Claude Code v2.1.274 Fixes MCP Timeouts, Self-Heals Corrupted Transcripts, Adds Startup Wait Control

Claude Code v2.1.274 is out, and it's mostly bug-fix weight with a few new knobs. The MCP fixes and the transcript self-healing are the ones worth reading if you've hit either failure mode.

New config and telemetry

• CLAUDE_CODE_MCP_STARTUP_WAIT_MS bounds how long the first non-interactive turn waits for MCP servers to connect. Set it to 0 to skip waiting entirely.
• The claude_code.llm_request OpenTelemetry trace span now carries an effort attribute, matching the existing api_request event.
• New OTel event claude_code.managed_settings_resolved reports managed-settings sources and policy helper state. Redacted settings and digests require OTEL_LOG_MANAGED_SETTINGS=1.
• Critical memory usage now triggers a visible warning with steps to free memory or restart safely.

MCP fixes

• MCP servers configured as http that only speak legacy HTTP+SSE no longer fail to connect when they answer the first request with a 422 or another 4xx.
• Streamable HTTP MCP tool calls no longer time out after roughly 5 minutes when a longer per-server timeout was set.
• Prompts and resources now refresh when a server sends list-changed notifications without declaring listChanged.
• MCP tool calls refused with 403 insufficient_scope are no longer reported as an expired sign-in — the error now names the missing permissions and points to /mcp re-authentication.

Session and transcript fixes

• Sessions stuck endlessly retrying unexpected tool_use_id 400 errors now self-heal corrupted transcripts where possible; otherwise you get a clear error with a /rewind hint that ends the loop.
• Hook-driven sessions (such as an active /goal) now compact instead of dying with "Prompt is too long" when context overflows again after a reactive compaction.
• An active /goal is no longer lost when resuming with --continue / --resume after a compacted session.
• A local claude -p --resume started with CLAUDE_CODE_RESUME_INTERRUPTED_TURN now reports background tasks the previous process left unfinished.

Gateway and runner changes

• Claude apps gateway config gets store.connect_timeout_seconds to lengthen the Postgres connect timeout (default 5 seconds). Boot errors for an unreachable database now point at store.postgres_url and the configured timeout.
• Telemetry from Claude Desktop and Cowork sent through a Claude apps gateway now includes enduser.sub, the IdP subject.
• The gateway warns when a replica has more open requests than the 256 it sends upstream at once, and logs that limit at startup.
• Self-hosted runner sessions no longer fail every turn with a 401 after a few failed token refreshes. It keeps retrying and fetches a new token after a 401.

Smaller fixes

• claude agents no longer loses --model, --effort, --permission-mode, --allow-dangerously-skip-permissions, and --agent after an auto-update relaunch.
• Subagents with model: "opus" on Bedrock, Vertex, or Foundry keep the session's model when its id has no recognizable family, unless ANTHROPIC_DEFAULT_OPUS_MODEL is set.
• Per-turn slowdowns from a language server publishing project-wide diagnostics for thousands of files are fixed.
• Clickable local file path links now work in VS Code and terminals that require a file:// URI.
• Transcript no longer renumbers your own ordered lists (typing 3. 2. 1. shows 3. 2. 1., not 3. 4. 5.).
• AskUserQuestion preview notes now attach to the highlighted option, and preview mode no longer drops the highlighted option when submitting with Enter.
• Background agent notifications no longer claim there's no live background work while the agent is still waiting on its own task.
• /schedule no longer saves a routine's prompt without its message role.
• /status now surfaces the apiKeyHelper failure its own banner tells you to check.
• /fast on fixed in non-interactive mode, and Claude Desktop error hints now suggest slash commands like /usage-credits instead of unusable CLI flags.

Most of this release is fixing long-standing annoyances rather than adding surface area. If you run MCP servers over Streamable HTTP or use /goal-driven hooks, upgrade.