Agent ignores explicit user input and unnecessarily parses session transcripts

Description

During a coding session, the agent exhibited erratic behavior when the user provided explicit parameter suffixes in their message. Instead of using the information already present in the conversation context, the agent:

1. Ignored explicit user-provided data: The user clearly specified autoParamSuffix values (_FRAUD_EVAL_EVENT_TOPIC_ARN and _FRAUD_ACTION_MANAGER_QUEUE_URL) in a plan message. The agent failed to apply them correctly, using the pre-existing values in the file instead.

1. Attempted to recover information from JSONL session transcripts: When the user pointed out the missing suffixes, instead of simply re-reading the current conversation context, the agent ran multiple Node.js scripts to parse .jsonl transcript files — searching for the information that was already present in the active conversation.

1. Context was clean: The user had used /clear to reset context, making transcript searches even more pointless since all relevant information was in the current conversation window.

Expected behavior

• The agent should prioritize information explicitly provided in the current conversation context.
• The agent should never attempt to parse session transcript .jsonl files to recover information that exists in the active conversation.
• When a user provides a detailed implementation plan with specific values, those values should be applied exactly as specified.

Environment

• Claude Code CLI
• Model: claude-opus-4-6