菜单
Spawn and exercise the native `claude-code` runtime against a test server, including the model names that work and the synthetic-tool-result gotcha. Run with /test-claude-runtime. Pairs with /launch-grackle.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Spawn the ACP-variant runtimes (claude-code-acp / codex-acp / copilot-acp). claude-code-acp validated; it surfaces real tool failures that native claude-code can't. Run with /test-acp-runtime. Pairs with /launch-grackle.
Spawn the `codex` runtime against a test server. Use model `gpt-5.5` (requires Codex SDK >= 0.135.0). Run with /test-codex-runtime. Pairs with /launch-grackle.
Spawn and exercise the `copilot` runtime against a test server. CRITICAL: gpt-4o does NOT work — use claude-sonnet-4.5. Run with /test-copilot-runtime. Pairs with /launch-grackle.
Investigates code bugs and files GitHub issues.
Pure coordinator that decomposes work and delegates to subagents.
Verifies a PR is fully mergeable (CI green, no unresolved comments, approved) then merges it.
| name | test-claude-runtime |
| description | Spawn and exercise the native `claude-code` runtime against a test server, including the model names that work and the synthetic-tool-result gotcha. Run with /test-claude-runtime. Pairs with /launch-grackle. |
How to spawn the native claude-code runtime against an isolated test server and observe its behavior. Assumes you've started a server with /launch-grackle (and have GRACKLE_URL + GRACKLE_API_KEY exported).
| Model | Works? | Notes |
|---|---|---|
sonnet | ✅ | Default persona (claude-code / "Software Engineer") uses this |
opus | ✅ | |
haiku | ✅ |
No model gating — native Claude Code works out of the box (subscription/OAuth). These are the values grackle runtimes advertises and they all spawn.
The default persona already uses claude-code:
grackle spawn local "<prompt>" # uses default persona (claude-code / sonnet)
Or make an explicit persona (e.g. to pick opus):
grackle persona create "Claude Opus Tester" --runtime claude-code --model opus --prompt "You are a test agent."
grackle spawn local "<prompt>" --persona claude-opus-tester
The Claude Agent SDK runs tools internally and (in streaming mode) does not emit real tool_result blocks with is_error. The adapter instead emits synthetic empty tool_result events that are always successful (flushPendingToolResults → content: "", is_error: false).
Verified live: prompting Claude Code to run a failing command (cat /nonexistent) produced tool_result events with content: {"is_ok":true,...} (empty) — the failure was not observable. This is the separate bug 1c (filed off #1355).
Implication: you cannot demonstrate a tool failure through native claude-code. The adapter's b.is_error read (packages/runtime-claude-code/src/claude-code.ts) is only exercised when the SDK does surface tool_result blocks (some non-streaming modes). To observe a real Claude tool failure, use the ACP variant (see /test-acp-runtime — now that #1366 is fixed) — its tool_call_update carries a real status: failed (validated live).
Tool outcomes land in the session's JSONL log:
$GRACKLE_HOME/.grackle/logs/<session-id>/stream.jsonl
Each tool_result event has content = {is_ok, content, past_tense_message}, and failures additionally carry a first-class "tool_error":true (AHP #1362). Success omits tool_error (proto3 drops default-false).
/launch-grackle — start the isolated server first/test-copilot-runtime, /test-codex-runtime, /test-acp-runtime — sibling runtimes (and their very different model-name rules)