agent-bridge
Skill0 starsChanged 56 days agoApache-2.0View on GitHub
Delegate implementation to an external coding agent through Agent Bridge. Use when Codex should supervise Claude Code or DeepSeek Harness instead of editing the target repo itself. Covers bridge_run, continue, respond, approve, apply, reject, cancel, doctor, agents, prune.
8 other projects use a version of this.
Adopted by
Most recently taken up by
- kununu/agent-bridgetheir own version, first seen Nov 12, 2026
- Latth/agent-bridgetheir own version, first seen Nov 8, 2026
- cocodrino/bridge-harnesstheir own version, first seen Oct 25, 2026
- ilamanov/skillstheir own version, first seen Oct 13, 2026
- avas-app/agent-bridgetheir own version, first seen Sep 24, 2026
- Mihirokte/agent-skillstheir own version, first seen Jul 31, 2026
Other versions8 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 8
- kununu/agent-bridge★ 18same name
- xyva-yuangui/agent-bridge★ 15same name
- EthanSK/agent-bridge★ 11same name
- cocodrino/bridge-harness★ 4same name
- avas-app/agent-bridge★ 3same name
- Latth/agent-bridge★ 2same name
- ilamanov/skills★ 1same name
- Mihirokte/agent-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
Files
Agent Bridge
Codex 是 Supervisor。Worker 是外部完整 Coding Agent。Bridge 不调用 LLM。
MCP server 名:agent-bridge
工具前缀:bridge_*
安装:npx -y codex-agent-bridge,然后新开 Codex。
什么时候用
用户要改一个 git 仓库,且应该由 Claude Code 或 DeepSeek Harness 执行。
什么时候不用
- 只问问题、读代码、做设计
- 没有 git 仓库
- OpenCode(当前 DEFERRED)
- Codex 自己改当前工作区即可
闭环
bridge_run
→ 若 state=WAITING_FOR_INPUT:读 pendingInput.options,bridge_respond(optionId + stateVersion)
→ 若 state=TASK_TIMED_OUT:wait 预算到了,Worker 已停,worktree 保留。读 ReviewPacket,然后 continue 或 reject
→ 重复直到 AWAITING_REVIEW
→ 读 structuredContent.reviewPacket(含 verification)
→ verification.passed=false 时不要 approve,除非用户明确要求
→ 不通过:bridge_continue(notes + stateVersion)
→ 再读 ReviewPacket(continue 过程中也可能再 WAITING_FOR_INPUT)
→ 通过:bridge_approve(stateVersion)
→ 用户要落到当前分支时:bridge_apply(cherry-pick,不是 merge)worker:调试用 replay;闸门自检可用 fake;真干活用 claude 或 deepseek。V1 只选 worker。不要传、不要编造 model / provider / effort。Agent Bridge不管理DeepSeek model/provider/reasoning effort。DeepSeek effective route由运行中的Harness ACP composition决定。当前legacy integration使用上游composition配置。Claude Code 沿用其原生配置。Claude 背后的官方 Anthropic、CC Switch 或其他第三方 Provider 是 Worker 内部实现,不要猜测 effective upstream model。
项目里如果有 .agent-bridge/verify.json,bridge_run 会默认跑其中全部 verifyId。也可显式传 verifyIds。示例:templates/verify.json。
破坏性调用必须带当前 task.stateVersion。冲突时不要重试旧版本。
每次 bridge_* 任务调用都必须带同一个 project 路径。任务 UUID 的字段名是 task,不是 taskId。bridge_respond 示例:project + task + stateVersion + optionId。漏掉 project 会查到空任务库或报 journal 不存在,看起来像状态丢失。
clientRequestId 相同且内容相同会返回同一 task;内容不同会 TASK_ALREADY_EXISTS。
MCP 默认 permissionMode=gate:Worker 要写文件/跑命令时会停在 WAITING_FOR_INPUT,不要自动当任务完成。选 allow_once,除非用户明确禁止。optionId 以 pendingInput.options 为准(Claude 可能是 allow 而不是 allow-once)。没有 live waiter 时(Core 重启)不要 respond,改 bridge_continue。
timeoutMs 到期会停 Worker、保留 worktree,状态变成 TASK_TIMED_OUT。这不是完成。可以 bridge_continue 或 bridge_reject。
CLI 默认 auto(一次进程无法跨调用停闸)。
掉线后先 bridge_status(needsAttention: true)或 bridge_doctor。Core 重启会把 QUEUED / STARTING / RUNNING / VERIFYING / WAITING_FOR_INPUT 收成 AWAITING_REVIEW + interrupted,不要 respond,改 bridge_continue(Claude 可尝试 session/load)。doctor 标出遗留 worktree 时用 bridge_prune。
硬规则
- turn 结束 ≠ 任务完成
- Worker 不得 commit / push / merge / rebase
- 不要用
base..HEAD当 diff;以 ReviewPacket /bridge_diff为准 - 不要信 Worker 自称测过;验证只认
.agent-bridge/verify.json的 verifyId - approve 后 worktree 会拆掉,checkpoint 留在任务分支;落到当前分支用
bridge_apply(cherry-pick) - 不要 merge
- crash / 死 pid / MCP 进程退出:不要 reattach 死 pid。hydrate 后看
interrupted。bridge_continue会尝试 Claudesession/load;失败或 DeepSeek 则 REHYDRATE - DeepSeek 没有
session/load;Claude 可以 cold load bridge_logs/ journal 会脱敏,不要把密钥写进 notes- V1 不管理 Worker 的 model / provider / effort;
bridge_run只选 worker
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

