2026-08-27 – Claude Code v2.1.247: SendFeedback, Cost Optimisation, and a 1M Compact Window

🧭 Claude Code v2.1.247: SendFeedback Tool, Admin API Skill Expansion, and a Wide Bug-Fix Sprint

Claude Code v2.1.247 arrived on August 27 with three meaningful additions: a tool that lets Claude draft session feedback reports, a major expansion of the Claude API skill to cover Admin API operations, and a batch of bug fixes that address long-standing friction points in keyboard handling, sandbox cleanup, and background session reliability.

SendFeedback: structured session health reports

Claude can now invoke a built-in SendFeedback tool to draft a feedback report when something goes wrong — repeated tool failures, self-detected reasoning errors, or user-flagged issues. Drafted reports are surfaced under /feedback, where you can review the content before sending. The feature is controlled by the feedbackDrafts setting (default: on).

Practical use

After a session that produced unexpected results, run /feedback to see if Claude drafted an issue report. These reports include session context that makes it much easier to pinpoint whether the problem was a tool-call failure, a model reasoning gap, or an environment issue. You can edit the report before sending — or simply use it as a debugging artefact yourself without submitting.

Claude API skill now covers Admin API

The built-in Claude API skill has been extended to include the full Admin API surface, adding coverage for:

Organisation members and invites
Workspaces management
API keys management
Rate limit reports
Workload identity federation
CMEK (Customer-Managed Encryption Keys)

This means Claude can now assist with enterprise account management tasks in the same session as API development work, without switching to a separate tool or documentation context.

Customisable spinner tips

Organisations can now inject their own tips into the Claude Code loading spinner via the tipsFile setting, with control over priority, cooldownSessions, and id fields. A new built-in tip points users toward auto mode when a Bash permission prompt appears.

Notable bug fixes

Bash sandbox and dotfiles — ~/.claude/settings.json managed as a symlink by nix/home-manager or GNU stow was being deleted on sandbox cleanup. Fixed.
/compact system prompt — "Summarize from here" and /compact were using the default system prompt instead of the conversation's own system prompt, causing compacted summaries to miss project-specific context. Fixed.
Cyrillic keyboard layouts — Ctrl keyboard shortcuts were not firing in kitty-protocol terminals when a non-Latin layout (e.g. Russian) was active. Fixed.
/terminal-setup and Zed — the command was overwriting the entire keymap.json for Zed users rather than merging its keybindings. Fixed.
Sub-agent model 404 errors — when a sub-agent's assigned model returned a 404, the sub-agent crashed on the first call. It now falls back through the session's model chain and reports the error clearly (type, status, request ID, model name).
Background sessions stuck on "opening…" — sessions that lost their terminal host were stuck displaying "opening…" indefinitely. They now fail within seconds with a reason.
/install-github-app over SSH — fixed URL handling and copy shortcuts for SSH-cloned repos.
History/model navigation with fast arrow + Enter — rapid keystrokes were acting on the wrong row in history search, /config, /mcp, /skills, /model, and background task lists. Fixed.

⭐⭐⭐ github.com/anthropics/claude-code

🧭 /claude-api cost-optimize: Profile Your API Spend — and Sonnet 5 Now Runs to 1M Tokens Before Compaction

Two features in v2.1.247 directly affect session economics: a new CLI command that systematically audits your Anthropic API costs, and a context-management change that lets Sonnet 5 sessions run significantly longer before auto-compaction kicks in.

/claude-api cost-optimize

Running /claude-api cost-optimize inside a Claude Code session profiles the project's existing API spend and walks you through five cost levers, measuring the impact of each:

Prompt caching — are you re-sending the same system prompt on every call? Cache-able prefixes reduce cost on repeat calls by 80–90%.
Token hygiene — are responses longer than needed? Reducing unnecessary verbosity is often the highest-return lever for high-volume applications.
Batch processing — workloads that don't need real-time responses can use the Message Batches API at 50% off standard pricing.
Effort levels — extended thinking can be scoped to the reasoning budget actually required; over-allocating effort tokens is a common source of runaway spend.
Model choice — the tool surfaces the cost delta of switching from Sonnet 5 to Haiku 4.5 for tasks that don't require frontier reasoning, using your actual prompts as the benchmark.

When to run this

Run /claude-api cost-optimize any time your project's /cost total surprises you, before you scale a workflow from prototype to production, or after adding a new tool that makes heavy API calls. The command reads your session's actual call history rather than using hypothetical estimates, so the recommendations are grounded in real usage patterns.

Sonnet 5 auto-compact window: 934K → 967K → 1M

v2.1.247 changes Sonnet 5's default auto-compact threshold to the full 1M-token context window, up from the previous default of approximately 934K tokens. Sessions now compact at ~967K tokens in practice (a small safety margin below the hard limit), meaning long agentic runs — deep code reviews, multi-file refactors, extended research loops — can proceed for roughly 7% longer before the context is summarised and truncated.

What this means for multi-step agents

If you run agents with many tool calls (web searches, file reads, shell commands), each turn's output accumulates in context. The expanded window gives agents more working memory before compaction discards earlier reasoning. However, compaction still happens — it is not disabled. If you need to guarantee that specific earlier outputs (a plan, a spec, a parsed dataset) survive across many turns, write them to a file rather than relying on the context window to retain them.

Cross-session peer messages: collapsed by default

Cross-session @mention messages now appear as a one-line preview (Message from @sender: <first line>) rather than the full message body. Press Ctrl+O to expand the full content. This keeps the terminal tidy when running multiple coordinated agents that send status updates to each other — a pattern that became much more common after the August 21 self-hosted runner launch.

⭐⭐⭐ github.com/anthropics/claude-code