fix(moa): scope the per-slot runtime cache to the active profile - #78185
Open
pierrenode wants to merge 1 commit into
Open
fix(moa): scope the per-slot runtime cache to the active profile#78185pierrenode wants to merge 1 commit into
pierrenode wants to merge 1 commit into
Conversation
_slot_runtime()'s runtime cache (added in NousResearch#77868/NousResearch#66811 to cut MoA cold-start latency) keys resolved (api_key, base_url, api_mode) only on (provider, model). A multiplex gateway resolves multiple profiles' agents in the same process; a MoA slot referencing the same (provider, model) pair in two profiles with different real credentials for that provider name gets the first profile's resolved runtime silently served to the second for up to _RUNTIME_CACHE_TTL_SECONDS (300s). Add hermes_home to the cache key so each profile resolves and caches its own runtime. Existing cold-start cache tests updated for the new key shape; new cross-profile regression test in tests/test_profile_isolation_runtime.py, alongside the existing profile-isolation suite for this exact bug class.
alt-glitch
added
type/bug
spfcraze
added a commit
to spfcraze/hermes-agent
that referenced
this pull request
Aug 9, 2026
The per-message memo in _memoized_resolve_runtime() keyed on config/auth file (mtime_ns, size) signatures only. A multiplex gateway resolves multiple profiles' agents in the same OS process (the desktop tui_gateway switches profiles per request via set_hermes_home_override), and 'hermes profile create --clone-all' copies the profile tree with mtime-preserving shutil.copy2 — so a cloned profile's config.yaml/auth.json carry the source's identical (mtime_ns, size) and compute the same memo signature, letting one profile receive the other's cached api_key/base_url for up to the 300s TTL. Add str(get_hermes_home()) to the memo key, the same profile-boundary fix open PR NousResearch#78185 applies to agent/moa_loop.py's sibling _runtime_cache. Tests: new tests/gateway/test_runtime_resolve_fastpath_profile_isolation.py drives _memoized_resolve_runtime under two profile overrides and asserts each profile gets its own credentials; both tests fail without the fix and pass with it. Same-profile memo hit preserved (2 invocations, 1 resolve).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
_slot_runtime()'s per-slot runtime cache (added in #77868, salvage of #66811, to cut MoA cold-start latency) keys the resolved(api_key, base_url, api_mode)runtime only on(provider, model). A multiplex gateway resolves multiple profiles' agents in the same OS process (see #77865, #77872 for the same anti-pattern in other modules). If two profiles both use MoA with a slot referencing the same(provider, model)pair but have different real credentials configured for that provider name, the second profile silently receives the first profile's resolvedapi_key/base_urlfor up to_RUNTIME_CACHE_TTL_SECONDS(300s).Reproduced empirically before writing the fix: resolving a slot under profile A's
HERMES_HOMEoverride, then under profile B's, returned profile A's cached credentials to profile B.Fix
Add the active
hermes_hometo the cache key ((hermes_home, provider, model)instead of(provider, model)), so each profile resolves and caches its own runtime independently._preset_cache(the sibling cache in the same module) was checked and does not have this issue — its key already includes the config file'sst_mtime_ns, and different profiles have different config file paths, so it self-invalidates correctly across profiles (at the cost of some cache-thrashing under interleaved multi-profile access, which is a missed optimization, not a correctness bug — left untouched to keep this fix minimal).Test plan
tests/agent/test_moa_cold_start_cache_66793.py) for the new 3-tuple cache key shape.TestMoASlotRuntimeCacheIsolationtotests/test_profile_isolation_runtime.py(the existing dedicated home for this exact bug class), reproducing the cross-profile leak against two realHERMES_HOMEoverrides.*moa*+ profile-isolation test files, 106 tests) passes.ruff checkclean on all changed files.