fix docker e2e

.github/workflows/ci-docker-e2e.yml

@@ -9,6 +9,10 @@ env:
   CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
   SKITTER_LLM_API_KEY: ${{ secrets.SKITTER_LLM_API_KEY }}
   SKITTER_LLM_MODEL: ${{ vars.SKITTER_LLM_MODEL || 'claude-haiku-4-5' }}
+  # CI runners are slower than dev machines; large prompts to the Claude CLI can
+  # take well over the 30s default before the agent emits its first artifact.
+  # Keep below the 120s outer timeout in send_and_collect to avoid racing it.
+  SKITTER_STREAM_IDLE_TIMEOUT: "90"
 
 jobs:
   docker-e2e:

skitter/agent_runner.py

@@ -114,34 +114,34 @@ def _build_cli_cmd(
             if agent.instructions:
                 cmd.extend(["-c", f"developer_instructions={agent.instructions}"])
             cmd.append(prompt)
+        return cmd
+
+    full_prompt = prompt
+    if agent.instructions:
+        full_prompt = f"{agent.instructions}\n\n{prompt}"
+    cmd = [
+        "claude",
+        "-p",
+        full_prompt,
+        "--output-format",
+        "stream-json",
+        "--verbose",
+    ]
+    if resume_id:
+        cmd.extend(["--resume", resume_id])
     else:
-        full_prompt = prompt
-        if agent.instructions:
-            full_prompt = f"{agent.instructions}\n\n{prompt}"
-        cmd = [
-            "claude",
-            "-p",
-            full_prompt,
-            "--output-format",
-            "stream-json",
-            "--verbose",
-        ]
-        if resume_id:
-            cmd.extend(["--resume", resume_id])
-        elif _PERMISSION_MODE == "bypassPermissions":
-            cmd.extend(
-                ["--dangerously-skip-permissions", "--settings", _SANDBOX_SETTINGS]
-            )
-        else:
-            cmd.extend(
-                ["--permission-mode", _PERMISSION_MODE, "--settings", _SANDBOX_SETTINGS]
-            )
-        if agent.model:
-            cmd.extend(["--model", agent.model])
-        if agent.max_turns:
-            cmd.extend(["--max-turns", str(agent.max_turns)])
-        if agent.tools:
-            cmd.extend(["--allowedTools", ",".join(agent.tools)])
+        # Use --permission-mode rather than --dangerously-skip-permissions even for
+        # bypassPermissions: the latter triggers a Claude CLI hang on large inputs
+        # combined with --output-format stream-json.
+        cmd.extend(
+            ["--permission-mode", _PERMISSION_MODE, "--settings", _SANDBOX_SETTINGS]
+        )
+    if agent.model:
+        cmd.extend(["--model", agent.model])
+    if agent.max_turns:
+        cmd.extend(["--max-turns", str(agent.max_turns)])
+    if agent.tools:
+        cmd.extend(["--allowedTools", ",".join(agent.tools)])
     return cmd
 
 

tests/conftest.py

@@ -8,16 +8,36 @@
 
 import asyncio
 import json
+import os
 import socket
 import subprocess
 import sys
+import tempfile
 import uuid
 from pathlib import Path
 
-import aiomqtt
-import pytest
-
-from skitter.a2a import (
+# Isolate tests from the developer's ~/.skitter/ config and from MQTT_*/SKITTER_*
+# values uv and skitter.mqtt pull out of the project .env. Point SKITTER_HOME at
+# a throwaway dir and blank the relevant env vars so load_config() falls back to
+# built-in defaults (localhost:1883, no auth, org=skitter unit=default), matching
+# what CI sees on a fresh runner. Set to empty string rather than popping so
+# skitter.mqtt's load_dotenv() call at import time leaves them alone (dotenv
+# respects already-set vars). Must happen before importing anything from skitter.*.
+os.environ["SKITTER_HOME"] = tempfile.mkdtemp(prefix="skitter-test-")
+for _var in (
+    "MQTT_BROKER_URL",
+    "MQTT_USERNAME",
+    "MQTT_PASSWORD",
+    "MQTT_CA_CERT",
+    "SKITTER_A2A_ORG",
+    "SKITTER_A2A_UNIT",
+):
+    os.environ[_var] = ""
+
+import aiomqtt  # noqa: E402 — after env scrubbing above
+import pytest  # noqa: E402
+
+from skitter.a2a import (  # noqa: E402
     a2a_org,
     a2a_unit,
     A2ARequest,
@@ -29,8 +49,8 @@
     topic_discovery,
     topic_reply,
 )
-from skitter.config import load_config as _load_config
-from skitter.mqtt import mqtt_client_kwargs
+from skitter.config import load_config as _load_config  # noqa: E402
+from skitter.mqtt import mqtt_client_kwargs  # noqa: E402
 
 PROJECT_ROOT = Path(__file__).parent.parent
 

tests/test_acceptance.py

@@ -163,6 +163,9 @@ def mock_codex_bin(tmp_home):
 def skitter_env(tmp_home, mock_claude_bin, mock_codex_bin):
     """Subprocess environment with isolated HOME and mock CLIs on PATH."""
     env = os.environ.copy()
+    # conftest pins SKITTER_HOME for in-process test code; subprocess tests here
+    # want skitter to derive its home from HOME, so drop the inherited override.
+    env.pop("SKITTER_HOME", None)
     env["HOME"] = str(tmp_home)
     env["PATH"] = f"{mock_claude_bin}:{mock_codex_bin}:{env.get('PATH', '')}"
     env["NO_COLOR"] = "1"

tests/test_docker_e2e.py

@@ -97,6 +97,9 @@
 def host_env(tmp_path):
     """Fresh isolated environment for each test that runs skitter CLI commands."""
     env = os.environ.copy()
+    # conftest pins SKITTER_HOME for in-process test code; subprocess tests here
+    # want skitter to derive its home from HOME, so drop the inherited override.
+    env.pop("SKITTER_HOME", None)
     env["HOME"] = str(tmp_path)
     env["MQTT_BROKER_URL"] = "mqtt://localhost:1883"
     env["NO_COLOR"] = "1"

tests/unit/test_agent_runner.py

@@ -161,8 +161,10 @@ def test_build_claude_cmd(self):
         assert "--agent" not in cmd
         assert "--model" in cmd
         assert "sonnet" in cmd
-        assert "--dangerously-skip-permissions" in cmd
-        assert "--permission-mode" not in cmd
+        # Use --permission-mode bypassPermissions: --dangerously-skip-permissions
+        # triggers a Claude CLI hang on large + stream-json inputs.
+        assert "--dangerously-skip-permissions" not in cmd
+        assert cmd[cmd.index("--permission-mode") + 1] == "bypassPermissions"
 
     def test_build_codex_cmd(self):
         from skitter.agent_runner import _build_cli_cmd