News

Releases & Models

Claude Code 2.1.283 strengthens governance, reliability, and project instructions

CONFIRMED: Claude Code 2.1.283 expands control over models, telemetry, plugins, and gateways, fixes SDK, MCP, fallback, and cloud-session failures, and preserves useful 2.1.277 history.

Claude Code 2.1.277 amplia instruções de projeto e endurece agentes

CONFIRMED. Anthropic released Claude Code 2.1.283 on September 25. The update shows the current phase of coding agents: the challenge is no longer only making the model answer, but controlling, observing, and recovering what happens next. It continues the 2.1.277 history, which brought AGENTS.md to projects without CLAUDE.md, proxy gateway controls, and fixes for headless sessions, plugins, resumes, and agent workflows.

Control, telemetry, and reliability

availableModelsMatch set to exact permits only explicitly listed versions, while deniedModels blocks specific versions. The x-claude-code-prompt-id header, with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1, groups requests from one prompt. OTEL_LOG_TOOL_CONTENT=1 records MCP, WebFetch, and WebSearch outputs in OpenTelemetry, requiring a retention policy to prevent sensitive-data exposure.

/doctor prompt-audit finds instructions, skills, agents, and commands written for older models. load_test_mode signs requests without sending them upstream, and the gateway accepts Mantle for Amazon Bedrock.

The SDK fixes lost tool calls, final results, and approvals after restarts. MCP preserves progress in long tasks, cleans up stdio servers, and recovers from transient 404s. Workflows started during fallback return to the configured model. Plugins validate names, missing paths, and files outside the allowed directory. From 2.1.277, invisible-Unicode cleanup, marked subagent output, and replacing TaskOutput with reading the output file using Read remain important.

Impact for SMBs

The gain is less rework and more predictability in remote sessions, MCP automations, and cloud integrations. This is not a capability revolution: it is infrastructure maintenance, precisely why it matters to teams already relying on agents. My reading: update development environments, audit prompts and plugins, and test fallback, MCP, telemetry, and resumption before expanding.

Sources

https://github.com/anthropics/claude-code/releases/tag/v2.1.283
https://code.claude.com/docs/en/changelog
https://github.com/anthropics/claude-code/releases/tag/v2.1.277
https://api.github.com/repos/anthropics/claude-code/releases/tags/v2.1.277
https://registry.npmjs.org/@anthropic-ai%2fclaude-code