eide-rebuild— agent skill

Use this skill when the user wants an EIDE project rebuilt from Codex.

Trigger patterns

• Natural language build requests such as 你自己编译验证下对不对
• Natural language build requests such as 帮我编译确认一下
• Natural language build requests such as 先 rebuild 看结果
• Explicit requests such as EIDE rebuild C:\work\demo\project.code-workspace
• Long-log requests such as EIDE subagent rebuild C:\work\demo\project.code-workspace

How to execute

1. Resolve the target path.
2. Prefer a user-provided .code-workspace path.
3. If the user points at a project directory, pass that directory to the runner and let the runner resolve the single workspace file.
4. If no path is supplied, inspect the current working directory. Use it when it contains exactly one .code-workspace file.
5. If the current working directory does not identify one workspace, ask the user for the path.

Decision table:

Run:

python scripts/eide_rebuild.py rebuild <workspace-or-project-path> --stdout minimal

The relative scripts/eide_rebuild.py path is relative to this installed skill directory. Host integrations can use their own installed absolute runner path.

Environment check:

python scripts/eide_rebuild.py doctor

Result handling

• Treat the runner as the source of truth.
• Prefer --stdout minimal for normal agent work. It prints bounded JSON status and still writes the complete result to resultPath.
• Use --stdout summary only when stdout needs per-target artifacts, memory, source stats, failures, or diagnostics without opening resultPath.
• Use --stdout full or omit --stdout only when the full JSON is explicitly needed on stdout.
• First inspect ok, exitCode, errorCode, summary, targetNames, targets[].ok, targets[].failureCount, targets[].diagnosticCount, targets[].artifactCount, and resultPath.
• On success, report only a low-noise summary by default: ok, exitCode, errorCode, summary, targetNames, failure count, diagnostic count, artifact count, and resultPath.
• Use targets[].artifacts[].sha256 as artifact identity/provenance data. Do not treat hash changes across different rebuilds as success/failure evidence unless the user explicitly asks for reproducibility or deterministic build checks.
• Keep the complete JSON available at resultPath; it includes compilerLog, steps, artifacts, and transcript.
• Do not paste full logs or long artifact lists on successful builds.
• Read compilerLog, transcript, and steps[].stdout/stderr from resultPath only when the build fails, the user asks for details, or the structured failures / diagnostics fields are not enough.
• Use exit code 0 for success, 6 for build failure, and the other exit codes for environment or tool errors.

Subagent guidance

• Prefer the eide-rebuild custom agent when the user explicitly asks for subagent/delegated rebuild work and the host supports delegation.
• Otherwise use direct runner execution for normal single-project requests.
• Use generic worker subagents only when the host supports delegation, policy allows it, and the user explicitly asks for subagent/parallel/delegated/background work.
• The worker should run the same Python runner with --stdout minimal.
• The worker should return only the minimal stdout plus a short factual conclusion, not paste the full result JSON, full logs, or long artifact lists.
• The main agent should keep resultPath available for follow-up analysis.
• Do not run multiple rebuild workers against the same project/build directory concurrently.