Jonathon's AI Wiki

What's New in Claude Code — Week 34 (August 17–18, 2026)

9 min read

Source: raw/anthropic-watch-claude-code-tag-v2-1-234.md (released 2026-08-17) and raw/anthropic-watch-claude-code-tag-v2-1-235.md (released 2026-08-18), both captured from the anthropics/claude-code releases feed on 2026-08-19.

Two releases so far this week. v2.1.234 is one of the largest single releases the wiki has recorded — roughly sixty entries — and its centre of gravity is unattended operation: sessions that continue themselves when a usage limit resets, /goal that checks in on stalled background work instead of waiting forever, and permission dialogs that can be opened and changed mid-turn. v2.1.235 is a smaller consolidation release. Continues Week 33.

Key Takeaways

  • Claude Code now continues your session automatically when a claude.ai usage limit resets (v2.1.234). Off-switch lives in /config → “Continue automatically at usage limit.” This is the single most consequential entry of the week for anyone running long unattended work — the limit stops being a hard stop and becomes a pause. It also lands in the middle of an active community argument about limits (see Kahn v. Anthropic).
  • The claude-api skill’s context cost dropped from ~200k+ tokens to ~25k by loading its reference docs on demand (v2.1.234). An 8× reduction on a single built-in skill, achieved purely by deferring reference material — the largest concrete progressive-disclosure number the wiki has on file, and a direct measurement of the principle argued in New Rules of Context Engineering.
  • Claude is now instructed to use your account email only to identify you, and not to send it to unrelated services unless you ask (v2.1.234). A first-party data-handling constraint written into the model’s instructions rather than enforced by a tool boundary.
  • /goal stops hanging on dead or stalled turns (v2.1.234). It clears itself with a notice when a turn dies on an unrecoverable error (revoked auth, exhausted credit balance, context overflow) instead of staying armed; and when background tasks keep a goal waiting 30+ minutes, Claude checks in on them rather than waiting indefinitely. Opt out with CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0.
  • Permission and configuration dialogs now open mid-turn. /permissions can be opened while Claude is working and rule changes apply to the rest of the current turn; /add-dir likewise; and /add-dir, /autocompact, /theme, /help, /config and /advisor all open mid-turn in the fullscreen TUI (v2.1.234). Previously a permission mistake meant interrupting the run.
  • Two permission-integrity fixes that were security-relevant, not cosmetic. Credential masking on relayed permission previews could previously hide commands, paths, or destinations from the approver — that is, the human was asked to approve something they could not fully see. Now oversized private-key blocks redact under full-strength redaction instead (v2.1.234). And Shift+Tab inside the permission prompt’s comment field approved the edit and granted session-wide edit permission rather than closing the field (v2.1.235).
  • The NT-namespace path hardening from W33 was extended to the remaining pre-approval file accesses (v2.1.234): remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads now all reject Windows \??\ paths. W33 closed this for UNC path validation; this closes the rest of the NTLM credential-leak vector.
  • Remote Control got materially more honest about state. Signing this computer into a different claude.ai account or organization now stops the running session within seconds with a reason, instead of surfacing a misleading HTTP 404 hours later. Sessions started from Desktop or VS Code keep phones and claude.ai/code updated on permission mode and model; effort picks made on a phone now apply to terminal- and Desktop-hosted sessions (v2.1.234).
  • SendMessage and ListAgents now say when your account’s session list was too long to check completely, instead of treating unseen sessions as absent (v2.1.234). A silent-wrong-answer failure mode turned into a visible one. SendMessage also refuses over-large messages up front rather than dropping them silently (v2.1.235).
  • The “Default teammate model” setting is gone from /config (v2.1.234). Agent-team teammates now use the leader’s model unless the spawn names one. If you relied on that setting to route teammates to a cheaper model, that routing must now be explicit at spawn time — see Agent Teams.
  • Todo/task list and GitLab work continued. A GitLab merge-request badge now appears in the footer and statusline (MR !N with draft/pending/green states) for repos with a GitLab remote and an authenticated glab CLI (v2.1.234), extending the GitLab push documented in W33.
  • Optional spellcheck in the prompt input (v2.1.235), underlining misspelled words as you type using an installed aspell, hunspell, or ispell.

Unattended operation is the theme

Three v2.1.234 entries point the same direction, and they are more interesting together than apart.

The usage limit stops being terminal. “Claude Code now continues your session automatically when a claude.ai usage limit reset” converts the most common hard stop in long-running work into a wait. Combined with /goal’s new 30-minute check-in on stalled background tasks and its self-clearing on unrecoverable errors, the failure modes that used to end an overnight run quietly now either recover or announce themselves.

Mid-turn control closes the other half of it. Being able to open /permissions and have rule changes apply to the rest of the current turn means a run that hits an unexpected permission wall no longer has to be killed and restarted. Same for /add-dir when the agent needs a directory nobody scoped in advance.

The honest caveat: auto-continue at limit reset is on by default, and it changes the cost profile of walking away from a session. If you deliberately used the usage limit as a budget stop, turn it off in /config.

The progressive-disclosure measurement

Reduced the context cost of loading the built-in claude-api skill from ~200k+ tokens to ~25k by loading reference docs on demand.

This deserves separating out because the wiki has argued the principle repeatedly without a first-party number attached to a specific artifact.

  • The reduction is ~8×, on a skill that ships with Claude Code.
  • Nothing was deleted — the reference docs still exist and still load, just on demand.
  • ~200k tokens is a substantial fraction of a context window spent before any work begins. The skill was, in effect, a worked example of the anti-pattern its own guidance warns about.

This is the same move the vault itself made when the schema was split into operations/ files, and the same one Agent Skills describes as the model for skill authorship. It also bears on a community argument that progressive disclosure still spends attention on the catalog — see Open Questions.

Selected fixes worth knowing

  • Auto mode in very long sessions was repeatedly re-checking and denying sandboxed commands’ network access after the conversation had been compacted (v2.1.234). If auto mode seemed to get more restrictive the longer a session ran, this was why.
  • Session-scoped permission answers, including denies, were dropped when answering background subagent tool permission prompts (v2.1.234). A deny you gave could silently fail to persist.
  • MCP diagnostics were printing resolved secrets (v2.1.234). Scope-conflict warnings now show the configured ${VAR} form; connection-failure details show only the server origin.
  • strictKnownMarketplaces allowlists accepted SCP-style git marketplace sources whose host differs from the one git would actually connect to (v2.1.234) — an allowlist that did not constrain what it appeared to constrain.
  • Whole-prompt-cache invalidation when a language server disconnected or reconnected mid-session (v2.1.235). A silent, expensive cache miss.
  • The Agent tool advertised a general-purpose default in sessions where that agent is unavailable (v2.1.235); an omitted subagent_type there now returns a clear error listing the available agents.
  • Repository detection mis-read the host of git remotes with unusual userinfo, producing links and repo-specific behavior for the wrong host (v2.1.234).

Try It

  • Decide whether you want auto-continue at usage limit. /config → “Continue automatically at usage limit.” Leave it on for long unattended runs; turn it off if you were using the limit as a spend brake.
  • If you run agent teams, check your model routing. The “Default teammate model” setting is gone — name the model at spawn time or teammates inherit the leader’s.
  • Re-test any workflow that relied on a session-scoped deny inside a background subagent, since those denies were previously being dropped.
  • If you author skills, read the claude-api reduction as a template: move reference material behind on-demand loading rather than trimming content.
  • On Windows, update. The \??\ path hardening now covers remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads.

Open Questions

  • What is the auto-continue behavior’s interaction with billing? The release note says the session continues when the limit resets; it does not say whether anything is queued, retried, or charged differently in the interim.
  • Does the claude-api skill’s ~8× reduction generalize? ~200k → ~25k is reported for one built-in skill with unusually large reference docs. No methodology or per-skill breakdown is published, so this is not a rate to expect elsewhere.
  • Is the “use your account email only to identify you” instruction enforced anywhere below the model? It is written as an instruction to Claude, which is a weaker guarantee than a tool-level or network-level block. The release note does not say.
  • v2.1.230 remains skipped, as noted in W33; v2.1.234 and v2.1.235 are consecutive.
  • Does the progressive-disclosure win answer the “context tax” objection? A community argument (raw/reddit-1vry6gk.md) holds that model attention should not serve as the retrieval index at all at 500+ skills, since the catalog itself grows. Deferring a skill’s reference docs shrinks the body, not the catalog entry — so the objection is narrowed by this change, not resolved by it.

Graph View

Backlinks