✅ Done! I have everything I need. Here's the full picture. Why the bot failed The reviewer automation didn't fail because of the PR — it failed on LLM quota exhaustion in your environment. The run logs (

agent-canvas.2026-09-10.log

, run

0ba1fb10…

) show the review conversation erroring out with:

litellm.RateLimitError: {"error":{"type":"usage_limit_reached",
  "message":"The usage limit has been reached","plan_type":"team",
  "resets_in_seconds":421890}}

The agent retried 4× (backing off each time), the conversation went to

error

, and the automation posted the generic

⚠️ …encountered a problem (status: error)

comment.

resets_in_seconds ≈ 421890

≈ ~4.9 days, so it will keep failing until the team-plan limit resets or is raised — unless a fallback model kicks in. Notably, the fix for exactly this is sitting unmerged in this same repo:

software-agent-sdk#4917

"Fail over to fallback LLM immediately on hard quota exhaustion" (fork branch

quota-fallback-immediate

). It adds

is_quota_exhaustion_error()

so

usage_limit_reached

/

insufficient_quota

skip the retry loop and fail over to a fallback LLM on the first attempt instead of after the full backoff. I reviewed that diff too — it's clean and correct (a

retry_if_exception_type(...) & retry_if_exception(lambda e: not is_quota_exhaustion_error(e))

tenacity predicate). Merging it only helps if a fallback model is actually configured; the automation's

model:

is currently

null

. The same failure hit

OpenHands#17221

(the identical PR in the OpenHands repo) at 04:33 — same quota error.

automation#431

and

extensions#451

are still pre-review (HUMAN-section/CI gates), so they haven't spent an LLM call yet. The PR itself (software-agent-sdk#4921) Docs-only: adds

.agents/skills/pr-design-doc/SKILL.md

(168 lines) +

references/html-craft.md

(322 lines). This is a skill that teaches the agent to author a self-contained HTML design doc for non-trivial PRs. Low risk, well written. The workflow (verify PR head → read both sides → before/after SVG → commit under

.pr/

→ link via htmlpreview) and the safety stance (reject symlinks, never publish private-repo docs, escape all source-derived text) are genuinely good. Issues worth raising (non-blocking): 1. Duplicated across four repos — this is the real smell. The exact same skill is landing in

software-agent-sdk

,

OpenHands

,

automation

, and already exists in

extensions#451

. All three new ones have a byte-identical

SKILL.md

(blob sha

62f02c6a…

) and near-identical

html-craft.md

.

extensions

is the canonical skills home — this copy-paste will drift (it's already drifted: 168 vs 137 lines SKILL, 322 vs 295 lines html-craft). Recommend one source of truth, with the others referencing it. 2.

test ! -L .pr && test ! -L .pr/design.html

is a reasonable guard but is advisory, not a security boundary — it's a TOCTOU check on a worktree the agent already owns. Fine, just don't over-claim. 3. Cleanup assumption is repo-specific. The skill correctly says to verify

.github/workflows/pr-artifacts.yml

exists before relying on cleanup. I confirmed it exists in

software-agent-sdk

(and in `OpenHands`/`automation` checkouts), and its

cleanup-on-approval

job only runs for same-repo PRs — for forks it deletes

.pr/

from base after merge. The skill's "fork PR → remove manually before merge" instruction matches, so this is consistent, but it means the doc can linger on fork PRs. 4. Trigger collision risk:

/design-doc

is a very generic trigger; combined with the long

description

keywords it may fire more broadly than intended (e.g. "add a design doc"). Worth tightening. Verdict: ✅ approve (as docs) once the maintainer decides on the dedup strategy; the bot failure itself is an infra/quota issue, not a code issue. The sibling PRs •

OpenHands/OpenHands#17221

— same content, same author (enyst),

Fixes #16304

. Same review applies verbatim. Currently stuck on the same quota error. •

OpenHands/automation#431

— same content (

Fixes

an automation issue). Bot has only posted the HUMAN-section nudge so far, not reviewed. •

OpenHands/extensions#451

— the original (smolpaws), in the older

skills/<name>/

plugin/marketplace layout (with `.claude-plugin`/`.codex-plugin`/`commands/`) rather than the newer

.agents/skills/

convention. This is the one the others should arguably consolidate into. Currently pre-review (HUMAN + CI gates). Recommended next step: merge

#4917

(the quota-failover fix) and/or raise/reset the OpenAI team quota, then re-trigger the reviews on

#4921

,

#17221

, and

#431

.