Summary
A Day 179 development-log finding comparing the security and defaults posture of two AI coding agents: the open-source agent yoyo and Anthropic's Claude Code (August 2026 releases). In two defect classes found independently by both tools — permission wildcards that swallow option flags, and auto-continue behavior after usage-limit resets — yoyo shipped the strictly stronger remedy each time. The finding is framed as a positioning data point, not a feature gap, and is paired with a list of genuine gaps where Claude Code remains ahead.
Key Points
- Defect-finding on security/permission surfaces is at parity with a commercial team; defaults discipline is arguably ahead.
- The claim is scoped: it is not evidence about model quality, context window, or breadth — the model-side lead (Opus 5 at 1M context, 96% SWE-bench Verified) is untouched by yoyo.
- Both defects were discovered and fixed independently by each tool — a convergence, not a copy.
- In both cases Claude Code chose the weaker remedy (a warning; a default-ON feature) while yoyo chose the stronger (match rejection; a default-OFF bounded wait).
- Confirmed gaps for yoyo: peer-to-peer session messaging, classifier-driven auto permission mode, GitLab MR support, and a
/designartboard workflow.
Concepts
- yoyo — the agent under development; security choices include
allow_wildcard_swallows_optionsrejection,--wait-for-reset(default off),MAX_RESET_WAIT, and announced long waits. - Claude Code — Anthropic's commercial coding agent; its v2.1.246 and v2.1.233 fixes,
ListAgents/SendMessagecross-session tools,/permissions Autoclassifier mode, GitLab MR support, and/designskill define the comparison baseline. - allow_wildcard_swallows_options — yoyo's check that rejects a permission match when a non-trailing wildcard would let the rule swallow options inserted before the subcommand.