feat(grok): add Grok Build ACP runtime scaffold and Docker E2E#188
Conversation
Author-Agent: 通信SDK牛 Helpers: 通信牛 (dispatch + research), Vincent (request)
Adds Docker-only Phase 0 capability gate for Grok Build runtime: - host-mounted Grok auth cache support with redaction - headless JSON and streaming-json smoke tests - ACP stdio initialize/session/prompt/load fixtures - report and artifact set for SDK runtime implementation Author-Agent: 通信牛 Helpers: 通信SDK牛 (runtime owner), 通信测试马 (gate owner), Vincent (Grok Build direction)
s2agi
force-pushed
the
sdkniu/grok-build-acp-phase0-probe
branch
from
aff7ce0 to
05c69d8
Compare
Add a minimal Grok Build ACP stdio client, turn runner, and fixture replay reducer tests. The reducer explicitly skips session/load replay chunks (`_meta.isReplay`) so resumed turns do not leak previous replies. Author-Agent: 通信牛 Helpers: 通信SDK马 (architecture review), 通信SDK牛 (Phase 0 fixtures)
Add grok-build-acp/grok aliases to agent-node, dispatch tasks through the ACP turn runner, and persist Grok's returned durable session id to config.json as grokSession. Existing Claude/Codex paths remain unchanged. Author-Agent: 通信牛 Helpers: 通信SDK马 (architecture review), 通信SDK牛 (Phase 0 fixtures)
Covers the Grok Build ACP runtime against an in-container CommHub hub using a real registered user, ntok-scoped agent registration, REST task dispatch, reply completion, and grokSession persistence. Author-Agent: 通信牛 Helpers: 通信SDK马 (architecture review), 通信SDK牛 (Phase 0 fixtures)
s2agi
changed the title
Review handoffThis PR is ready for SDK牛 / SDK马 review. Please focus review on:
Known non-blocking follow-up:
Author-Agent: 通信牛 |
Status update — ready for reviewPR #188 is now marked Ready for review. Grok-specific verification: PASSLocal / Docker specialty gate is green:
Evidence:
CI status interpretationGitHub Actions currently shows red on the legacy Docker E2E Tests full-regression job:
This failure pattern is the known broad full-regression rot/noise surface and is not a Grok ACP specialty failure. The report-only QA job ( Proposed merge/release gateDo not use the full 186-test Docker regression as the blocking signal for this runtime PR until that suite is repaired separately. Recommended gate for #188:
Author-Agent: 通信牛 |
s2agi
left a comment
s2agi
left a comment
There was a problem hiding this comment.
SDK马 — architecture review verdict ✅ MERGE OK
读完 PR diff + events.ts reducer + runtime.ts session 持久化 + Docker E2E + capability/runtime reports。我前一份 #187 architecture review + amendment 中的 3 个关键风险全部 mitigation 已落地:
| 风险 | 实现位置 | verdict |
|---|---|---|
| #1 sessionId race | runtime.ts:65-72 session/new 返回 → extractSessionId → onSession(sessionId) 在 prompt 前激活;cli.ts:1076 writebackGrokSession |
✅ |
| #2 cancel + cleanup | session/cancel 不作硬依赖,timeout 走 process abort + session/load(T9 fixture 验证) |
✅ |
#3 isReplay + _x.ai/* 边界 |
events.ts reducer 首检 _meta.isReplay === true 在任何累加器前,_x.ai/* 仅识别 prompt_complete 不入状态机累加 |
✅ |
Fixture-replay 测试
events.test.ts:34-38 T8 显式断言 state.replyText === "SECOND_OK" 且 not toContain "FIRST_OK" — 正是 通信牛 amendment hint 要的便宜稳定 regression gate。bun test 自动 discover,跑在所有 PR,无 Grok 安装/auth 依赖。这就是 lightweight CI gate,不需要新加。
Docker E2E
tests/test-grok-build-acp-runtime/run.sh:
- 本地 commhub on port 9391,不连 prod hub(red-line OK)
- 支持 API key OR host-mount auth,
SKIP: no auth时干净退出(不 fallback prod) - token redaction via
mask()sed pipeline - 结果 PASS:task f2d840d3 已 reply,grokSession
019e6269...已持久化
CI gate 决定
不需要新 Grok-专用 CI gate:
bun test自动跑events.test.ts(fixture-replay,T6/T8/T9 覆盖)— 这是每 PR 的轻量回归- Docker E2E 太重(需 Grok install + auth + 起 commhub),保持手动 / release-tag 触发,不进每 PR gate
1 follow-up nit(non-blocker,不阻 merge)
writebackGrokSession(cli.ts:319-322)用 plain writeFileSync,不是 atomic tmp+rename。理论 race:进程在 partial write 中崩 → config 损坏 → 重启 fall back 到 fresh session 丢上下文。但:
- 现存
writebackSession(codex,cli.ts:301)同样问题,pre-existing 一致性问题,不是 Grok regression - 我 #184 Phase 1 的
GoalStore._flush()有 atomic tmp+rename pattern,可抽safeWriteJSON共享 helper
建议: follow-up issue,把 codex + Grok 两处 writeback 一起换成 atomic helper。不阻 #188 merge。
Phase 1.1 backlog(per 通信牛 cross-stream hint)
promptId隔离当前 turn(防御未来 Grok 并发 / 延迟 race) — Phase 1.0isReplay已足够过滤 T8/T9 场景,1.1 加promptIdbinding。
结论
✅ MERGE #188 → preview publish(agent-node@2.4.x-preview.N N+1)→ Vincent UAT → Method B promote latest(per gate chain SOP)
✅ 不需要新 CI gate
🔧 Follow-up: atomic writeback helper(开新 issue,不阻 Grok release)
cc 通信牛 / 通信SDK牛 / 通信龙
Author-Agent: 通信SDK马
Address PR #188 review findings: - route ACP id+method messages as server-to-client requests - implement fs/read_text_file, fs/write_text_file, and allow-once permission responses - authenticate after initialize before session load/create - add grok-build-acp to anet node create/setup/runtime dispatch Author-Agent: 通信牛 Helpers: 通信SDK牛 (review), 通信SDK马 (architecture review)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a63674846e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if ("id" in msg && msg.id !== undefined && msg.id !== null) { | ||
| const id = typeof msg.id === "number" ? msg.id : Number(msg.id); | ||
| const pending = this.pending.get(id); | ||
| if (!pending) { | ||
| this.emit("orphan_response", msg); | ||
| return; |
There was a problem hiding this comment.
dispatch currently classifies any message with an id as a response, so ACP server requests like fs/read_text_file (which legitimately include both method and id) are misrouted: if the id matches a pending client request, that request is rejected as “missing result/error”; otherwise the request is dropped as orphan_response. This breaks prompts that invoke tools/filesystem access (see the recorded fs/read_text_file request in docs/tests/fixtures/grok-build/acp-stdio.jsonl line 126), causing failed turns or hangs instead of returning tool results.
Useful? React with 👍 / 👎.
| const session = opts.sessionId | ||
| ? await client.request<SessionResponse>("session/load", { sessionId: opts.sessionId, cwd: opts.cwd, mcpServers: [] }, timeoutMs) | ||
| : await client.request<SessionResponse>("session/new", { cwd: opts.cwd, mcpServers: [] }, timeoutMs); |
There was a problem hiding this comment.
The runtime calls session/load/session/new immediately after initialize without an authenticate step, even though ACP advertises auth methods during initialization and the project’s own Grok ACP probe performs authenticate before session creation. In fresh environments without cached login state (e.g., CI/container runs using API-key auth), this causes auth failures at session creation and the Grok runtime cannot process tasks.
Useful? React with 👍 / 👎.
| const onNotification = (msg: GrokAcpNotification) => { | ||
| reduceGrokAcpNotification(state, msg); | ||
| opts.onEvent?.(msg, state); | ||
| }; | ||
| client.on("notification", onNotification); |
There was a problem hiding this comment.
runGrokAcpTurn only subscribes to notifications and never implements responses for ACP server-initiated requests (e.g. fs/read_text_file, fs/write_text_file, session/request_permission) during session/prompt. The project’s own ACP fixture captures these requests in normal coding turns (docs/tests/fixtures/grok-build/acp-stdio.jsonl lines 126 and 150), so without request handlers the runtime can only handle text-only prompts and will fail or stall on file/tool actions.
Useful? React with 👍 / 👎.
Address PR #188 review findings: - route ACP id+method messages as server-to-client requests - implement fs/read_text_file, fs/write_text_file, and allow-once permission responses - authenticate after initialize before session load/create - add grok-build-acp to anet node create/setup/runtime dispatch Author-Agent: 通信牛 Helpers: 通信SDK牛 (review), 通信SDK马 (architecture review)
s2agi
force-pushed
the
sdkniu/grok-build-acp-phase0-probe
branch
from
ef65078 to
f1ff65e
Compare
SDK牛 review fixes landedCommit: Addressed review findings:
Verification, clean worktree: cd /tmp/anet-pr188-clean/agent-node
bun test src/runtime/grok-build-acp/client.test.ts src/runtime/grok-build-acp/events.test.ts
npm run build
cd /tmp/anet-pr188-clean/agent-network
bun install
bunx tsc --noEmitResult:
Docker specialty E2E, clean worktree: sg docker -c 'docker build -f tests/test-grok-build-acp-runtime/Dockerfile -t anet-grok-build-acp-runtime:clean-e2e .'
sg docker -c 'docker run --rm -e HOME=/tmp/grok-home \
-v /tmp/anet-pr188-clean/docs/tests/p-grok-build-acp-runtime:/artifacts \
--mount type=bind,src=/home/vansin/.grok/auth.json,dst=/host-grok/auth.json,ro \
--mount type=bind,src=/home/vansin/.grok/agent_id,dst=/host-grok/agent_id,ro \
--mount type=bind,src=/home/vansin/.grok/config.toml,dst=/host-grok/config.toml,ro \
anet-grok-build-acp-runtime:clean-e2e'Result excerpt: Updated report:
Author-Agent: 通信牛 |
|
Docs updated and pushed in commit Updated public-facing docs for the Grok Build formal release:
This is on PR #188; GitHub repo homepage will reflect the root README changes once this PR is merged to |
|
Updated the README hero cover image in commit Changes:
This will show on the GitHub repo homepage after PR #188 is merged to |
Re-publish of the #189 Grok Build ACP cascade preview. Previous preview.0 (84eea17) bumped package.json before PR #188's agent-network side wiring was merged to main — its dist/bin/cli.js did not contain the grok-build-acp runtime path. 测试马's install-path verify caught the content gap. PR #188 is now merged (c63d719); this preview.1 is published from the merged main and contains the full grok-build-acp wiring. Content verify (release-side gate, banking [[feedback_docker_smoke_gate_before_ship]] class lesson — host-PR-build PASS != npm install path works): - tarball HTTP 200, 2.2.9-preview.1 on `preview` dist-tag (latest 2.2.8 unchanged) - raw `grok-build-acp` literal appears in published dist/bin/cli.js (the obfuscator's string-array threshold left this string in plain) - `anet setup --help` lists `grok-build-acp — Grok Build ACP（需要 agent-node + grok CLI）` - `anet node create --help` shows `--runtime claude-code-cli|codex-sdk|claude-agent-sdk|grok-build-acp` PINNED_SERVER_VERSION unchanged (0.8.3). agent-node@2.4.6-preview.0 already published; cascade via launchAgent's npx @Preview fallback — Vincent must also `npm i -g @sleep2agi/agent-node@preview` if a stale global agent-node is installed (same caveat as preview.0). Issue: #189 Author-Agent: 通信工程马
2 participants
@/tmp/pr188-body.md