fix(agent): throttle cross-turn fallback replay after non-retryable exhaustion (#24996) - #53909
Merged
Merged
Conversation
When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
Contributor
🔎 Lint report:
|
| Rule | Count |
|---|---|
invalid-argument-type |
1 |
First entries
tests/run_agent/test_24996_fallback_exhaustion_cooldown.py:38: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `dict[str, Any]`, found `Unknown | None`
✅ Fixed issues: none
Unchanged: 6064 pre-existing issues carried over.
Diagnostics are surfaced as warnings — this check never fails the build.
tonydwb
approved these changes
Jun 28, 2026
tonydwb
left a comment
tonydwb
left a comment
There was a problem hiding this comment.
Code Review Summary
Verdict: Approved
Clean fix for fallback chain exhaustion replay storm (#24996). Well-scoped change:
- Arms a 5-second cooldown when the fallback chain exhausts on non-rate-limit failures
- Uses existing
_rate_limited_untilgate to prevent cross-turn re-marshaling of the full context - Rate-limit/billing failures keep their existing 60s cooldown unaffected
- The
max()pattern ensures the cooldown never shrinks an existing window
4 dedicated regression tests covering: exhaustion cooldown, empty chain (no false gating), rate_limit preserves 60s, max cooldown never shrinks. Good documentation with clear issue reference.
No security or correctness concerns. Ready to merge.
Reviewed by Hermes Agent
teknium1
changed the title
alt-glitch
added
type/bug
pai-scaffolde
pushed a commit
to Scaffolde/hermes-agent
that referenced
this pull request
Jun 28, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
waefrebeorn
pushed a commit
to waefrebeorn/slermes
that referenced
this pull request
Jul 2, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
caozuohua
pushed a commit
to caozuohua/hermes-agent
that referenced
this pull request
Jul 2, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
Jasper6439
pushed a commit
to Jasper6439/hermes-agent
that referenced
this pull request
Jul 5, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
habarmc1223-sudo
pushed a commit
to habarmc1223-sudo/hermes-agent-fluxmem
that referenced
this pull request
Jul 8, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
santhreal
pushed a commit
to santhreal/hermes-agent
that referenced
this pull request
Jul 13, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
Gravezzz
pushed a commit
to Gravezzz/hermes-agent
that referenced
this pull request
Jul 21, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
leewenjie
pushed a commit
to leewenjie/hermes-agent
that referenced
this pull request
Aug 7, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
melon-xf
added a commit
to melon-xf/hermes-agent
that referenced
this pull request
Sep 3, 2026
…earch#24996) (NousResearch#53909) When every provider in the fallback chain fails non-retryably back-to-back (e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is already bounded — _fallback_index advances monotonically and the loop aborts when the chain exhausts. The damaging mode is cross-turn: restore_primary_ runtime resets _fallback_index=0 every turn, so a client that re-submits immediately replays the entire chain, re-marshaling the full (potentially 80k-token) context once per provider every turn with no throttle on the non-rate-limit path. On constrained hosts this exhausts memory/swap. Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until; the gap was the non-rate-limit case. Now, when the chain exhausts on a non- rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the same _rate_limited_until gate (max(), never shrinking an existing window). The next turn's restore stays gated and does NOT reset the index, so the chain isn't replayed until the cooldown clears. No new state, no thread sleep, no false-trip on legitimately long chains (those walk normally within a turn). Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
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
3 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
A fallback chain that exhausts on a non-rate-limit failure no longer replays the entire chain on the next turn, so a re-submitting client can't drive a constrained host into memory/swap exhaustion (#24996).
Root cause: the within-turn fallback walk is already bounded (
_fallback_indexadvances monotonically, the loop aborts when the chain exhausts). The damaging mode is cross-turn —restore_primary_runtimeresets_fallback_index = 0every turn, so a client that re-submits immediately replays the whole chain, re-marshaling the full (potentially 80k-token) context once per provider every turn. Rate-limit/billing failures already arm a 60s cooldown via_rate_limited_until; the non-retryable client-error path (HTTP 400 etc.) had no throttle.Changes
agent/chat_completion_helpers.py(try_activate_fallback): when the chain exhausts on a non-rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the existing_rate_limited_untilgate usingmax()(never shrinks an existing window). The next turn'srestore_primary_runtimestays gated and does not reset the index, so the chain isn't replayed until the cooldown clears. Rate-limit/billing keep their own 60s cooldown; an empty chain (no fallback configured) is never penalized.tests/run_agent/test_24996_fallback_exhaustion_cooldown.py: new regression coverage.No new state, no thread
sleep, no magic activation counter — and no false-trip on a legitimately long chain (long chains walk normally within a turn; only cross-turn replay is gated).Validation
max()preserves it)tests/run_agent/test_24996_fallback_exhaustion_cooldown.py+test_provider_fallback.py+test_primary_runtime_restore.py→ 57 passed, 0 failed.restore_primary_runtimereturns False and keeps_fallback_index == 2during cooldown → after cooldown clears, restore returns True and resets index to 0.Infographic