For AI agents

This page is the primary entry point for authors of OpenCode / spcode, Claude Code, Codex, Cursor, and other agent hosts that invoke SoftProbe through shell tools.

把文档喂给你的 Agent

整套文档以 llmstxt.org 格式发布，Agent 无需抓取 HTML 即可直接消费：

URL	内容
/llms.txt	带描述的全站页面索引 —— 体积小的入口
/llms-full.txt	整套文档正文合并为单个纯文本文件
‹任意页面›.md	单个页面的 Markdown 源（在 URL 后追加 .md）

先让 Agent 读 /llms.txt，它会按需跟进到全文或单页 .md。

Design goals

1. One process, one job — each tool call runs a single sp command with explicit flags. Do not rely on shell aliases or interactive prompts.
2. Always pass --json for machine parsing unless you are showing output to a human in chat.
3. Use artifacts for large payloads — diff bodies, log downloads, and record queries write files under --out-dir; stdout returns paths and summaries only.
4. Backend data only — use API commands (sp app, sp replay, sp record, …) for cluster state.

Recommended tool shape

Wrap sp as a shell tool with a fixed argv prefix:

sp --json --profile "${SP_PROFILE:-default}" <subcommand> ...

Environment variables agents should set in the host config:

Variable	Purpose
SP_API_URL	Backend base URL (e.g. http://127.0.0.1:8090)
SP_TOKEN	JWT from sp auth login or CI secret
SP_PROFILE	Named profile in ${XDG_CONFIG_HOME}/softprobe/config.jsonc or sp.jsonc
SP_CONFIG	Extra config file path loaded after sp.jsonc and before --config

sp also reads ${XDG_CONFIG_HOME:-~/.config}/softprobe/config.jsonc for shared SoftProbe settings and ${XDG_CONFIG_HOME:-~/.config}/softprobe/sp.jsonc for CLI-specific overrides. Agents should prefer SP_API_URL and SP_TOKEN when running in ephemeral CI containers.

Lifecycle command order

Use these job-oriented commands before low-level building blocks:

1. sp doctor --json
2. sp app create <name> --json → save data.appId
3. sp policy recording apply -f … --json
4. 从 install.softprobe.ai 下载 sp-agent.jar，然后运行 sp agent command --app <id> --agent-jar ./sp-agent.jar --json
5. Start the app with data.startCommand; send traffic
6. sp record case list --app <id> --since -1h --json
7. sp policy mock apply / sp policy compare apply
8. sp replay run --app <id> --env <service-base-url> --json
9. sp replay status <planId> --json or sp diagnose replay <planId> --json

Typical skill flows

Diagnose a failed replay

Commands: see Diagnose replay failure.

Change policy and re-run

1. sp policy recording validate -f policy.yaml --json
2. sp policy recording apply -f policy.yaml --json
3. sp replay run --app … --env <service-base-url> --json
4. sp replay status <planId> --watch --json

Migrating from sp_api (spcode plugin)

The OpenCode @spcode/plugin tool sp_api uses named endpoints (diff_detail, query_replay_case, …). The sp CLI exposes the same operations as stable subcommands.

Old	New
sp_api endpoint=list_applications	sp app list --json
sp_api endpoint=diff_detail diffId=…	sp replay diff get … --out-dir … --json
sp_api endpoint=find_traces_by_attr …	sp trace find … --json

Skills should be updated to call sp so behavior is consistent outside OpenCode.

Error handling for agents

Exit code	Meaning	Agent action
0	Success	Parse stdout JSON
1	API or business error	Read stderr JSON error field; may retry or ask user
2	Usage / missing config / auth	Fix argv or run sp auth login / set SP_TOKEN
3	Auth required	Refresh token

See Output contract and Exit codes.

What not to call

• Agent write APIs (POST /api/storage/record/save, batch saves) — instrumentation only
• OTLP ingestion (/v1/traces) — use collectors
• Mutating sp ops without --confirm — blocked by design

Related pages

• Output contract
• Versioning
• Commands