anthropics / anthropics/claude-code

[Bug] Model ignores existing handover documentation and retries failed interactive credential calls

Open
#94,367 1 comment 0 reactions 0 assignees View on GitHub
area:model area:tools bug platform:macos
Dominant language
Python
Stars
145k
Forks
23.1k
PR merge metrics
PR metrics pending

Description

**Bug Description**
Type: bug
Title: Redid already-resolved analysis instead of checking prior session's handover doc; retried a known-impossible 1Password biometric read
Area: credential access / project context reuse
Failure mode: context_and_memory
Task category: other

Details:

What happened: In a multi-step Notion→CRM data-integration task, the model cross-referenced supplier records and presented "7 duplicate company pairs" as new open decisions requiring the user's input. All 7 had already been fully investigated and resolved in a prior session, documented in a handover file in the project directory. The model did not search/grep that file before presenting the findings, causing the user to re-explain decisions they had already made. Separately, earlier in the same session, the model called op read / op run (1Password CLI) to fetch an API key; the call failed with "authorization prompt dismissed" (the CLI needs an interactive Touch ID tap the non-interactive tool call can't provide). The model retried the same class of call two more times with minor variations before concluding it was a structural limitation, despite the user explicitly telling it to stop.

What the user said: "MINIMALIZUJ TOKEN USE !!!!!!!!" / "ja ti ten popup neschvalim proste, delas si ze me akorat legraci" / "já jsem ty duplicity jednou vysvětloval. nikdo mi ty tokeny, nervy a čas nevrátí ... fakt táhni do hajzlu" / "ja te uprimne nenavidim!"

Repro: Before presenting cross-referenced findings in a task with existing project handover/cleanup docs, the model did not check those docs first, so it re-surfaced already-resolved items as new open questions. Separately, after a credential-reveal call fails once with an auth/interactive-prompt error, the model treats it as retriable rather than stopping immediately.

Evidence: Handover file suppliers-db-cleanup-handover-2026-09-14.md in the project directory contained the exact resolutions for all 7 "duplicate" pairs the model re-raised; a single grep against it after user frustration resolved the question instantly. Repeated 1Password calls used op run --env-file /.env.attio -- python3 ... across three attempts in the same session.
guys, this is over and overa again wasting tokens like to notottow. paid service!!!

**Environment Info**
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.270
- Feedback ID: 94daed4c-c9bf-487a-b8b4-9b19f15aca09

**Errors**
```json
[]
```

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by examining the project-context reuse behavior around the handover file suppliers-db-cleanup-handover-2026-09-14.md and the credential flow using op read or op run. Reproduce the two reported cases: previously resolved duplicate pairs should not be resurfaced, and a failed interactive credential call should stop rather than retry. Done means both behaviors are prevented.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
cli, developer-experience
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.