What happened
Anthropic’s Claude Code changelog lists version 2.1.329 with additions, fixes and behavior changes across the CLI, integrations and Code Mode. One notable addition lets users enable Code Mode while OpenTelemetry events and metrics are enabled. The changelog says each tool call made by the code sends its events, while tracing still keeps Code Mode off. This is a specific distinction: telemetry events and metrics can coexist with the experimental mode, but tracing remains incompatible according to the release note.
Several Code Mode fixes address reliability when scripts interact with tools and long-running sessions. A script’s call to a tool that takes no arguments no longer fails with a missing-parameter error. Scripts can retain access to tools through tools in sessions with more than 4,096 tools, and a script can call an MCP tool’s request to end the turn when Claude is waiting for that script. Code Mode also no longer loses finished edits and the stopped command after Escape, and it avoids retaining every large text file read by code or growing memory each time a background subagent wakes.
The release adds limits and clarifies hook behavior. Code Mode refuses a script’s read of a line range over 256 MiB, which the changelog says could use gigabytes of memory, and fixes timeouts on reads of very long lines. A runaway loop in Code Mode is now stopped at a limit. Separately, a PreToolUse hook that errors no longer refuses a script’s tool call by default, matching behavior outside Code Mode. Anthropic says administrators who need a hook failure to block an action should set "onFailure": "block" on a command, HTTP or MCP tool hook.
Other changes address waiting, permissions and monitoring. Code Mode scripts no longer stop simply because a permission prompt, plan approval or question remains open for ten minutes in the cases described by the changelog. The release also improves how quickly the terminal recovers after a script makes thousands of tool calls, and changes subagent guidance so Claude receives the same waiting guidance inside Code Mode as outside it. The entry is one part of a larger release: 2.1.329 also includes changes to tasks, MCP diagnostics, marketplace status, code review and session behavior.
What the archive adds
New Runtime’s earlier coverage includes Claude Code 2.1.310’s runtime and cloud-session changes (2026-10-24) and version 2.1.319’s auto-mode and MCP task controls (2026-11-08). This is a new release event in an ongoing Claude Code thread, with a new operational fact: Code Mode can emit OpenTelemetry events and metrics for its tool calls.
Trend signal
This supports harness architecture outlives model choice: the release is about the execution layer around model output—tool calls, hooks, memory and observability—rather than a new model. The signal is modest: changelog improvements show Anthropic refining the harness, but do not by themselves establish how teams use these controls in production.
Why now
The changelog’s 2.1.329 entry couples a new telemetry path for Code Mode tool calls with a set of fixes to script limits, memory use and MCP interactions. Together, these changes address practical concerns that arise when a tool-calling mode is used in longer or more complex sessions. Anthropic does not frame the release as a general-availability change; Code Mode remains described as experimental.
Why it matters
A scripted mode can reduce the friction of coordinating many tool calls, but teams need to see what those scripts do and keep failures bounded. OpenTelemetry events for Code Mode tool calls give observability systems a route to capture activity that might otherwise be harder to inspect alongside ordinary session events. The caveat around tracing matters: teams should not assume all telemetry features are interchangeable or that tracing works with the mode.
The hook change is equally consequential for teams setting policy. A hook that errors is no longer an implicit denial for a Code Mode script’s tool call; a team that depends on fail-closed behavior must configure onFailure: block for applicable command, HTTP or MCP hooks. The memory and loop limits address failure modes that could make experimentation costly on a developer’s machine. These are useful operational controls, not evidence that the experimental mode is ready for every production workflow.
Limits and uncertainty
The changelog describes implementation changes, not measured adoption, reliability rates or performance comparisons. It does not establish how much additional visibility OpenTelemetry provides in a particular deployment, or whether all relevant tool-call events are captured by a team’s exporter and dashboards. Code Mode remains experimental, and the listed fixes do not quantify the frequency or severity of the problems they address. Teams should validate their own telemetry, hook configuration and resource limits before relying on this behavior in consequential workflows.
Skill to build now
If your team is evaluating Code Mode, test hook failures deliberately: configure "onFailure": "block" wherever an error must prevent a command, HTTP or MCP action, then confirm that a failed hook blocks as intended. In the same test session, compare your OpenTelemetry events and metrics with a known sequence of scripted tool calls. Keep tracing out of that test, since Anthropic says tracing still disables Code Mode.
Reader takeaway
Claude Code 2.1.329 makes experimental Code Mode more observable and more bounded, while making hook failure behavior explicit. Teams testing it should verify telemetry coverage and configure fail-closed hooks themselves rather than assuming an error automatically denies a scripted tool call.