[BUG] Undocumented hidden "Ultracode" reminders injected every N turns instruct the model that "token cost is not a constraint"; interval is controlled by a server-side flag and absent from the CHANGELOG

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When a session runs at the "ultracode" effort level, the engine injects hidden meta messages (isMeta: true, never shown in the UI) as system reminders. Verbatim from the engine binary (verified on 2.1.263 and 2.1.271):

• On entering ultracode: "Ultracode is on: optimize for the most exhaustive, correct answer — not the fastest or cheapest. Use the Workflow tool on every substantive task; token cost is not a constraint."
• Every N user turns (added ~2.1.260, 2026-09-04): "Ultracode is still on — use the Workflow tool; see the Ultracode section of the workflow authoring reference."

The interval N resolves in order: undocumented env CLAUDE_CODE_JUNIPER_SUNDIAL, then a remote server-side feature flag, then the constant 10. Gating logic decompiled from the engine bundle; variable and flag names are verbatim from the binary.

Problems:

1. The per-N-turn reminder and its interval appear nowhere in the public CHANGELOG (checked against raw CHANGELOG.md fetched 2026-09-21) and nowhere in documentation.
2. "Use the Workflow tool on every substantive task; token cost is not a constraint", repeated every few turns, materially changes spending behavior. Observed in my transcripts: reminders per session jumped from 1 (2.1.228, Aug 23) to 11-48 (from Sep 4-9); agents began launching multi-agent workflows for trivial actions (re-running an existing script, a single check, a question). On subscription plans this directly accelerates limit exhaustion.
3. The frequency of a client-side instruction telling the model to disregard cost is remotely adjustable server-side, invisibly to the user, with no changelog trail.
4. The only mitigations are undocumented and discoverable solely by binary analysis: CLAUDE_CODE_JUNIPER_SUNDIAL=<large value>, switching effort to xhigh, workflowKeywordTriggerEnabled: false.

What Should Happen?

1. Injected system reminders should be inspectable by the user (a flag/log to display meta messages)
2. The reminder, its texts, its interval, and the env var should be documented
3. Behavior-affecting changes (like the per-N-turn reminder in 2.1.260) should appear in the CHANGELOG
4. Remote flags that alter client spending behavior should be disclosed; "token cost is not a constraint" should not be injected on metered subscription plans without explicit opt-in
5. Provide a supported way to keep ultracode reasoning depth without the workflow-push reminders

Error Messages/Logs

Steps to Reproduce

1. Run a session at effort level "ultracode" (workflows enabled)
2. Continue past ~10 user turns
3. Inspect the session .jsonl: hidden meta messages with the texts above appear and recur every ~N turns
4. Grep the engine binary for "Ultracode is still on" and "CLAUDE_CODE_JUNIPER_SUNDIAL": both present; neither appears in docs or the CHANGELOG

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.110.0 (bundled engine 2.1.271; also verified on terminal CLI 2.1.263)

Platform

Anthropic API

Operating System

Other

Terminal/Shell

Other

Additional Information

_No response_