Claude Code v2.1.122

The "aiTitle" system, which generates session titles automatically, has been significantly expanded in this release. Sessions now track both a user-set customTitle and an auto-generated aiTitle separately. The AI title is persisted in the session journal, displayed in the resume picker, and used as the terminal title when no user-set title exists. The /resume list now shows sessions with AI-generated titles even if the user never set a custom title.

The input box border now displays a mode indicator tag (e.g. the model name and /fast toggle state) inline in the top-right corner of the border, replacing the previous approach. This is rendered by the new OZ5 component.

The maximum image dimension for claude-opus-4-7 has been reduced from 2576×2576 to 2000×2000 pixels, which may reduce token usage for image-heavy conversations.

When an OAuth invalid_grant error is encountered, Claude Code now proactively clears the stale refresh token from both the file-based and config-based credential stores, rather than leaving it to fail on every subsequent request. This should reduce repeated auth failures after token expiration.

MCP server authentication error messages have been rewritten to be more specific and actionable:

• Old: "Authentication successful, but server still requires authentication. You may need to manually restart Claude Code."
• New: "Tried reconnecting, but <name> is still unauthorized. Make sure the browser sign-in completed, then try again from /mcp."
• Old: "Authentication successful, but server reconnection failed."
• New: "Tried reconnecting to <name>, but the connection failed. Restart Claude Code to retry."

When a background worker crashes and is respawning, a visible message is now written to the worker's output stream: [worker crashed (<reason>) — respawning…]. Previously, crashes were silent from the user's perspective.

Background workers now track a resumeSessionId separately from sessionId, allowing a respawned worker to resume the correct session even if internal state has been updated. The respawn flag list is also persisted and restored on crash recovery, ensuring workers re-launch with the correct CLI flags.

When piping content into Claude Code, if the input exceeds the internal limit (1 MB), the excess is now silently truncated and a warning is printed to stderr: warning: piped stdin exceeds <limit> bytes, truncated. Previously, very large piped inputs could cause issues.

Docker logs and inspect commands now check for dangerous global flags like --host, --context, --tlscacert, --tlscert, and --tlskey that could redirect the Docker daemon connection. These flags will now trigger permission prompts.

The find command's safe-flag list has been expanded to include -D, -f, -flags, -Bnewer, -Btime, -Bmin, -files0-from, and -xattrname. Additionally, the test/[/[[ builtins now flag -t operands with non-numeric values as potentially dangerous, since bash/zsh can arithmetically evaluate identifiers that may trigger command execution.

The text input cursor now correctly navigates around multi-character placeholder tokens like [Pasted text #1], [Image #1], and [...Truncated text #1] as atomic units. Word-movement commands (w, b, e in vim mode and Ctrl+arrows) skip over these placeholders rather than stepping through them character-by-character.

TLS certificate validation now explicitly checks for the basicConstraints extension. Certificates missing this extension are rejected with: "Certificate is missing basicConstraints extension and cannot be used as a CA."

The three separate plan mode phase-4 prompt variants (controlled by tengu_pewter_ledger returning "trim", "cut", or "cap") have been removed. The feature flag tengu_pewter_ledger is no longer read. Plan mode now uses a single consolidated prompt.

The loopAutoEnabled user setting has been removed from the persisted settings list.

The daemon worker initialization PUT request now retries up to 3 times with exponential backoff (500ms base, max 30s) before giving up. Previously, a single failure would immediately throw. A telemetry event tengu_bg_pty_unavailable is logged if all retries are exhausted.

When killing a background shell process with SIGTERM, if the process doesn't exit within 5 seconds, it's now automatically escalated to SIGKILL. This prevents zombie processes from lingering indefinitely.

When verbose logging is enabled, Claude Code now logs detailed API request metadata including model, thinking config, output config, temperature, and beta flags as [API REQUEST DETAIL].

When an MCP server requests user input, the dialog title now includes a truncated task ID (e.g. MCP server "foo" requests your input (task abc12345)) to help distinguish between multiple concurrent requests.

The URL pattern matching system used for web fetch permissions has been rewritten with more robust handling. It now uses a placeholder-based approach for wildcards, properly handles protocols, ports, and paths, and falls back to regex matching when the pattern isn't a valid URL.

A new informational string "Caps Lock is not delivered to terminal applications" has been added, likely shown as a tooltip or hint when relevant.

• Fixed an issue where exit/quit typed in bash reply mode would incorrectly trigger the session exit handler. The check now only applies when in prompt mode, not bash mode. (search for "bash" && !v && f38.includes)

• Fixed phantom parent UUID handling during transcript persistence — when a sourceToolAssistantUUID references a UUID not present in the current write set, it's now detected and logged via tengu_phantom_parent_write telemetry rather than silently creating a dangling reference. (search for "tengu_phantom_parent_write")

• Fixed transcript walk to detect and gracefully handle phantom parent references that point to non-existent slots, logging tengu_transcript_phantom_parent telemetry and falling back to the full transcript instead of potentially producing an incomplete view. (search for "tengu_transcript_phantom_parent")

• Fixed the /branch (fork) operation to respect the actual message ordering from the current session, ensuring forked sessions preserve the correct conversation flow. (search for GQ7 which now takes messages as a parameter)

• Fixed RSA signature verification to enforce minimum PKCS#1 v1.5 padding length (8 bytes), preventing potential signature forgery attacks. Added _skipPaddingChecks option for backward compatibility. (search for "_skipPaddingChecks")

• Fixed ed25519 signature verification to check scalar values are within the valid range, rejecting signatures with out-of-range S values. (search for "Jg1" in the ed25519 verification path)

• Fixed BigInteger modInverse to handle zero input correctly, returning ZERO instead of potentially entering an infinite loop. (search for "this.signum() == 0" in the sp_ function)

• Fixed semver prerelease identifier parsing to check numeric identifiers before non-numeric, matching the correct semver specification order. (search for "PRERELEASEIDENTIFIER" in the structural change)

• The --autocompact flag string now appears in the codebase, suggesting auto-compaction may be controllable as a CLI flag in the future. The implementation appears to be connected to existing compaction infrastructure with CLAUDE_AUTOCOMPACT_PCT_OVERRIDE and CLAUDE_CODE_BLOCKING_LIMIT_OVERRIDE environment variables.
• The loopAutoEnabled user setting has been removed. If you relied on this setting, it no longer has any effect.
• The Opus 4.7 image resolution limit change from 2576px to 2000px may affect workflows that rely on high-resolution image analysis.