PR Review Room Demo

anet demo pr-review is a built-in demo in the anet CLI (PR #41 merged 2026-05-13; already wired into the CLI's demoCommand). The CLI spawns 4 ephemeral agents — 3 parallel reviewers (security / performance / style) plus 1 final judge — and writes a markdown PR review with an LGTM / Request Changes / Comment verdict that you can paste straight back to a GitHub PR.

Full design contract: docs/demos/pr-review-room-proposal.md.

One-line run

# Run an in-repo sample (any of three)
anet demo pr-review --diff tests/test28-pr-review-room/samples/typo-pr.diff

# Run against your local git diff
anet demo pr-review --ref origin/main

# Run against a live GitHub PR
anet demo pr-review --pr https://github.com/sleep2agi/agent-network/pull/40

Expect 1-3 minutes (3 reviewers in parallel + judge serial). Requires anet login first and a MiniMax API Key:

export MINIMAX_KEY=sk-cp-...

Roles (4 agents)

Role alias	Focus	Output
reviewer-security-<suffix>	injection / credential leakage / authz bypass / SSRF / deserialization / command injection / unsafe reflection / privilege escalation	- [severity] file:line — issue — fix lines + ## 安全 issue 数: N
reviewer-performance-<suffix>	N+1 queries / O(n²) / unnecessary IO / blocking await / large objects / memory leaks / lock contention / missing caches	same shape + ## 性能 issue 数: N
reviewer-style-<suffix>	naming / comments / abstraction levels / dead code / complexity / duplication / type signatures	same shape + ## 风格 issue 数: N
judge-<suffix>	dedup by (file:line) tuple → sort by severity → pick LGTM / Request Changes / Comment	top-of-file **决议：** field + stats + three issue sections + ## 终审说明

<suffix> is 4 hex characters (e.g. 7f3a), overridable with --suffix.

Orchestration timeline (6 steps)

[1/6]  Create 4 agents (alias suffix -<suffix>)
[2/6]  Start 4 agents in tmux sessions → SSE connected
[3/6]  Broadcast review task to 3 reviewers (parallel)
       └─ Each reviewer logs duration + char count on completion
       └─ Parallel total vs serial estimate (saved ~Xs)
[4/6]  Barrier: collect 3 reviews, pack and dispatch to judge
[5/6]  Judge integrates → final verdict
[6/6]  Write ./pr-review-<suffix>-<ts>.md
       └─ Clean up 4 agents + delete dedicated network (unless --keep)
       └─ Print next steps: less to view + gh pr comment one-liner

Parallelism payoff: 3 reviewers run simultaneously, so total review time ≈ the slowest single dimension, not the sum of all three.

Flags

Flag	Default	Description
--diff <file>	—	Local .diff / .patch file
--ref <ref>	—	Auto-derive patch via git diff <ref>..HEAD (e.g. --ref origin/main)
--pr <url>	—	GitHub PR URL, fetched via gh CLI (requires local gh)
--key <key>	$MINIMAX_KEY / $ANTHROPIC_AUTH_TOKEN	MiniMax API key
--out <path>	./pr-review-<suffix>-<ts>.md	Review output path
--keep	false	Retain 4 agents + dedicated network after run
--step-timeout <s>	180	Per-reviewer / judge timeout in seconds
--suffix <s>	random 4 hex	Alias suffix
--no-network	false	Run in current / default network
--network <id>	—	Use an existing network

At least one diff entry flag is required.

Three in-repo sample diffs

tests/test28-pr-review-room/samples/ ships three fixture diffs covering typical scenarios:

File	Scenario	Expected verdict
good-pr.diff	Single-line README typo fix	LGTM
typo-pr.diff	1-2 minor style issues (dead helper / over-commented)	Comment
cross-file-pr.diff	SQL injection + N+1 query + hardcoded password + path traversal	Request Changes

Run it:

anet demo pr-review --diff tests/test28-pr-review-room/samples/cross-file-pr.diff

Expected output structure is asserted in tests/test28-pr-review-room/expected/assertions.json — verdict regex / per-dimension issue count ranges / required sections.

⚠️ LLM output is non-deterministic. What's shown below is the output structure and fields (verdict / issue counts / required sections / console pacing) — exact text, ordering, and tone shift with model state, prompt hit rate, and sampling temperature. The structural assertions in assertions.json (regex + count ranges) are more reliable than verbatim string matches.

Output markdown structure

./pr-review-<suffix>-<ts>.md follows a fixed shape (see cli.ts lines 4876-4891):

# PR Review

**来源**: <local file / git diff / owner/repo#N>
**大小**: X.X KB
**时间**: <localized datetime>
**Run**: <suffix>
**总耗时**: Ns

**决议：** LGTM | Request Changes | Comment
**统计：** 安全 N1 处 / 性能 N2 处 / 风格 N3 处

## 安全
- [严重] src/auth.ts:42 — User input concatenated into SQL — use parameterized query
- [中] ...

## 性能
- [中] src/api/users.ts:88 — DB query inside for-loop (N+1) — batch the query

## 风格
- [低] src/utils/format.ts:12 — Name `tmp` is unclear — rename to `formattedRows`

## 终审说明
This PR introduces one severe security risk (SQL injection); recommend switching
to a parameterized query before merge. Performance and style issues are non-blocking
and can be addressed in a follow-up PR.

---
## 附：3 reviewer 原始输出

### reviewer-security (reviewer-security-<suffix>, Xs)
<raw markdown>

### reviewer-performance (reviewer-performance-<suffix>, Xs)
<raw markdown>

### reviewer-style (reviewer-style-<suffix>, Xs)
<raw markdown>

The text content above is illustrative — actual reviewer findings come from the LLM and vary every run. Section headers are in Chinese because the agent prompts use Chinese (preserved verbatim by judge).

Console output cadence

What you see in the terminal once the run starts (see cli.ts lines 4711-4893):

  🔍 PR diff: local file tests/test28-pr-review-room/samples/typo-pr.diff
  📏 Size:   1.0 KB
  📡 Hub:    http://127.0.0.1:9200
  📂 Net:    (pr-review-7f3a net_ab12cd34...)
  🆔 Run:    7f3a

  [1/6] 创建 4 个 agent (alias 后缀 -7f3a)...
        ✓ 创建/更新 4 个 agent
  [2/6] 启动 4 个 agent (tmux session)...
        ✓ 4 agent 全部 SSE connected
  [3/6] 广播 review task 给 3 reviewer (parallel)...
        ✓ reviewer-security-7f3a       38s, 412 字
        ✓ reviewer-performance-7f3a    41s, 287 字
        ✓ reviewer-style-7f3a          52s, 503 字
        ─ 并行总耗时 52s (估串行 131s, 节省 ~79s)
  [4/6] barrier 收齐 3 份 review，整包派给 judge...
  [5/6] judge 整合 + 终审...
        ✓ judge-7f3a 19s, 786 字
  [6/6] 写入 review: ./pr-review-7f3a-1714766365.md
        ✓ 1842 字写入 ./pr-review-7f3a-1714766365.md

  🧹 清理 4 个 agent (用 --keep 跳过)...
        ✓ 删除独立 network (net_ab12cd34...)
        ✓ 清理完成

  🏁 完成！review: ./pr-review-7f3a-1714766365.md

     下一步建议:
       1. 查看: less ./pr-review-7f3a-1714766365.md
       2. 贴到 GitHub PR: gh pr comment <PR-N> --repo <owner>/<repo> -F ./pr-review-7f3a-1714766365.md

Console output is in Chinese (CLI is bilingual: code-path strings + log lines are Chinese; flags / docs are bilingual). Durations / char counts / saved-seconds shown are illustrative — actual LLM runs vary ±50%. The proposed +seconds prefix from the v3 §UX feedback section is planned for a follow-up iteration.

When using --pr, the trailing gh pr comment line auto-fills the PR number + repo from the URL — copy-paste as is.

Network isolation

Same machinery as the Debate Demo:

1. Auto-creates a dedicated pr-review-<suffix> network
2. Provisions 4 ephemeral agents into it
3. Includes network_id in dispatched tasks
4. Cascade-deletes all 4 agents + network on exit (tasks, inboxes, node data all cleared)

To keep the run for Dashboard inspection:

anet demo pr-review --diff <file> --keep

With --keep, the CLI prints manual cleanup commands at the end:

  📌 已保留 4 个 agent (alias 后缀 -7f3a)。手动清理:
     tmux kill-session -t pr-review-7f3a-*
     anet node delete reviewer-security-7f3a reviewer-performance-7f3a reviewer-style-7f3a judge-7f3a
     anet network delete net_ab12cd34...

To reuse an existing network:

anet demo pr-review --diff <file> --network net_xxx

To fall back to the current / default network (legacy behavior):

anet demo pr-review --diff <file> --no-network

Troubleshooting

Each entry below quotes the actual error text from cli.ts:

❌ 没有 hub. 先 'anet init' 或 'anet hub start'.

No init or hub not running. anet init to configure the hub URL, then anet hub start to launch commhub-server.

❌ 没有 token. 先 'anet login'.

Not logged in. Run anet login to obtain a user token.

❌ 需要 --diff <file> / --ref <ref> / --pr <github-url> 之一

One of the three diff entry flags is required. --diff is most reliable (no network or gh-CLI dependency), then --ref, then --pr.

❌ --diff 文件不存在: <path>

--diff path is wrong. Check relative path vs current working directory. Bundled samples live at tests/test28-pr-review-room/samples/.

❌ git diff 失败: ...

--ref failed to run git diff <ref>..HEAD. Common causes: ref doesn't exist (run git fetch), or current directory is not a git repo (cd to project root).

❌ git diff <ref>..HEAD 输出为空

No diff between <ref> and HEAD. Try a different ref, e.g. --ref main or --ref HEAD~5.

❌ --pr 需要本地装 gh CLI

--pr requires gh CLI. After install, run gh auth login once. Or switch to --diff / --ref to bypass.

❌ --pr 不是合法 GitHub PR URL

URL must be https://github.com/<owner>/<repo>/pull/<N>. Note: it's the PR URL, not an issue URL.

❌ gh api 拉 PR diff 失败

gh CLI not logged in, missing permissions, or private repo. Run gh auth status to check.

⚠️ diff 大小 X KB > 30 KB

Non-fatal warning, but MiniMax context may not fit a large diff. Use gh api -X GET 'repos/.../files' to pick key files and run them separately.

❌ 需要 MiniMax key

No env var or --key was provided. export MINIMAX_KEY=sk-cp-... or pass --key sk-cp-... once.

❌ 创建 network 失败

Hub returned a non-OK response. Run anet status (or anet doctor to diagnose hub reachability), or pass --no-network to fall back to the default network.

❌ tmux <alias>: ...

An agent's tmux session failed to start. Check tmux ls for stale same-named sessions; tmux kill-server and retry in the worst case.

❌ timeout waiting for <alias> reply

A reviewer or the judge didn't respond within --step-timeout (default 180s). Usual causes: MiniMax quota exhausted, model failure, or slow inference on a large diff. Bump --step-timeout 360 or run anet logs <alias> for the SDK error.

Advanced usage

Paste output back to a GitHub PR

With --pr, the closing message gives you a ready-to-copy gh pr comment line (PR number + repo pre-filled), e.g.:

gh pr comment 40 --repo sleep2agi/agent-network -F ./pr-review-7f3a-1714766365.md

For --diff / --ref modes, the closing message gives a template — fill in <PR-N> and <owner>/<repo> yourself.

Watch it in Dashboard

Start the Dashboard first:

anet dashboard

Then run the demo in another terminal and open http://127.0.0.1:5173/network/<network_id> to see the topology + live message flow of all 4 agents. With --keep, the scene is preserved for replay.

Switch model

The default is MiniMax-M2.7. Swap the 4 agents' ANTHROPIC_* env vars to any other Anthropic-compatible endpoint (DeepSeek / GLM / Kimi / Claude). See Multi-model. A side-by-side review-quality comparison across models is planned for v2.

Next

More demos:

• Debate — 6 agents, 9-step turn-based orchestration
• Hello World — minimal two-agent conversation
• Translation Pipeline — 3-agent chain

Customize and dig in:

• Curious why there's no dedicated dispatcher agent? See §Roles in the full proposal.
• Want to modify the demo? Source at agent-network/bin/cli.ts:4585 demoPrReviewCommand.
• Why a dedicated network per run? See Networks & Nodes.

Watch in Dashboard:

• anet demo pr-review --diff <file> --keep → open the Dashboard → in the topology you'll see the funnel shape: 3 reviewers in parallel → judge serial.
• The task panel walks you through the I/O of all 4 agents.