✅ Done! Reviewed all three PRs and left a verdict on each: #17555 — ✅ Approved. The diagnosis is right: the old oracle polled the agent-server's global

/api/bash/bash_events

stream, which only holds commands the UI runs itself, so any exit-0

BashOutput

(even

stdout: null

) satisfied it. Scoping success to a

TerminalObservation

for the exact command, with the token required in that same observation, is the correct fix, and the 14 unit cases cover the tricky paths. Answering the SDK's title-generation request outside the scripted trajectory (mirroring the pre-flight ping) is a solid companion fix. Only ask: the positive Pull/Push assertion in

mock-llm-files-and-git.spec.ts

is a product-behaviour claim — make sure the maintainer decision behind it is visible on the PR. #17876 — 🔁 Changes requested. The fixture fixes are sound (recognizing the system + one-token user

ping

validation format, and writing npm-path user skills under

dirname(STATE_DIR)/skills

). But removing all six quarantines is pre-emptive: the PR's own evidence shows the full Linux workflow still exposes four tracked failures, so this just turns

skipped

into

failed

on

main

until the dependencies land. Also,

Validate PR description

is red here. #17889 — ✅ Approved. This is a real production fix, not a test-only restore: onboarding built the new profile from the dirty-field diff, dropping an endpoint/API key that already matched saved settings and silently falling back to the default provider (the

LLMAuthenticationError

in the recording). Reading the full saved settings and preserving unchanged values is right. Two notes:

Validate PR description

is red (needs the

bug

label on #17884), and the trajectory-activation overlap with #17876 should be spelled out. One clarification: #17872 is the tracking issue, not a PR, so there was nothing to approve there — it's best left open until all five failing specs actually pass on

main

. Also worth flagging: #17555 and #17876 both edit

tests/e2e/mock-llm/scripts/mock-llm-server.py

(title handling vs. validation handling), and both #17876/#17889 touch mock-LLM trajectory activation — so the merge order matters to avoid churn.

✅ Done! Reviewed PR #18209 and approved it. 👍 It's a tidy, low-risk harness change confined to

.agents/skills/verify-openhands/

(plus

.pr/

evidence): the new pure helper

lib/agent-profile-repoint.mjs

decides when

llm preset

should repoint the seeded

default

agent profile, with unit tests covering the move, the already-there/missing-ref cases, and the ACP case. It's wired into

preset

only —

llm set

, named profiles, and fresh-run seeds are deliberately left alone, which matches the rationale in the description. All CI checks are green (the earlier "Validate PR description" failure was re-run and now passes), and I confirmed the helper's behavior against a few edge cases myself. No major issues found.