrapper/nesting scheme
would need the strip logic to be updated in lockstep, and a missed
case fails OPEN (reasoning re-enabled on a route that must never show
it), which is the wrong failure mode for a security-adjacent guard.
This commit removes that class of bug instead of patching its latest
instance: _nested_route_reasoning_denied() now searches for the
vertex/gemini- or gemini_cli/gemini- pattern ANYWHERE in the string at
a non-alphanumeric boundary, rather than requiring it at position 0.
Correctness no longer depends on _strip_provider_hint_for_reasoning()
having stripped exactly the right prefix first — any number of opaque
wrapper layers (@provider:, a named custom-provider slug, or any
nesting scheme not yet invented) can precede the route and the deny
still fires.
Verified: all historical bypass strings (round 1 and round 2, plus a
hypothetical deeper double-wrapped case) now correctly deny; embedded
substrings that must NOT match (e.g. 'notvertex/gemini-x') correctly
don't, thanks to the boundary lookbehind.
Added test_nested_route_deny_is_boundary_based_not_prefix_based
locking in the structural invariant directly, independent of any
particular wrapper scheme.
Verification:
./scripts/test.sh tests/test_reasoning_effort_model_capabilities.py tests/test_custom_provider_bare_model_reasoning.py
# 54 passed in 1.31s
Full suite: 11257 passed, 64 pre-existing failures (identical file/test
set as the prior commit's baseline — network/credential/profile-isolation/fd-leak
tests, unrelated to reasoning_efforts).
Resolve gate certifier feedback on PR #5313 (deeper bypass, round 3):
A provider-qualified hint like `@custom:<slug>:vertex/gemini-image-1.0`
still bypassed `_nested_route_reasoning_denied()`. The old
`_strip_provider_hint_for_reasoning()` did a naive first-colon split,
which only removed the leading `@custom:` wrapper and left the named
provider's slug (e.g. `agg:vertex/gemini-image-1.0`) attached to the
model id. That leftover slug fragment no longer starts with
`vertex/gemini-`, so the nested-route deny missed it and a configured
`providers.<name>.reasoning_efforts` / `custom_providers[].reasoning_efforts`
allowlist re-enabled reasoning controls on Gemini image/embedding routes
that must never expose them.
Fix: `_strip_provider_hint_for_reasoning()` now accepts the resolved
provider id and strips the exact `@{provider}:` prefix first (e.g.
`@custom:agg:`), so both the wrapper and the slug are removed in one
pass before the nested-route deny check runs. Falls back to the
original first-colon split when no provider is supplied, preserving
existing behavior for plain `@provider:model` hints.
Verified in-process: both
`@custom:agg:vertex/gemini-image-1.0` and
`@custom:agg:vertex/gemini-embedding-001` now correctly resolve to []
instead of leaking the configured ["low", "high"].
Added regression test extending the existing nested-route-deny test to
cover provider-qualified hinted models.
Verification:
`./scripts/test.sh tests/test_reasoning_effort_model_capabilities.py tests/test_custom_provider_bare_model_reasoning.py`
-> 53 passed in 1.23s
Full suite: 11258 passed (63 pre-existing failures unrelated to this
change — network/credential/profile-isolation/fd-leak tests; confirmed
identical failure set on unmodified branch).
Resolve review feedback on PR #5313:
- Read reasoning_efforts from named custom_providers entries for custom:<name>
providers instead of looking only in providers:<name>.
- Preserve the nested-route deny ordering before the provider-config shortcut
so Gemini image/embedding routes cannot be re-enabled by config.
- Deduplicate configured effort values before returning them.
Add focused regression tests for bare provider config, named custom provider
config, all-invalid fallthrough, ACP hard guards, and nested route denies.
Addresses Greptile review feedback on PR #5313:
1. Move cursor-acp/copilot-acp guard before step 0 so a stray config
entry can't surface unsupported effort options for those providers.
2. Only short-circuit when the filtered list is non-empty; an all-invalid
list (e.g. typos) falls through to heuristics instead of hiding
reasoning support from the UI.
Add a config-driven path to resolve_model_reasoning_efforts() that reads
providers.<name>.reasoning_efforts from config.yaml. This lets users
explicitly list valid reasoning effort levels per provider, so the
WebUI dropdown only shows options the model actually supports.
Handles both custom:<name> and bare registered provider names.
Falls through to existing heuristics (Copilot per-model lists, LM Studio
live API, models.dev) when no config entry is present.
Reapply footer control ordering on current upstream/master while preserving the required situational chip renderer. Persist composer_control_order with backend validation, keep the settings descriptions reorder-aware, and make primary/situational chip renderers participate in same-group drag ordering.
Verified with: node --check static/boot.js; node --check static/panels.js; git diff --check; ./scripts/test.sh tests/test_issue4598_composer_control_visibility.py
Add a default-OFF 'project_quick_create_buttons' setting that gates the
per-project '+' quick-create buttons on sidebar project chips. The chip
render site only calls _attachProjectQuickCreateButton when the new
window._projectQuickCreate runtime flag is enabled, so the buttons do
not appear unless the user opts in via Settings.
Mirrors the large_text_paste_as_attachment opt-in plumbing (defaults,
allowlist, settings collect/restore/save, window hydration at both sites,
checkbox row, i18n across all locales).
Add a user setting 'large_text_paste_as_attachment' (default: on) that lets
users paste large text inline instead of having it auto-converted to .md file
attachments. The toggle lives in Control Center → Chat tab and persists via
settings.json, following the established upstream pattern for boolean settings.
Mirror the worker's LRU-eviction ACTIVE_RUNS liveness check in
_evict_session_agent so truncate/clear/model-switch can't close the SessionDB
out from under an in-flight turn on the same session (#5096 Bug D follow-up).
The cache handle still drops (harmless — worker holds a local ref); only the
lifecycle commit + _session_db.close() are skipped while a run is active.
Co-authored-by: claw-io <claw-io@users.noreply.github.com>
The streaming path already honored a max_tokens cap from the active profile's
config.yaml, but there was no UI to manage it. Settings > Preferences now has a
"Max output tokens" field that writes a root-level max_tokens override through
the profile config (not settings.json, which the streaming worker never reads
for this), preserving unrelated YAML keys under _cfg_lock. Blank clears the
override so agent.max_tokens fallback resumes; get_max_tokens_status() surfaces
root / effective / fallback so the field shows what a new turn would use.
Co-authored-by: rodboev <rodboev@users.noreply.github.com>
Two gate-found CORE/SILENT edges fixed on-branch:
- api/config.py: switching the main default to a different custom:* provider kept
the OLD provider's base_url, routing New Chat to the wrong endpoint. The
base_url-clear on provider-change now covers custom:* too (was exempted).
- static/sessions.js: newSession() skipped the configured default when S.session
was null but the picker held stale state (e.g. after deleting the last session),
inheriting the stale model. Default now wins over legacy picker regardless of
loaded-session state (explicit empty-composer override still takes priority).
- Updated 2 stale source-grep precedence assertions + widened a brittle 6000-char
newSession() slice window to 7000 (assignment block grew; intent unchanged).
- 2 new non-vacuous regression tests (custom base_url clear; precedence).
Co-authored-by: Rod Boev <rod.boev@gmail.com>
