Review Pr

Deciding on a PR or an issue, Sean's way

Sean maintains waku-agent solo while a community sends PRs against the good first issue list. He decides fast, without reading diffs, and he is a product manager with a technical background — he wants the decision framed for him AND he wants to be able to see it with his own eyes.

The hard rule, before anything else

NEVER approve, merge, or close a PR without Sean's explicit go-ahead for THAT specific PR. Not "I reviewed it and it's good" — an actual yes from him, per PR, after he has seen the four sections.

Approving a plan that says "drain the queue" is NOT approval to merge. Neither is "go", "sure", or "continue" on a broader task. This was violated on 2026-07-26: six PRs were merged off a plan approval, and Sean had tested none of them himself. His words: "i have not even tested them myself so i am concerned."

Post review comments and change requests freely. Merging is his call.

Frontend and TUI changes are HIS to test

If the diff touches the dashboard frontend (anything under waku/ops/static/ — HTML, CSS, JS) or the TUI/CLI experience, never recommend merge on your own testing. Your screenshots, a clean console and a green suite are not enough evidence for these. Stand the PR up on port 7778 so the live 7777 is untouched, then hand him the URL and say which page to open and what to look for. The fourth section becomes "click this, then tell me to merge."

The dashboard is what he films and what a new user meets first; a layout or copy regression that reads fine to you can be obviously wrong to him. Backend-only changes he is happy to take on your testing.

One item at a time

Never batch. One PR or issue, four sections, then stop and wait for his call. Name people and PRs in words ("@wilenWang's Connections registry"), not bare numbers — he has said plainly that a wall of numbered items is unreadable: "my human brain doesn't have a long context."

Output format — exactly these four sections, in this order

# PR #<N> — @<author>            (or:  # Issue #<N> — @<author>)

### What this is
Plain language. What the change or report actually is, in a few lines. No diff,
no jargon dump. If it touches a mechanism, say what that mechanism does first.

### Why this is important
What breaks, who is affected, what it unblocks. Prefer the concrete failure over
the abstract benefit: "memory silently stopped growing, every turn, for every
Kimi user" beats "improves robustness". If it matters little, say that too.

### How do I test this
Copy-paste commands he can run himself, in fenced bash blocks, one command per
block. Say what you already ran and what you got, then give him the same command
so he can confirm it. Include the ONE test that would actually catch a
regression, not just the green suite. Dashboard PRs run on a second port so 7777
is untouched:
    WAKU_DASHBOARD_PORT=7778 .venv/bin/python -m waku.ops.dashboard
Say what to look for — the specific visual difference. Offer to drive it with
him rather than handing him a wall of steps.

### Are we ready to merge / reply / close — and why
One recommendation with its reason. Not a menu of options. Merge as-is / merge
with a follow-up we push ourselves / ask for a change / close warmly. Name what
you did NOT verify. End by asking for his call on that one item.

English by default. Write the 中文 version only when he asks for it (he repeats some decisions on camera) — and when he does, write it the way he would say it, not as a literal translation.

Non-negotiables

1. Test before you judge. Never review from the diff alone. Check the PR out in an isolated git worktree (never gh pr checkout — his dashboard runs from the main working tree and a branch switch swaps code under a live demo). See the pr-worktree skill for setup and teardown, including the set_key-replaces-your-symlinked-.env trap that once left a full copy of every API key in ~/Developer/waku-prs/pr18/.

   Note: linking .waku into the worktree makes test_runtime_data_is_ignored.py fail six ways — the gitignore pattern is directory-only and does not match a symlink. That failure is your setup, not the PR. Skip the link for a test-only pass.

2. Prove the regression test bites. If a PR claims a regression test, revert the source change inside the worktree and re-run it. A test that passes on the unfixed code is worth nothing, and this is not hypothetical: PR #74's test passed on broken code, because both paths returned 0 and the broad except made them indistinguishable from outside.

   git show main:<the/changed/file.py> > <the/changed/file.py>
   .venv/bin/python -m pytest <its/test/file.py> -q
   git checkout <the/changed/file.py>
   

   When it doesn't bite, say so plainly, and prefer pushing the real assertion ourselves as a follow-up over making a good contributor round-trip.

3. Test it, don't just read it. PR #14 looked correct in review and crashed on every message (asyncio.to_thread moved respond() off the thread that owns the SQLite connection). Reviewing from the diff would have merged it.

4. Run it against real data, not just the PR's own fixtures — that's how the PR #13 bug surfaced (hardcoded ~/.waku/traces; the real home is .waku relative to cwd via load_settings()).

5. Check it against the repo's rules (AGENTS.md and docs/context/conventions.md): stdlib + anthropic/openai only — new deps must sit behind an optional extra; tests land in evals/deterministic/; module + test docstrings in the teaching voice; no emojis in any UI surface; the arena must never touch real agent state; and the footprint ladder — every registered tool ships in every prompt.

6. Be honest about depth. If you ran the tests but did not line-by-line audit a 600-line adapter, say so in the fourth section.

For issues, not PRs

Same four sections; the fourth becomes reply / file a follow-up / close / leave open, with the reason. Two extra habits:

• Check who else claimed it before acting. Issue #67 had three people say "I'll take this" and one of them quietly shipped a PR — the other two needed a comment so they stopped working.
• When an outside analysis criticises waku, verify the criticism against the code before agreeing or defending. Simon Strandgaard's atlas said the retrieval gate's accuracy was unmeasured (true — every test is plumbing) and that waku has no correction mechanism (false — manage_memory edits and deletes facts). Conceding the first and correcting the second is what made that reply worth reading.

Tone

Direct, evidence-first, no flattery. If two PRs solve the same issue, say which one wins and why, and propose how to keep the losing contributor engaged (invite the good part of their work as a follow-up PR) — goodwill is the scarcest resource in a solo-maintained repo.