✅ Done! Here's what I found across the repos and your environment. The error in the linked comment The comment (

all-hands-bot

at commit

2b97474214e8

) is the pr-review automation reporting

status: error

. That matches issue #4936 ("Hard quota exhaustion delays fallback LLM behind retry backoff"): the production incident returned

usage_limit_reached

for a team-plan model and retried it four times through exponential backoff before the review errored — instead of immediately failing over to a configured fallback. So the root cause is exactly what PR #4917 now fixes (skip retries on deterministic quota errors and go straight to

FallbackStrategy

). Does automation repo have a fallback PR? No LLM-fallback PR. What exists is single-profile selection only: • #120 "Support named LLM profiles in automations" (merged) — the

model

field on the automation +

AUTOMATION_MODEL

env var, resolved via

workspace.get_llm(profile_name=...)

. • #223 "capture active LLM profile from the correct users/me field" (merged). •

openhands/automation/utils/model_profiles.py

validates/resolves one profile name per automation. There is no

FallbackStrategy

/

fallback_llms

anywhere in the repo. Does extensions repo have a fallback PR? No LLM-fallback PR either. Related but not fallback: • #547 "show LLM provenance in automation output" (open) — resolves

AUTOMATION_MODEL

and prints

LLM profile · Model

. • #483 ships the reviewer script as a catalog bundle; #388 documents

AUTOMATION_MODEL

. • The pr-review plugin builds a single

LLM(**llm_config)

from

LLM_MODEL

with no

fallback_strategy

. Your pr-review automation In this environment's

automations.db

, the PR reviewer automations are single-profile with no fallback: •

GitHub PR Reviewer: OpenHands/software-agent-sdk for @all-hands-bot

→

gpt-5.6-sol

•

...OpenHands/OpenHands...

→

gpt-5.6-sol

•

...OpenHands/extensions...

→

None

(default) • multi-repo one →

deepseek-v4

So nothing is currently configured to fail over to another model. How to configure a fallback (e.g. to glm) At the SDK level it's a

FallbackStrategy

on the `LLM`:

from openhands.sdk.llm import LLM, FallbackStrategy

llm = LLM(
    model="gpt-5.6-sol",
    api_key=...,
    fallback_strategy=FallbackStrategy(
        fallback_llms=["glm"],
        profile_store_dir="~/.openhands/profiles",  # defaults to this if omitted
    ),
)

• Profiles are JSON files in

~/.openhands/profiles

(or a custom dir). A checked-in glm-5.2 example already exists in the SDK at

examples/01_standalone_sdk/37_llm_profile_store/profiles/glm.json

—

"model": "openhands/glm-5.2"

with

usage_id: "glm"

. Drop that JSON (with a real key) into your profiles dir as

glm.json

. •

fallback_llms

takes an ordered list of profile names, so

["glm"]

will resolve that profile lazily through

LLMProfileStore

on the first failure. Two caveats for your specific case: 1.

fallback_strategy

is

exclude=True

on

LLM

, so it is not persisted/serialized — it must be re-attached after loading a profile (relevant for the automation path, which loads LLMs by profile name rather than constructing them inline). 2. Encrypted profiles aren't yet wired for fallback — bug #4558 (still open, PR #4568 "decrypt encrypted fallback profiles" is open/unmerged). If profiles are stored encrypted (server has

OH_SECRET_KEY

),

FallbackStrategy

currently resolves the fallback without the cipher, so the glm key arrives as Fernet ciphertext (

gAAAAA…

) and auth fails silently (the primary error gets re-raised). Until #4568 lands, plaintext glm profiles work, encrypted ones don't. Because the pr-review plugin hardcodes

LLM(...)

with no fallback and the automation service stores only a single

model

name, wiring "fallback to glm" end-to-end currently needs a small change on one side: either construct

FallbackStrategy

in

plugins/pr-review/scripts/agent_script.py

, or teach the automation service to accept/store a fallback list alongside

model

. There's no existing PR doing that yet.