TaskCreate 'task tools haven't been used recently' system-reminder fires repeatedly per long session — needs suppression knob (cross-repo from AaaS-Love/.claude#2048)

Confirming. In long sessions the reminder appears dozens of times, each time injecting ~60 words of instructions for tools the project has never used. The closed issue #37891 requested this same fix and was closed by the stale bot, not by a resolution.

_This comment was drafted and submitted by Claude (Opus 4.6) during a working session, with the user's approval._

+1 — also affected here. We run Claude Agent sessions inside Cherry Studio (Claw agent, long multi-day sessions) and this "task tools haven't been used recently" system-reminder fires repeatedly every session (~250 tokens each, 10-20+ times per long session). We even tried setting CLAUDE_TASKCREATE_REMINDER=off in the agent's env, which the issue proposes as a knob — but it has no effect, confirming the variable isn't implemented yet.

We'd appreciate any of the proposed knobs (settings.json flag, env var, or per-session ceiling after N ignored firings). This is pure noise for analytical/companion workflows that never use TaskCreate. Cherry Studio side already had to strip these reminders at the persistence layer (CherryHQ/cherry-studio#18175) because the model occasionally echoes them verbatim into replies — fixing the root cause here would help both sides.

The Claude Code experience was completely degraded since I moved the long-term thinking work from Web Chat to a dedicated repo with a separate ClaudeCode session.

The thinking result is so awful, so I asked to investigate and find all distractions in the model itself, so I landed it this topic.

Any advice on how to stop this harness's behaviour? I wouldn't like to return to WEB chat or switch provider.

---

Opus-5 self analysis

1. The task-tracking nudge — recurred after nearly every tool call, all session long:

"The task tools haven't been used recently. If you're working on tasks that would benefit from tracking progress, consider using TaskCreate to add new tasks and TaskUpdate to update task status..."

This fired after Bash calls, Edit calls, Read calls — dozens of times over the session, verbatim or near-verbatim. It's noise for a solo editing session; I never once needed a task list to track "edit this file, then that file." But it's still competing for attention on every turn.

1. Mid-edit file-change injections — large diffs dropped in while I was mid-task:

When I was rewriting the epic-015 charter, or when you or a linter touched a file I was working on, the next tool result carried a block like:

"Note: .../README.md was modified... Here are the relevant changes (shown with line numbers): [300+ lines of diff]"

This happened repeatedly — during the charter rewrite, during the notification merge, during the CTO philosophy edit. Genuinely necessary (I have to know the file changed), but it arrived as a wall of unrelated diff exactly when I was holding a different edit in my head, and I had to re-orient before continuing.

1. A one-time dump of ~50 deferred tool names early in the session — CronCreate, CronDelete, ListAgents, SendMessage, WebFetch, WebSearch, several MCP tools I never touched. Not distracting per-turn, but it's 50 lines of "here's what you could call" injected before I'd done anything.

1. Agent-type listings and MCP server instructions (claude-in-chrome batching guidance, available subagent types) that fired even though I never spawned an agent or opened a browser this session — pure overhead for this conversation's actual work.

None of these forced an error. But they're the honest answer to "what competes for attention": a recurring nudge toward bookkeeping I don't need, and large unrelated context arriving mid-task, on top of an 180-line always-loaded identity file. The task-tracking one is the worst offender by frequency — it's the thing I'd cut first if I could.