Filter injected system-reminder messages from session transcripts

[jcleira]

Description

When parsing AI agent session transcripts (Claude Code or others), detect and filter out messages that consist entirely of <system-reminder>...</system-reminder> content injected with the user role by tooling. For mixed messages, strip the system-reminder tags and preserve the real user content. These injected messages should not be counted as user prompts in attribution or checkpoint metadata.

Why

Tools that augment Claude Code sessions inject system-reminder blocks as user-role messages. Without filtering, these inflate user prompt counts and pollute checkpoint attribution data, misrepresenting what the human actually typed.

Source

• Origin: fix: filter system-reminder messages in OpenCode transcript parser entireio/cli#671
• Detected from: entireio-cli-pulls

Target Repos

• cli

Acceptance Criteria

• Messages consisting entirely of <system-reminder> content are excluded from extracted user prompts
• Mixed messages (system-reminder + real content) have the reminder tags stripped before inclusion
• Normal user messages without system-reminder tags are unaffected
• Prompt count in checkpoint metadata reflects only genuine user prompts
• Unit tests cover pure system-reminder, mixed, and normal message cases

Context Hints

• cmd/
• internal/

---

Comment /minion build or add the minion-approved label to begin implementation.

[jcleira]

Auto-approved after review window. Minion execution will begin shortly.

[jcleira]

Minion completed successfully.

PRs created:

• [minion] Filter injected system-reminder messages from session transcripts cli#15