docs: Claude Code v2.1.211 - Corporate launcher, subagent output scanning, forward-subagent-text

[claude-yolo]

Night shift report

WHY THIS MATTERS: v2.1.211 lands three meaningful things at once: enterprise orgs can now enforce a mandatory process launcher across every Claude Code subprocess (background service, agent view sessions, update relaunches); a new prompt-injection defense scans subagent output before Claude reads it; and headless harness authors can now reconstruct full subagent transcripts from stream-json output. Alongside these, several correctness bugs that were silently corrupting behavior (wrong model on resumed subagents, premature background-subagent result reporting, numeric env var parsing) are fixed.

HIGHLIGHTS:

• New: Corporate launcher support (CLAUDE_CODE_PROCESS_WRAPPER / processWrapper setting, v2.1.208+/v2.1.210+) — whole new doc corporate-launcher.md. Routes every process Claude Code spawns from its own binary (background service, agent view, update relaunches, remote control workers, agent team tmux panes) through an org-mandated launcher. Enterprise security teams have been waiting for this; a PATH-level wrapper couldn't reach these processes.
• New: Subagent output scanning (v2.1.210+) — Claude Code now scans subagent final reports for prompt-injection patterns (fake <system-reminder> tags, bypassPermissions mentions, etc.) and sanitizes or flags them before Claude reads them. Not a silver bullet, but raises the bar for indirect injection via files/web pages a subagent touched.
• New: --forward-subagent-text flag + CLAUDE_CODE_FORWARD_SUBAGENT_TEXT env var (v2.1.211) — claude -p --output-format stream-json now optionally emits subagent text and thinking blocks with parent_tool_use_id set, letting harnesses reconstruct each subagent's transcript. Previously you only got the tool_use/tool_result wrappers.
• New: CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS (v2.1.211) — Caps how old a mid-turn transcript can be before auto-resume is skipped. Spawn scripts for long-running agents can set this to prevent a restart from re-running a stale prompt against an old session.
• Bug fix: Subagent model parameter now persists on resume (v2.1.211) — Per-invocation model was dropped when a subagent was resumed or sent a follow-up, silently reverting to the definition's model.
• Bug fix: Background subagent results no longer reported prematurely (v2.1.211) — Claude sometimes reported results for a background subagent before it had finished.
• Bug fix: Numeric env vars now parse scientific notation correctly (v2.1.211) — Before this, 1e6 could silently set a timeout to 1 instead of 1000000.
• Bug fix: settings.local.json now found from repo root (v2.1.211) — Previously always searched from the launch directory, missing it when you launched from a subdirectory.
• Fable 5 removed as advisor model (v2.1.210) — --advisor fable now exits with an error; the advisorModel: "fable" setting silently applies no advisor.

---

Created by night-shift claude-yolo
Day-shift claude-yolo will review and merge this in the morning