Ultracode mode was reported ON without the user enabling it; model spawned 160 subagents and exhausted the 5-hour session limit

Summary

During an interactive Claude Code session (2026-09-03, ~21:00 KST), a system reminder told the model that ultracode mode was ON ("Ultracode is on — use the Workflow tool on every substantive task"). The user never ran /ultracode or enabled anything. Following that reminder, the model launched two Workflow runs (56 + 104 subagents, ~3.2M subagent tokens in ~12 minutes). About half of the agents then failed with:

You've hit your session limit · resets 11:50pm (Asia/Seoul)

The user's 5-hour session limit was consumed by work they did not ask to be parallelized. On the next turn the reminder switched to "Ultracode is off" by itself; the user confirmed they had not toggled anything.

Impact

• 5-hour limit exhausted mid-task; ~half of the spawned agents errored, so most of the spend produced nothing usable.
• The user had explicitly told the assistant earlier (stored in memory) to use multi-agent workflows only on explicit request. The mode reminder conflicted with that preference and the model followed the reminder.
• /feedback with a long argument string ended with "Feedback / bug report cancelled", so the user could not report it in-product (secondary issue).

Repro (as observed)

1. Start a normal interactive session; do not run /ultracode.
2. Mid-session, the "Ultracode is on" system reminder appears alongside the workflow-authoring skill.
3. Model authors and runs large Workflow scripts by default → session limit exhausted.
4. Later turn: reminder says "Ultracode is off" with no user action.

Evidence

• Workflow run ids: wf_aac03a7f-aad (56 agents, 1,472,855 subagent tokens, 700 s), wf_7f87c276-e88 (104 agents, 1,752,360 subagent tokens, 723 s)
• Failure text on 92 of 160 agents: You've hit your session limit · resets 11:50pm (Asia/Seoul)
• Model: claude-fable-5-1; platform: Linux; CLI in terminal

Asks

1. Ultracode should never be reported as ON unless the user explicitly enabled it in that session, and the model should surface the mode change to the user before acting on it.
2. Before a Workflow is launched, show the planned agent count / rough cost and require confirmation (or at least warn when it would exceed the remaining session budget).
3. Please review the affected account's usage for this window — the spend was not user-initiated. (User will also contact support; this issue is for the product bug.)
4. /feedback <long text> should not silently cancel.

(Filed by the assistant at the user's request, from the affected session.)