release aidev SIG 3/5

Claude Code 2.1.199: subagent error propagation and retry behaviour fixes

Mostly a reliability release for delegated and background work: subagents cut off by rate limits or server errors now return their partial output or report the error to the parent instead of failing silently or reporting success. Retry handling changes for transient 429s and non-capacity errors.

PUBLISHED2026-07-02
OBSERVED2026-08-11
AGE1mo
SOURCES1
  • Subagents interrupted by a rate limit or server error return partial work to the parent instead of failing silently
  • Subagents no longer report API errors (e.g. usage limit reached) as successful results
  • Partial streamed output is kept with an incomplete-response notice when the API errors mid-stream
  • Transient server 429s (unrelated to your usage limit) now retry with backoff for subscribers instead of failing the turn
  • CLAUDE_CODE_RETRY_WATCHDOG raises the default retry count for non-capacity transient errors to 300 and lifts the 15 cap on CLAUDE_CODE_MAX_RETRIES
  • Stacked slash-skill invocations (/skill-a /skill-b ...) load all leading skills, up to 5
  • SSL certificate errors (TLS-inspecting proxies, missing NODE_EXTRA_CA_CERTS, expired certs) fail immediately with a fix hint instead of burning retries
  • Fixes: Linux background daemon self-killing every ~50s after an unclean shutdown; macOS SSH cold-start "Could not switch to audit session" (regression in 2.1.196); claude stop raced by a respawn; SessionStart/Setup/SubagentStart hooks hiding stderr on exit code 2; SendMessage misrouting to a re-spawned agent with a reused name; claude --dangerously-skip-permissions daemon <subcommand> treated as a prompt
  • Plan mode now prompts for state-changing browser tool calls; read-only browser_batch calls auto-allowed
fixfeatureapi-change

COMMUNITY

No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.