0d95b5098a
The 2026-04-30 staging incident traced back to workspaces booting with ANTHROPIC_BASE_URL pointing at a non-Anthropic shim (MiniMax / OpenAI gateway) but no explicit model configured. The adapter silently fell back to "sonnet" — an Anthropic-native alias the upstream didn't recognize — and the SDK --print probe hung 30s before timing out. Platform's phantom-busy sweep then nuked the workspace at 10min, producing "every workspace dead" with the root cause buried in a 30s subprocess hang. Pre-validate the combo at adapter boot: when ANTHROPIC_BASE_URL host is non-Anthropic AND no explicit model is set, raise ValueError with an actionable message pointing to MODEL_PROVIDER / runtime_config.model. Also log the resolved model + base_url_host every boot so future failures explain themselves in the workspace logs without digging into the SDK subprocess. Tests live under tests/ with their own pytest.ini that anchors rootdir there — keeps pytest from importing the package __init__.py (which does the runtime-discovery relative import that requires molecule_runtime installed). 7 tests cover: misconfig raises with the right message, Anthropic-native passes, no-base-url passes, custom-url + explicit model passes, dataclass + dict shapes, unparseable URL no-crash. CI runs them on every push/PR. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
268 lines
12 KiB
Python
268 lines
12 KiB
Python
"""Claude Code adapter — wraps the Claude Code CLI as an agent runtime."""
|
|
|
|
import json
|
|
import os
|
|
import logging
|
|
from pathlib import Path
|
|
|
|
from molecule_runtime.adapters.base import BaseAdapter, AdapterConfig, RuntimeCapabilities
|
|
from a2a.server.agent_execution import AgentExecutor
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# Cap one transcript response at 1000 lines so a paranoid client can't OOM
|
|
# the workspace by polling /transcript?limit=999999.
|
|
_TRANSCRIPT_MAX_LIMIT = 1000
|
|
|
|
|
|
class ClaudeCodeAdapter(BaseAdapter):
|
|
|
|
@staticmethod
|
|
def name() -> str:
|
|
return "claude-code"
|
|
|
|
@staticmethod
|
|
def display_name() -> str:
|
|
return "Claude Code"
|
|
|
|
@staticmethod
|
|
def description() -> str:
|
|
return "Claude Code CLI — full agentic coding with hooks, CLAUDE.md, auto-memory, and MCP support"
|
|
|
|
@staticmethod
|
|
def get_config_schema() -> dict:
|
|
return {
|
|
"model": {"type": "string", "description": "Claude model (e.g. sonnet, opus, haiku)", "default": "sonnet"},
|
|
"required_env": {"type": "array", "description": "Required env vars", "default": ["CLAUDE_CODE_OAUTH_TOKEN"]},
|
|
"timeout": {"type": "integer", "description": "Timeout in seconds (0 = no timeout)", "default": 0},
|
|
}
|
|
|
|
def capabilities(self) -> RuntimeCapabilities:
|
|
"""Claude-code SDK owns session lifecycle natively — see project
|
|
memory `project_runtime_native_pluggable.md`.
|
|
|
|
provides_native_session=True
|
|
The claude-agent-sdk maintains a long-lived streaming session
|
|
with its own ClaudeSDKClient state. The platform's a2a_queue
|
|
would double-buffer the same in-flight state — declaring
|
|
native_session lets the platform skip enqueueing and dispatch
|
|
directly. Validates capability primitive #5 once that
|
|
consumer lands.
|
|
|
|
Other capabilities stay False (platform fallback owns them):
|
|
- provides_native_heartbeat: the SDK's session events don't map
|
|
cleanly to our 30s heartbeat cadence; heartbeat.py keeps
|
|
emitting WORKSPACE_HEARTBEAT so the canvas idle indicator and
|
|
a2a_proxy idle-timer reset behavior keep working.
|
|
- provides_native_scheduler: claude-code has no built-in cron;
|
|
platform scheduler keeps owning it.
|
|
- provides_native_status_mgmt: claude-code wedge detection IS
|
|
adapter-driven (claude_sdk_executor sets is_wedged + heartbeat
|
|
forwards as runtime_state="wedged"), but the rest of the
|
|
status state machine (online/degraded recovery via error_rate)
|
|
stays platform-owned. Reconsider once the SDK exposes its own
|
|
ready-signal hook.
|
|
- provides_native_retry / activity_decoration / channel_dispatch:
|
|
not implemented in the SDK surface — platform fallback applies.
|
|
"""
|
|
return RuntimeCapabilities(
|
|
provides_native_session=True,
|
|
)
|
|
|
|
def idle_timeout_override(self) -> int:
|
|
"""Claude-code synthesis on Opus + multi-step tool use legitimately
|
|
runs 8-10 min between broadcaster events. The pre-capability
|
|
bug PR #2128 patched at the env-var layer hit this exact issue:
|
|
`context canceled` mid-flight when the platform's 5min idle
|
|
timer fired during a long packaging step. Override to 15 min
|
|
to cover the long tail without leaving genuinely-wedged runs
|
|
hanging too long.
|
|
|
|
Capability primitive #2 — see workspace/adapter_base.py:
|
|
idle_timeout_override and PR #2139 for the platform-side
|
|
consumer in a2a_proxy.dispatchA2A.
|
|
"""
|
|
return 900 # 15 minutes
|
|
|
|
async def setup(self, config: AdapterConfig) -> None:
|
|
"""Install plugins via the per-runtime adaptor registry.
|
|
|
|
The legacy claude-code-specific ``inject_plugins()`` override is gone:
|
|
each plugin now ships (or has registered in the platform registry) a
|
|
per-runtime adaptor, and ``BaseAdapter.install_plugins_via_registry``
|
|
routes installs through it. The Claude Code SDK still reads
|
|
``CLAUDE.md`` and ``/configs/skills/`` natively, and the default
|
|
:class:`AgentskillsAdaptor` writes to both.
|
|
"""
|
|
from molecule_runtime.plugins import load_plugins
|
|
workspace_plugins_dir = os.path.join(config.config_path, "plugins")
|
|
plugins = load_plugins(
|
|
workspace_plugins_dir=workspace_plugins_dir,
|
|
shared_plugins_dir=os.environ.get("PLUGINS_DIR", "/plugins"),
|
|
)
|
|
await self.install_plugins_via_registry(config, plugins)
|
|
|
|
async def create_executor(self, config: AdapterConfig) -> AgentExecutor:
|
|
from claude_sdk_executor import ClaudeSDKExecutor
|
|
|
|
# Load system prompt if exists
|
|
system_prompt = config.system_prompt
|
|
if not system_prompt:
|
|
prompt_file = os.path.join(config.config_path, "system-prompt.md")
|
|
if os.path.exists(prompt_file):
|
|
with open(prompt_file) as f:
|
|
system_prompt = f.read()
|
|
|
|
# runtime_config may arrive as a dict (from main.py vars(...)) or as a
|
|
# RuntimeConfig dataclass. Read `model` defensively from either shape.
|
|
rc = config.runtime_config
|
|
if isinstance(rc, dict):
|
|
explicit_model = rc.get("model") or ""
|
|
else:
|
|
explicit_model = getattr(rc, "model", None) or ""
|
|
|
|
# Pre-validation: detect the misconfiguration combo that drove the
|
|
# 2026-04-30 staging incident — ANTHROPIC_BASE_URL pointed at a
|
|
# non-Anthropic upstream (MiniMax / OpenAI shim) but no explicit
|
|
# model was set, so we'd silently fall back to "sonnet" and the
|
|
# upstream would hang on the SDK --print probe for 30s before
|
|
# timing out. The platform's phantom-busy sweep then resets the
|
|
# workspace at the 10min mark — the user-visible failure is "every
|
|
# workspace dead" but the root cause is one missing env var.
|
|
#
|
|
# Fail fast here with an actionable message so the operator sees
|
|
# exactly what to fix instead of chasing ghosts in workspace logs.
|
|
# We only fire when ALL three are true:
|
|
# 1. ANTHROPIC_BASE_URL is set (custom upstream is in play)
|
|
# 2. The host is NOT api.anthropic.com (real Anthropic accepts
|
|
# "sonnet" as a known alias, so the fallback is fine there)
|
|
# 3. The user did NOT set an explicit model (the check we want)
|
|
# Anthropic-native users with no model picked still get the
|
|
# "sonnet" fallback — that's correct behavior, no error.
|
|
base_url = os.environ.get("ANTHROPIC_BASE_URL", "").strip()
|
|
if base_url and not explicit_model:
|
|
from urllib.parse import urlparse
|
|
host = urlparse(base_url).hostname or ""
|
|
if host and host != "api.anthropic.com":
|
|
raise ValueError(
|
|
"claude-code adapter: ANTHROPIC_BASE_URL points at a "
|
|
f"non-Anthropic host ({host}) but no model is configured. "
|
|
"The default fallback ('sonnet') is an Anthropic-native "
|
|
"alias; non-Anthropic shims (MiniMax, OpenAI gateways, "
|
|
"etc.) won't recognize it and the SDK --print probe will "
|
|
"hang for 30s before timing out. Fix: set MODEL_PROVIDER "
|
|
"as a workspace secret (canvas: Save+Restart with model "
|
|
"picked) or set runtime_config.model in /configs/config.yaml."
|
|
)
|
|
|
|
model = explicit_model or "sonnet"
|
|
# Surface what we resolved to in logs — when the workspace agent
|
|
# eventually fails, this single line in the logs explains "the
|
|
# adapter sent X to Y" without having to dig into the SDK
|
|
# subprocess. Cheap diagnostic, no runtime cost.
|
|
if base_url:
|
|
from urllib.parse import urlparse
|
|
logger.info(
|
|
"claude-code: model=%s base_url_host=%s%s",
|
|
model,
|
|
urlparse(base_url).hostname or "<unparseable>",
|
|
" (custom upstream)" if base_url else "",
|
|
)
|
|
else:
|
|
logger.info("claude-code: model=%s base_url=anthropic-default", model)
|
|
|
|
return ClaudeSDKExecutor(
|
|
system_prompt=system_prompt,
|
|
config_path=config.config_path,
|
|
heartbeat=config.heartbeat,
|
|
model=model,
|
|
)
|
|
|
|
async def transcript_lines(self, since: int = 0, limit: int = 100) -> dict:
|
|
"""Read the live Claude Code session transcript.
|
|
|
|
Claude Code writes every session to
|
|
``$HOME/.claude/projects/<cwd-as-dirname>/<session-uuid>.jsonl`` —
|
|
every line is a JSON event (user/assistant/tool_use/attachment/etc).
|
|
We pick the most-recently-modified .jsonl in the projects dir for
|
|
the agent's working directory, then return ``[since:since+limit]``.
|
|
|
|
Returns ``supported: True`` even if no .jsonl exists yet (empty
|
|
``lines`` + ``cursor=0``) so the canvas can show "agent hasn't
|
|
produced output yet" instead of "feature unavailable".
|
|
"""
|
|
limit = max(1, min(limit, _TRANSCRIPT_MAX_LIMIT))
|
|
since = max(0, since)
|
|
|
|
# Resolve the projects-dir name. Claude Code maps cwd → dirname by
|
|
# replacing "/" with "-" (so "/configs" → "-configs"). The exact
|
|
# rule lives inside the CLI binary, but the leading-dash + path-
|
|
# without-trailing-slash pattern is stable across versions.
|
|
#
|
|
# Match ClaudeSDKExecutor._resolve_cwd: prefer /workspace if populated,
|
|
# else /configs. Override via CLAUDE_PROJECT_CWD for tests.
|
|
WORKSPACE_MOUNT = "/workspace"
|
|
CONFIG_MOUNT = "/configs"
|
|
cwd_override = os.environ.get("CLAUDE_PROJECT_CWD")
|
|
if cwd_override:
|
|
cwd = cwd_override
|
|
elif os.path.isdir(WORKSPACE_MOUNT) and os.listdir(WORKSPACE_MOUNT):
|
|
cwd = WORKSPACE_MOUNT
|
|
else:
|
|
cwd = CONFIG_MOUNT
|
|
|
|
# Normalize: strip trailing slash, replace path separators with "-"
|
|
cwd_norm = cwd.rstrip("/") or "/"
|
|
projdir_name = cwd_norm.replace("/", "-") # "/configs" → "-configs"
|
|
|
|
home = Path(os.environ.get("HOME", "/home/agent"))
|
|
projdir = home / ".claude" / "projects" / projdir_name
|
|
result_base = {
|
|
"runtime": self.name(),
|
|
"supported": True,
|
|
"lines": [],
|
|
"cursor": since,
|
|
"more": False,
|
|
"source": str(projdir),
|
|
}
|
|
|
|
if not projdir.is_dir():
|
|
return result_base
|
|
|
|
# Pick most-recently-modified .jsonl
|
|
candidates = sorted(projdir.glob("*.jsonl"), key=lambda p: p.stat().st_mtime, reverse=True)
|
|
if not candidates:
|
|
return result_base
|
|
target = candidates[0]
|
|
result_base["source"] = str(target)
|
|
|
|
lines = []
|
|
more = False
|
|
try:
|
|
with target.open("r") as f:
|
|
for i, raw in enumerate(f):
|
|
if i < since:
|
|
continue
|
|
if len(lines) >= limit:
|
|
more = True
|
|
break
|
|
raw = raw.strip()
|
|
if not raw:
|
|
continue
|
|
try:
|
|
lines.append(json.loads(raw))
|
|
except json.JSONDecodeError:
|
|
# Skip malformed lines but keep cursor advancing
|
|
lines.append({"_parse_error": True, "_raw": raw[:200]})
|
|
except OSError as exc:
|
|
logger.warning("transcript_lines: read failed for %s: %s", target, exc)
|
|
return result_base
|
|
|
|
result_base["lines"] = lines
|
|
result_base["cursor"] = since + len(lines)
|
|
result_base["more"] = more
|
|
return result_base
|
|
|
|
|
|
Adapter = ClaudeCodeAdapter
|