Bash learning: discover repeated shell patterns and support explicit approval before redirecting

[willwashburn]

Problem

Wash has a fixed warn-only list for Bash commands with structured equivalents. Over time, users may repeatedly run project-specific shell commands that produce large output and could benefit from structured handling or stronger nudges. Auto-blocking unknown Bash patterns would be risky, but surfacing candidates is useful.

Goal

Teach the learning layer to discover repeated high-waste Bash patterns and present them for explicit approval.

Scope

• Observe Bash commands that match no existing redirect pattern.
• Redact or normalize sensitive command parts.
• Group commands into pattern candidates.
• Estimate output waste from command result size when available.
• Write candidates to pendingBashPatterns in the profile.
• Add an approval command that promotes selected patterns into active redirect hints.

Approval flow

Suggested commands:

• wash learn bash-candidates
• wash learn approve-pattern <id>
• wash learn reject-pattern <id>

Suggested slash command:

• /wash-approve-patterns

Safety rules

• Never auto-block a discovered pattern.
• Never store full commands that include quoted secrets or obvious token values.
• Prefer normalized patterns such as docker logs <arg> over raw command strings.
• Approval should be reversible.

Acceptance criteria

• Repeated unmatched Bash commands are grouped into candidates.
• Candidates include count, estimated output bytes, and suggested structured replacement when known.
• Approved patterns produce warn-only hints in PreToolUse.
• Tests cover grouping, redaction, approval, rejection, and profile persistence.