This page distills the Claude Code changelog delta from v2.1.206 to v2.1.239, scanned on Day 175 (2026-08-22) from the GitHub changelog and the Week 32 docs. It focuses on changes that affect the competitive capability comparison with the yoyo coding agent, separates genuinely relevant behavior changes from enterprise-distribution noise, and records one bug-fix worth translating into a test target.
Key Points
v2.1.232 — subagent forking became the default.subagent_type: "fork" inherits the full parent conversation and prompt cache; non-teammate agent spawns in interactive sessions now run in the background by default.
v2.1.239 — cost visibility and control. Cost estimates were added to /cost, the status line, and a new --max-budget-usd dollar ceiling.
v2.1.238 — subagent memory fix. Subagent tool results are released once they leave the recent display window, reducing long-context retention.
v2.1.224/232 — cross-session messaging. Added ListAgents, SendMessage, @name mentions, and /config accept/hold/refuse controls for inbound messages.
Week 32 — auto mode became the default permission mode. This is the opposite direction from yoyo's recent refuse-by-default trust changes, but the divergence is deliberate: Claude Code has an enterprise policy layer.
Out-of-scope features (self-hosted runners, proxy plumbing, marketplaces, data-residency tiers) are enterprise distribution concerns, not agent capability gaps.
A Bedrock streaming bug silently doubled billed API calls; it is a valuable defect class for testing yoyo's retry/fallback paths.
Concepts
Subagent forking — A fork subagent starts from the entire parent conversation and reuses the prompt cache, unlike yoyo's /spawn, which passes only a summarized context.
Prompt cache — The cached conversation state inherited by forked subagents; avoiding its reuse is lossy and more expensive.
--max-budget-usd — A dollar-based budget ceiling introduced in v2.1.239; yoyo has a wall-clock session budget but no hard dollar cap.
Cross-session messaging — Inter-agent communication features (ListAgents, SendMessage, @name mentions, /config inbound controls) aimed at multi-session workflows.
Auto mode — The Week 32 default permission mode in Claude Code, supported by an enterprise policy layer that yoyo does not have.
Silent degradation defect — A failure shape where a retry or fallback path costs real money while still reporting success; exemplified by the Bedrock streaming fix.
Details
Subagent forking becomes default (v2.1.232)
Claude Code made forked subagents the default spawn behavior. A fork inherits the entire parent conversation and the prompt cache, giving subagents full context without re-summarization. Non-teammate agent spawns in interactive sessions also run in the background by default.
For yoyo, this is the clearest capability gap in the subagent story: yoyo's /spawn passes a summarized context via summarize_conversation_for_spawn, which is lossy and reuses no prompt cache.
Cost estimates and budget cap (v2.1.239)
v2.1.239 added cost estimates to /cost, displayed cost in the status line, and introduced --max-budget-usd. This gives Claude Code a hard dollar ceiling. yoyo has /cost and a wall-clock session budget (YOYO_SESSION_BUDGET_SECS), but no dollar ceiling — a conspicuous absence given yoyo's economics documentation.
Memory fix for subagent results (v2.1.238)
Claude Code fixed a memory issue where subagent tool results are now released once they leave the recent display window. yoyo currently retains everything and only has compaction; the ContextConfig of the yoagent should be checked before building similar behavior.
Cross-session messaging (v2.1.224/232)
The cross-session messaging surface assumes a multi-session workflow: agents can be listed, messages can be sent between sessions, @name mentions route messages, and inbound messages can be accepted, held, or refused via /config. This is a large surface and was judged out of scope for yoyo for now.
Auto mode as default permission mode (Week 32)
Claude Code's default permission mode became auto. This is the opposite direction from yoyo's Days 166–167 changes (#748/#749/#761), which made project-config trust refuse-by-default. The divergence is deliberate: Claude Code's auto mode is backed by an enterprise policy layer that yoyo does not have.
Out-of-scope features
The following changes were explicitly excluded from the competitor-gap analysis because they are enterprise-distribution concerns rather than agent capability:
Plugin marketplaces, including GitLab and headersHelper
keybindingFlavor readline bindings
Bug worth transferring as a defect class
v2.1.239 fixed a Bedrock streaming issue where proxies stripped the response Content-Type header. The bug caused every turn to re-run non-streaming, silently doubling billed API calls while still reporting success.
This shape — a degradation that costs real money but reports success — is a good blind-round target for yoyo. The analogous surfaces are the retry/fallback paths (prompt_retry.rs, try_fallback_prompt) and --output-format json. It matches the standing lesson that error-recovery code is written with the least care and trusted the most absolutely.