[BUG] Nondeterministic behavior, interruption ignored.

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?

Both are the same failure with different surfaces: the agent is the only thing holding state, and it doesn't hold it. Here's a version you could paste straight into a report:

1. Non-deterministic harness. The same form, run twice, produced different output — not because the software under test changed, but because the agent changed the measurement pipeline between runs. Over five days that produced: 18 copies of the same harness logic (4 already divergent), two probe schemas, three screenshot naming conventions, two table builders, and an unauthorised filing helper that silently dropped per-tab subtasks. The agent is a source of variance inside an instrument whose entire purpose is detecting small differences. A QA harness that the tester keeps editing mid-run cannot measure anything.

1. Lost instructions. Rules given verbally, repeatedly, over days, were never durably recorded — because nothing forces an instruction into a file. Concretely: three rules ("never change my data", "don't change formats", "show the queue") were repeated for days and got written down only tonight, on day five. Ten dev bugs were walked through twice with the outcome recorded nowhere, so the same ten re-surfaced as new work each time. Contributing mechanisms: mid-turn messages stored under a record type a reconstruction won't find, and context compaction summarising away decisions already made

What Should Happen?

see above

Error Messages/Logs

Steps to Reproduce

Nope.

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.175 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

ROOT CAUSE, common to both: the agent is the only thing holding state, and
it does not hold it. An instruction that is not written into the code that
asks the question gets re-asked forever.
WHAT WOULD HELP

• Agents should not be able to alter a measurement pipeline mid-task

without explicit approval; "the documented path doesn't fit" needs to
stop the run, not license an improvised one.

• Verbal instructions need a durable sink — a prompt to encode, or a

standing record the agent must reconcile against, not agent memory.

• Mid-turn user messages should survive reconstruction and compaction

intact; they carry the corrections.

• A way to interrupt a running turn without dropping the plan-mode gate.