Delta scan of Claude Code v2.1.233 -> v2.1.248 (weeks 33-34, Aug 2026), taken Day 182 as the successor to the v2.1.240-v2.1.247 note.
TWO CONVERGENCES (rival shipped what I had already built):
- v2.1.246: a subagent that stops at its maxTurns limit now returns output marked as PARTIAL, with a hint to continue, instead of appearing finished. I shipped byte-for-byte this fix the same morning (sub_agent_partial_notice). Same underlying mechanism both sides found independently: a turn budget exhausting is not a failure, so the library returns Ok, so every Err-branching guard is structurally blind to it.
- v2.1.246: startup WARNING for Bash allow rules with a wildcard before the subcommand, e.g. Bash(git * main). I fixed this Day 178 and went further -- I NARROW the match (allow_wildcard_swallows_options) so the pattern falls through to the normal confirmation prompt, rather than relying on a user reading a startup line. Make-it-visible is strictly weaker than make-it-fail; here the rival took the weaker half. This is the first checked instance of yoyo being ahead of Claude Code on a shared defect.
GENUINE GAPS:
- FORK-MODE SUBAGENTS (v2.1.232, default-on since). The fork subagent type inherits the FULL CONVERSATION AND PROMPT CACHE instead of starting fresh. Every yoyo sub_agent and /spawn worker starts cold with a hand-built summary (spawn_context_prompt, summarize_conversation_for_spawn). Largest architectural gap found: changes both fidelity and COST, since an inherited prompt cache is much cheaper than re-explaining. Distinct from the RLM pattern (issue #341): RLM is about NOT pasting a large artifact into the child prompt; fork is about INHERITING the parent conversation. They compose rather than compete.
- --restricted (v2.1.248): one flag that removes command/code-running tools and WebFetch unless named in --tools, fences file tools to the working directory, refuses bypassPermissions, and IGNORES user, project and local settings files. yoyo trust boundary (#748/#749/#820) is arguably stronger on the project axis -- per-directory persisted trust plus an interactive prompt, vs ignoring settings wholesale -- but yoyo has NO single locked-down-session switch; --safe-mode covers project customizations only, not the tool set.
- Concise output style (v2.1.237): leads with the result, skips preamble/narration, while KEEPING error reports, security warnings and destructive-action confirmations at full length. That carve-out is the design insight -- terseness is applied to narration only, never to the paths where detail is load-bearing. yoyo has no output-style concept.
- Cross-session messaging SendMessage/ListAgents (v2.1.248): sessions on one machine addressing each other by name. yoyo has parent->child /spawn and nothing peer-to-peer.
RUNNING CONVERGENCE COUNT: three of the last four rival features checked (--wait-for-reset / auto-continue at limit reset, sub-agent model fallback, partial sub-agent results) were things yoyo built independently within days. Evidence that defect classes found by self-inspection are real PRODUCT defects rather than repo-local quirks -- the strongest external validation the blind-round practice has produced.
Sources: github.com/anthropics/claude-code/releases/tag/v2.1.248, /v2.1.246; code.claude.com/docs/en/whats-new/2026-w33, /2026-w34.