Patch release with tighter agent guidance (brevity, safer sharing, correction memory) plus internal knobs for effort, event polling, plugins, and subagents.
Changelog
[changed] Agent guidance pushes harder toward substance over narration: less restating of the request or step-by-step play-by-play, more outcomes, decisions, and anything the user must act on — including when the user has chosen brevity.
[changed] Stronger default against unsolicited external posts: routine messages to chat platforms or tickets only when the user directed them, and secrets only with explicit authorization for both the secret and the destination.
[changed] Correction handling is stricter about durability — when a user correction or preference lands, memory writes are expected in the same turn that engages the fix, not deferred until later.
[changed] Subagent instructions clarify background-by-default behavior: keep work in the background unless the next action truly depends on the result; never invent or predict a pending agent’s outcome before the real notification arrives.
[changed] Long-session context guidance spells out that large conversations may be summarized mid-work and continued from the summary plus remaining context, so the agent should not wrap up early solely because the thread is long.
[changed] Focus mode wording is clearer that the user only sees the final text of each response (not tool calls, results, or mid-turn chatter), so everything they need belongs in that final message.
[changed] SDK compatibility matrix no longer lists wrapper 0.3.198 among tested versions (still covers recent 0.3.199+ builds).
Models & prompts
Built-in system guidance adds an explicit “cut narration, keep substance” bar and a short autonomous style cue (minimize interruptions; prefer action over planning when that mode applies).
Safety copy around external sharing and secrets is more direct, including PR follow-ups that only offer Slack posting when the user’s project docs call for it and a Slack send tool is actually available.
Classifier / permission-denial and plan-mode prompts keep the “err on the side of blocking” and no-edits-until-approved posture, with small wording cleanups rather than a new permission model.
Agent and coordinator prompts still emphasize finishing the task fully, reporting concisely for the caller, and treating repeated user reaffirmation after a concern as a decision to proceed (refusals reserved for genuinely harmful or prohibited requests).
Under the hood
New environment toggles for upcoming or experimental paths, including effort level (CLAUDE_CODE_EFFORT_LEVEL), subagent model override (CLAUDE_CODE_SUBAGENT_MODEL), event polling (CLAUDE_CODE_POLL_EVENTS), gateway model discovery, SDK file checkpointing, artifact type controls, loop keepalive, scroll speed, tmux truecolor, token-reminder / workflow size budgets, and bash working-directory maintenance.
Plugin and skill install paths gain explicit sync/download stall and install timeout controls, plus options to skip selected plugin MCP servers during sync.
Additional internal flags cover quiet/startup timing, background-shell pressure reaping, extra metadata, mock 429 behavior, workflow name-only modes, and Foundry base URL configuration — not required for normal use.
Feature-gate churn is small (survey/memory and related experiments); no user-facing model ID swap ships in this build.
Upgrade notes
No action required for typical CLI upgrades.
If you pin an older Claude Code SDK wrapper, note that 0.3.198 is outside the tested set in this release; prefer a current 0.3.199+ wrapper if you hit compatibility issues.
content/
Browse captured content/
Same tree as claude-code/latest/content for this version. Open the full browser for sidebar navigation between docs and other files.
