Claude Code 2.1.261 Adds /skill-doctor and 128K Output Limit

Claude Code Releases Version 2.1.261 on 4 September 2026

Claude Code version 2.1.261 shipped on 4 September 2026, containing 67 CLI changes: 46 fixes, 9 new features, 7 performance improvements, 3 security changes, and 2 breaking changes. The release builds on the preceding v2.1.259 and v2.1.260 updates, which together introduced a fullscreen diff panel and a security fix for permission rules. The v2.1.261 update extends those changes with improvements targeting output limit handling, plugin management, and session control — all addressing friction points that engineering teams running Claude Code at scale or in CI/CD pipelines encountered in production use.

Claude Code's release cadence shortened across 2026, with multiple minor version updates shipping each week. The development team has prioritised fast iteration on specific usability and security fixes rather than batching changes into larger monthly releases. Version 2.1.261 continues that pattern.

/skill-doctor: Auditing Plugin Overhead

The most practically significant new feature in v2.1.261 is /skill-doctor. Claude Code sessions load skill plugins — files that extend Claude Code's behaviour with task-specific instructions, tools, and context — and as teams add more plugins over time, the combined contextual overhead of those plugins consumes tokens before any user input reaches the model. /skill-doctor addresses this by listing all loaded skills, identifying which ones are unused given the current session's task type, and reporting each plugin's contextual cost in tokens. Teams can then prune plugins that are not contributing to their current work, reducing the base token cost of every session.

For organisations running Claude Code with a large set of enterprise plugins or shared skill libraries across multiple projects, the ability to audit plugin load at runtime — and understand exactly which plugins are consuming context budget without being invoked — is the kind of operational visibility that makes a meaningful difference at the scale of hundreds of agent sessions per day. Before /skill-doctor, teams had no straightforward method to inspect the cumulative contextual footprint of their loaded plugin set from within a session.

128K Inline Output Limit

The second headline change in v2.1.261 is the addition of two settings — bashOutputMaxChars and taskOutputMaxChars — that raise the inline output limit for terminal commands and background tasks to 128,000 characters. Previously, large command outputs were saved to a file rather than returned inline to the model, which meant Claude lost direct access to the output content in its active context window. For test suites, build pipelines, and data processing scripts that produce hundreds of lines of output, this meant the model needed to read the file as a separate tool call to access the content.

The 128K limit means most realistic command outputs from test runners, compilers, and linters will be returned inline. The model can reason directly over large stack traces, test failure logs, and build output in the same context turn, reducing the number of tool calls per debugging cycle. For iterative workflows where an agent writes code, runs tests, reads failure output, and fixes the code, keeping the test output inline shortens the round-trip time and reduces the risk of context fragmentation across multiple tool calls.

The Diff Panel from v2.1.259 and v2.1.260

Version 2.1.259 introduced a fullscreen diff panel accessible via the /diff command. The panel opens beside the active conversation and shows uncommitted changes as Claude edits files, giving developers a real-time view of what the model has changed without switching to a separate terminal or git interface. Version 2.1.260 added a security fix for permission rules and extended /reload-plugins to headless and desktop Claude Code sessions, making it available in the same command lists regardless of the execution environment.

SDK and Cloud Session Interrupt Handling

Version 2.1.261 fixes a reliability issue in SDK and cloud sessions: Stop and interrupt signals sent immediately after the first prompt are now honoured. Previously, interrupts sent in the first moments of a session were sometimes ignored, resulting in responses that could not be halted until the first response completed. Teams running Claude Code in SDK orchestration pipelines where task cancellation needs to work reliably across all session phases will see this as a correctness fix. Reliable interrupt handling reduces wasted compute on cancelled tasks and enables tighter control over autonomous agent sessions where early termination is a routine part of the workflow.

Organisation Policy Status Line

The release also adds an "Organisation policy" line to the /status command and to claude doctor's diagnostic output. When an organisation's network policy cannot be loaded — typically because a proxy is blocking the required endpoint — Claude Code now reports the specific reason rather than failing silently. For development teams operating Claude Code behind corporate proxies or in restricted network environments, this simplifies diagnosing access failures by identifying the exact configuration point that is failing.

What This Means for Development Teams in India

The v2.1.261 release addresses three friction points directly relevant for Indian engineering teams deploying Claude Code in production: plugin overhead management for large shared skill libraries, context loss on large command outputs, and silent failure modes in corporate network environments where proxy restrictions are common.

The 128K output limit improvement is particularly valuable for teams running Claude Code against large monorepos or test suites that produce verbose output. Indian product companies managing multiple services with substantial test coverage regularly encounter test failure logs that exceed previous inline limits, forcing the model to read output indirectly. The 128K limit resolves this for most practical test and build scenarios.

The /skill-doctor feature benefits any team that has built a substantial shared plugin library over time and is operating Claude Code at scale. Pruning unused plugins reduces per-session token costs, which matters for teams running hundreds of automated coding sessions per day. For organisations where Claude Code is part of a CI pipeline rather than an interactive tool, the token cost of loaded plugins is a fixed overhead on every run — /skill-doctor makes that overhead visible and actionable.

The Bottom Line

Claude Code shipped version 2.1.261 on 4 September 2026 with 67 CLI changes across 46 fixes, 9 new features, 7 improvements, 3 security changes, and 2 breaking changes. /skill-doctor lists unused loaded skills and their contextual token cost to enable plugin overhead reduction in sessions with large libraries. The bashOutputMaxChars and taskOutputMaxChars settings raise the inline output limit to 128,000 characters, keeping large command and test outputs in context rather than saving them to file. SDK and cloud sessions now honour interrupts sent immediately after the first prompt. The /status command and claude doctor now surface the reason when an organisation's network policy cannot be loaded. Preceding versions 2.1.259 and 2.1.260, also shipped on 4 September 2026, added the /diff fullscreen diff panel and a permission-rules security fix.