# Bug Report — Scheduled automation runs fail with API 400 ("tool" role error)

[Arjunan323]

Reporter: Meesho supplier automation user (chat 1503962408)
Date: 2026-09-10 (IST)
Severity: High — scheduled runs fail silently; platform reports them as OK
Status: Recurring (2 consecutive scheduled runs, 2 different tasks)

---

Error (exact)

Error code: 400 - {'error': {'message': "Messages with role 'tool' must be a response to a preceding message with 'tool_calls'", 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}

What happens

A scheduled (cron) task run starts, then aborts mid-execution with the error above.
The run produces no output artifacts (no report file, no Telegram message), yet the
task's last_status is recorded as "ok" and run_count increments — so there is no
visible failure signal other than the missing outputs.

Recurrence — both affected runs

#	Task	Schedule	First run?	When (IST)	Result
1	Meesho daily pending-orders accept	daily 23:00	No (had run ok before manually-created)	2026-09-09 23:00	400 error, no outputs
2	Meesho Label Download — Every 4 Hours	0 */4 * * *	Yes — task created ~25 min earlier, never ran before	2026-09-10 00:00	400 error, no outputs

Key observation: run #2 was the first ever run of a brand-new task — so the failure is
not caused by long-running state, accumulated context, or the task's instructions.

What works

Interactive (chat) sessions running the same steps against the same website complete
successfully every time (same tools: browser automation via Puppeteer MCP, file writes,
send_message). The failure is specific to the scheduled-run execution path.

Hypothesis for the platform team

The scheduled-run engine compacts/rewrites the agent's message history mid-run; afterwards a
tool role message remains that no longer follows a matching tool_calls message, and the
model API rejects the request with this 400. Additionally, the runner does not propagate
this failure into last_status (reports "ok" instead of an error).

Requested fix

1. Prevent tool-role messages from surviving context compaction without their tool_calls
   parent (drop or re-pair them before sending to the API).
2. Record genuinely failed runs as failed (last_status error) so users/alerts can detect
   missed runs instead of discovering them from missing outputs.

Current state (so support can triage)

• Tasks affected: task-272f869009 (Label Download — Every 4 Hours), task-ac6a7c2f06
  (Meesho daily pending-orders accept)
• Both schedules remain enabled and will keep failing until the runner is fixed.
• Business impact so far: 2 missed runs, recovered manually (~2 min each); no orders lost.