molecule-core/docs/adapters/medo-smoke-test-log.md

MeDo Smoke Test Log — 2026-04-13 (Run 4)

Tester: PM (direct execution)
Goal: Install Miaoda App Builder skill → build "Hello Molecule AI" landing page → publish → URL.
Credits spent: 0 across all four runs.

---

Run Summary

Run	Blocker	Resolution
1	workspace-template:openclaw image not built	✅ Operator rebuilt image
2	Adapter key lookup ignores AISTUDIO_API_KEY / QIANFAN_API_KEY	✅ Code fix committed (d779e16)
3	Executor creates fresh OpenClaw session per A2A message	✅ Code fix committed (9466943)
4	payloads: [] on every response — agent never returns text via --json mode	❌ Root cause below

---

Run 4 — Detailed Findings

Environment — all green

Check	Result
Platform health	✅
workspace-template:openclaw image	✅ boots in 31s
AISTUDIO_API_KEY + gemini-2.0-flash	✅ confirmed in every response meta
Stable session ID (workspace ID)	✅ sessionKey: agent:main:explicit:a507780d-... consistent across all calls

Messages Sent and Responses

Message	Response	Duration
Install skill	payloads: [], livenessState: working	1.7s
Build Hello Molecule AI	payloads: [], livenessState: working	0.8s
Check status (sessions_list)	LLM request failed: provider rejected request schema/payload	—
Reply with exactly: STATUS_OK	payloads: [], livenessState: working (after restart)	1.8s

The "Reply with exactly: STATUS_OK" response is decisive. A vanilla LLM call with no tool use should produce a text payload. It didn't. This rules out skill complexity or message ambiguity as the cause.

Root Cause — openclaw agent --json Does Not Surface Agent Text in payloads

The OpenClaw agent processes messages using background session dispatch (sessions_spawn / sessions_yield). In this mode:

1. Main session receives message → immediately spawns background session → calls sessions_yield
2. openclaw agent --json exits with payloads: [], livenessState: 'working'
3. Background session processes the actual work and produces text — but only visible in interactive/streaming mode, not in the --json subprocess call

Evidence: Even "Reply with exactly: STATUS_OK" returns payloads: []. The agent is using background sessions for everything, including trivial echo requests.

Likely cause: OpenClaw's default SOUL.md / BOOTSTRAP.md workspace config instructs the agent to always use async session patterns. In a terminal session these background responses appear naturally; via subprocess --json, only the main session's synchronous output is captured.

Transient issue: LLM request failed

After 3+ rapid A2A calls (install → build → status check), the Gemini AI Studio API returned a schema/payload rejection. Resolved by restarting the workspace (POST /workspaces/:id/restart). Likely a rate-limit or context-size rejection from Gemini. Restarted in 30s, normal on next call.

---

4. Required Fix — OpenClawA2AExecutor Response Capture

The executor must retrieve the agent's text response from session history after the main session yields. The sessions_history CLI command (exposed as session_history tool) retrieves past messages.

Proposed change to workspace/adapters/openclaw/adapter.py (execute() method):

# After proc.communicate() returns with payloads=[]:
if not reply or reply.startswith("{'payloads': []"):
    # Agent yielded without responding — fetch last message from session history
    await asyncio.sleep(2)  # brief wait for background session to complete short tasks
    hist_proc = await asyncio.create_subprocess_exec(
        "openclaw", "sessions", "history",
        "--session-id", self._session_id,
        "--limit", "1", "--json",
        stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE,
        env={**os.environ, "PATH": f"{os.path.expanduser('~/.local/bin')}:{os.environ.get('PATH', '')}"}
    )
    hist_stdout, _ = await asyncio.wait_for(hist_proc.communicate(), timeout=15)
    hist_data = json.loads(hist_stdout.decode().strip() or "{}")
    last_msg = (hist_data.get("messages") or [{}])[-1]
    reply = last_msg.get("content", reply)  # fall back to original if no history

Note on long tasks (5–8 min builds): Session history won't have the build result until it completes. For Miaoda App Builder, PM must poll: send a follow-up "What is the status of the Hello Molecule AI app build?" message every 60s until the response contains a URL or error.

---

5. Open Questions Status

5-C — Rate limits

UNKNOWN. Never reached skill invocation.
New data: Gemini AI Studio hit a schema/payload rejection after 3 rapid calls. This may be a Gemini-specific issue with large tool schemas (OpenClaw's cron schema is 6311 chars). Worth filing separately.

5-D — Failure recovery

UNKNOWN. Never reached app generation.

---

6. Issues to File

#	Issue	Status	Location
A	fix(openclaw): use stable workspace session ID	✅ fixed in 9466943	adapter.py
B	fix(openclaw): extend key lookup for AISTUDIO/QIANFAN	✅ fixed in d779e16	adapter.py
C	fix(provisioner): surface Docker errors in last_sample_error	❌ open	provisioner.go
D	fix(openclaw): capture agent response via session history when payloads=[]	❌ open — see §4	adapter.py
E	fix(openclaw): Gemini rejects request after N rapid calls with large tool schema	❌ open — investigate cron schema size	adapter.py

---

7. Next Steps (before Run 5)

• Dev Lead: Implement §4 session-history fallback in OpenClawA2AExecutor.execute()
• Dev Lead (optional): Trim cron tool schema to reduce Gemini schema-size rejection risk
• Operator: Rebuild image: bash workspace/build-all.sh openclaw
• PM (Run 5): Re-run smoke test — expected to finally reach skill install confirmation