Backlog triage: low-risk closes and quick wins to thin the PR queue

[IsmaelMartinez]

Hi @naorpeled,

I've been onboarding pr-agent across a few repos and ended up doing a full pass over the open PRs. There are a lot of them, and from experience a single maintainer carrying that kind of volume gets noisy fast, so I thought I'd offer a hand rather than add to the pile.

Below is a deliberately conservative first cut: only the extra-small and small PRs, split into "looks safe to close" and "small and basically ready". Everything here is a suggestion, your call entirely.

(I've obviously left my own open PRs out of all of this, triaging those would just be me reviewing my own work :D, so the list below is everyone else's.)

Candidates to close (stale or already steered elsewhere)

• fix: initialize max_tokens before try block in _calc_claude_tokens #2269 — you'd already asked for a different approach (let it fall through), and there's been no follow-up since.
• fix: Apply suggestions_score_threshold filter to inline code suggestions in dual publishing mode #2121 — roughly 7 months idle, author stopped engaging with the review notes.
• fix: YAML parser error from diff-style lines in nested lists #2061 — roughly 7 months idle, no review. Sound fix, but it looks abandoned.

Closing these, or nudging the authors first, drops the count with effectively no risk.

Quick wins (small, low-risk, near-ready)

Just needs a review:

• fix: make GitHub PR-description ticket selection deterministic #2422 — deterministic ticket ordering in GitHub PR descriptions (tests in, CI green).
• fix(litellm): detect GPT-5 family when model has openai/ or azure/ prefix #2401 — GPT-5 detection with openai/ and azure/ prefixes (well tested, green).

One small fix from ready:

• chore: improve pr-agent maintenance path #2431 — convert the new asserts to double quotes (Ruff).
• fix(gitlab): add null check for target_file in publish_code_suggestions #2316 — null-check target_file, and sanitise one log line.
• fix(description): use regex for minute suffix replacement instead of blanket .replace #2318 — regex minute-suffix, plus an isinstance guard.
• fix: preserve original indentation character in /improve suggestions #2429 — preserve tab indentation in /improve, plus an edge-case test.
• fix(gerrit): prevent path traversal in publish_code_suggestions #2314 — Gerrit path-traversal guard (security). Logic is right, wants tests.
• fix(gitea): tolerate binary file payloads (#2380) #2440 — tolerate binary payloads in Gitea, just needs a decision on non-UTF-8 text.

Correct micro-fixes needing a test or rebase:

• fix: use False as default for draft field in pull request event check #2273, fix: correct inverted condition in DefaultDictWithTimeout TTL refresh #2274, fix: avoid UnboundLocalError in _calc_claude_tokens error handler #2256, fix: use pr_url variable instead of literal "pr_url" for provider cache lookup #2257 — draft default, TTL operator, UnboundLocalError, cache-key lookup.
• fix: propagate exception from verified inline comment publishing #2258 — propagate the inline-publish exception. I'm happy to add the regression test I offered there.

If this is useful, I'm glad to keep going. I've triaged the whole backlog (the larger features, the two duplicate MCP PRs, a couple of security ones) and can write that up too, or help review and close in whatever order is most useful. Only if it's welcome, I don't want to add noise.

I hope this helps, and do let me know if it needs any changes.

[naorpeled]

Thanks Ismael!!!

I'll take a deeper look at the list during the weekend but thank you so much!

I really appreciate it ❤️