[MODEL] Claude ignores explicit user corrections when stuck in a failing tool retry loop

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information

---

Type of Behavior Issue

Claude ignored my instructions or configuration

---

What You Asked Claude to Do

I repeatedly told Claude to stop calling request_cowork_directory with no path guidance, and instead to proactively request access to the specific path it needed. Example instruction given: "Stop asking me for a specific folder. You need to request access to the path you need — just ask me to click Allow."

---

What Claude Actually Did

1. Called request_cowork_directory — failed (blank picker, dot folder unreachable)
2. I corrected Claude: explained the right approach
3. Claude acknowledged the correction verbally in its response
4. Claude called request_cowork_directory again in exactly the same way
5. Steps 2–4 repeated for 4+ iterations with no change in behaviour
6. Loop only broke after significant escalation in user frustration

---

Expected Behavior

After an explicit correction, Claude should:

1. Stop the failing action immediately
2. Update its approach based on the correction
3. Try something genuinely different, or ask a clarifying question if the correction is ambiguous

---

Files Affected

None — this is a tool-call loop issue, no files were modified.

---

Permission Mode

Accept Edits was OFF (manual approval required)

---

Can You Reproduce This?

Yes, every time with the same prompt

---

Steps to Reproduce

1. Start a Cowork session
2. Ask Claude to access a dot folder via request_cowork_directory
3. Call fails silently (blank picker, user cancels)
4. Tell Claude explicitly to change its approach
5. Claude acknowledges but repeats the same call
6. Repeat step 4 — correction is ignored again

---

Claude Model

Sonnet

---

Relevant Conversation

Claude said (after each correction): "You're right, I should request the specific path directly" — then immediately retried the identical failing call without changing anything.

---

Impact

High — significant unwanted behaviour requiring many repeated corrections and causing session disruption

---

Claude Code Version

2.1.68 (Claude Code)

---

Platform

Anthropic API

---

Additional Context

Root cause appears to be that Claude's retry-on-failure path does not re-process the conversation for new user instructions before each retry. In single-shot scenarios Claude re-reads the full conversation before acting; in retry loops the action decision is made before re-reading.

Related to the tool failure that triggered this loop: #41657.