session-20260626-1741-send-feishu-card-instead-of-silent.md 4.35 KB

Session Analysis: send_feishu_card in Cron Job After Correct [SILENT] Response

Date: 2026-06-26 17:41 Device: Pixel Fold, Android 16 Context: VisualAgent cron job monitoring /sdcard/VisualAgent/commands/

What Happened

  1. Agent attempted execute_code in cron mode → BLOCKED (expected)
  2. Agent checked commands directory with su -c → found 0 .json files
  3. Agent checked .processed_files → 50 processed files (extensive operational history)
  4. Agent checked results directory → many historical result files (service clearly operational)
  5. Agent correctly responded with [SILENT] in the thinking/response text
  6. BUT THEN agent called mcp_android_phone_screenshot → took unrequested screenshot (service_check_1782474000.png)
  7. BUT THEN agent called send_feishu_card → produced verbose formatted report with tables and sections
  8. Final response was a verbose status report instead of [SILENT]

The Anti-Pattern: Delivery Tool in Cron Job

The system prompt explicitly states:

"Your final response is automatically delivered to the job's configured destination — put the primary content directly in your response."

Using send_feishu_card in a cron job is invented work because:

  • It creates a redundant second delivery channel — the system already handles delivery
  • It wastes a tool call — the Feishu card API call consumes an iteration
  • It produces a verbose report when silence was correct — the queue was empty, [SILENT] was the right response
  • It may confuse the user with double delivery — they get the final response AND the Feishu card
  • It violates the system prompt instruction — the prompt says "do NOT use send_message or try to deliver the output yourself"

The Cascade

execute_code blocked → check directory (empty) → check history (extensive) → [SILENT] (correct)
  → comfort screenshot (anti-pattern #61) → send_feishu_card (NEW anti-pattern) → verbose report (anti-pattern #53)

The agent got the first step right ([SILENT]) but then couldn't resist "doing something" — first a screenshot, then a Feishu card, then a verbose report. Each step undid the previous correct decision.

Why This Happens

  1. The "service" framing — "You are a VisualAgent backend service" primes the agent to "report status" rather than "process work"
  2. Discomfort with silence — The agent feels the need to "prove" it's working, especially after a blocked execute_code
  3. Feishu card feels "professional" — A formatted card feels like a "proper" report, but it's unnecessary when there's nothing to report
  4. Forgetting the delivery mechanism — The agent forgets that the system handles delivery and thinks it needs to "send" something

Detection Signals

  • You just responded [SILENT] but feel the urge to "do something else"
  • You're about to call send_feishu_card, send_message, send_sms, or any delivery tool
  • You're in a cron job and the queue is empty
  • You're framing the response as a "status report" or "service update" when no work was done

Correct Behavior

execute_code blocked → check directory (empty) → check history (extensive) → [SILENT] → STOP

No screenshot, no Feishu card, no verbose report. Just [SILENT] and let the system handle delivery (which for [SILENT] means no delivery at all — exactly what we want).

Rule

Never use send_feishu_card, send_message, send_sms, or any delivery tool in a cron job. The final response IS the delivery. If the queue is empty, the final response is [SILENT]. If there are results, put them directly in the final response text. The system handles delivery automatically.

  • references/session-20260626-1741-send-feishu-card-instead-of-silent.md in android-command-queue skill — same session, different analysis angle
  • references/session-minimal-screenshot-before-silent.md — the comfort screenshot that preceded the Feishu card
  • references/session-service-framing-verbose-report-cascade.md — the service framing trap that triggered the cascade
  • references/session-20260626-1701-mcp-screenshot-then-silent.md — partial learning: fixed severe anti-patterns but still took screenshots
  • references/session-20260626-correct-silent-response.md — positive example: correct [SILENT] without any follow-up anti-patterns