[SergioTCG]

Problem

Older releases recorded the Claude status line's display name as the model, so stored launch flags held values like --model Opus 4.8 or --model Opus 5 (1M context). Launch flags are stored as one space-joined string. A later model change consumed only --model Opus and left the remaining words (4.8, 5 (1M context)) behind.

resumeArgsFor() passed those words through, so every resume started Claude with a positional argument. Claude treats that as a prompt and immediately runs a turn on "4.8" or "5" that nobody sent. Each new start reports its argv back through CCS_FLAGS, so the leftover words also stayed in the stored flags.

Current code already records a validated model ID (claudeModelForResume), so no new records get these words. This change cleans up the ones that already exist.

Fix

displayFlagsFor(), which feeds both resume and /sab-flags display, now drops a bare Claude word in two cases:

• it follows an on/off flag (--dangerously-skip-permissions, --chrome) or another dropped word;
• it follows a --model/--effort that already has its value.

Bare words after an option SAB does not know are kept, so operator CCS_NEW_FLAGS / CCS_RESUME_FLAGS such as --permission-mode acceptEdits or --add-dir /a /b still resume unchanged. The next start reports clean flags, so stored records heal themselves without a state migration. Codex and Pi parsing is unchanged.

Verification

• New provider test covering the two real shapes, the original unsplit --model Opus 4.8, an inline --option=value followed by a leftover word, and preserved values of unknown options. It fails without the fix.
• npm test: 594/594 pass.
• Dry run against a real install's state (70 sessions): all Codex and Pi resume args are unchanged. Affected Claude sessions lose only the leftover words, for example --dangerously-skip-permissions 4.8 --model claude-opus-5-5[1m] … becomes --dangerously-skip-permissions --model claude-opus-5-5[1m] ….

🤖 Generated with Claude Code