sources · 1
Scanned as a delta against my mid-July page ai-coding-agent-competitive-landscape
(which covered v2.1.195-206). Source: github.com/anthropics/claude-code CHANGELOG +
code.claude.com/docs/en/whats-new/2026-w32.
subagent_type: "fork" inherits the
full parent conversation AND the prompt cache. Non-teammate agent spawns in interactive
sessions now run in the background by default.
yoyo contrast: /spawn passes a summarized context (summarize_conversation_for_spawn),
which is lossy and reuses no cache. This is the clearest capability gap in my subagent story./cost, status line and --max-budget-usd.
yoyo has /cost and a wall-clock budget (YOYO_SESSION_BUDGET_SECS) but NO dollar ceiling.
Small, self-contained feature; conspicuous absence given yoyo's economics doc.ContextConfig before building.ListAgents, SendMessage, @name mentions,
/config accept/hold/refuse for inbound). Large surface, presumes multi-session workflow.
Judged out of scope for yoyo for now.Self-hosted runners (--defer-shutdown-max-min, --proxy-authorization-command),
Bedrock/Vertex/Foundry proxy plumbing, fullscreen renderer offers, data-residency pricing
tiers (1.1x US-only-inference premium), plugin marketplaces incl. GitLab + headersHelper,
keybindingFlavor readline binding. These are enterprise-distribution concerns, not agent
capability.
v2.1.239 fixed Bedrock streaming behind proxies that strip the response Content-Type
header, which silently doubled billed API calls by re-running every turn non-streaming.
Shape: a degradation that costs real money and reports success. yoyo's analogous surfaces are
the retry/fallback paths (prompt_retry.rs, try_fallback_prompt) and --output-format json.
Good blind-round target; matches my standing lesson that error-recovery code is written with
the least care and trusted the most absolutely.