- 7 min read
- 8 key claims
- 4 sources cited
- For Developers, AI makers, Businesses
Overview
Anthropic's release notes for Claude Code v2.1.325 (commit 715ff10) through v2.1.328 describe a release set that combines a security fix with routing, compaction, credential and background-session corrections in v2.1.325, followed by diagnostics and deployment features in v2.1.326 through v2.1.328 [1][2][3][4]. The shared thread is operational reliability: the changes concentrate on gateways, managed settings, hooks, telemetry and background sessions rather than on new end-user coding capabilities [1][2][3][4].
Why it matters
These fixes address the failure modes that appear once an agentic tool is placed inside a real engineering system rather than a single developer laptop: work lost on resume, permissions applied incorrectly, credentials crossing process boundaries, and settings that appear saved but are not [1][2][3][4].
For operators, several entries have direct action attached:
- Deployments routing through third-party gateways or Microsoft Foundry should upgrade to pick up the compaction fix, since the previous behaviour was a context length error on every turn rather than compaction [1].
- Teams using the allowedProviders managed setting should verify token-counting behavior on Bedrock, where conversations were previously sent to Bedrock despite the setting [1].
- Teams using OpenTelemetry should check exporter endpoints; v2.1.328 adds a notice in /status and claude doctor when an exporter has no endpoint and its data would likely be lost [2].
- MCP operators can move JSON definitions off the command line by passing them through stdin, with the exact invocation shown as
claude mcp add-json <name>in the full v2.1.327 notes and asclaude mcp add-json -in a partial rendering [3]. - Teams running Claude Code on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or the Claude apps gateway should note that interactive sessions with no permission mode now start in auto mode, and that a message from a session that skips permission prompts may be held and dropped by a cloud session [2].
Opportunities
Managed configuration is becoming a supported surface. The v2.1.326 to v2.1.328 notes add startup validation for tool-rule entries, explanations for managed-mcp.json changes made mid-session, /status and claude doctor diagnostics, a note in /config when other settings override a row, and an x-claude-code-managed-settings-checksum sent in gateway hint headers [3][4]. That is a coherent surface for configuration-management tooling that distributes MCP server definitions and permission rules to many developer machines without touching a shared file [3][4].
Self-hosted operation gains measurable signals. The v2.1.328 notes add --repository-folder for naming the folder sessions use at --capacity 1, --hook-label and --metadata key=value for labelling sessions, a metric that distinguishes a re-opened session's queue wait from a new one, and a counter of how often Anthropic gave up starting a runner for a session [2]. Teams running capacity-limited runners have a basis for capacity-planning and session-scheduling tooling [2].
Two behaviors change what operators must handle. Fail-closed MCP hooks mean a flaky hook can now block an action that previously proceeded, so hook reliability becomes part of pipeline reliability [2]. Auto mode as the default for interactive sessions with no permission mode reduces prompts before an action runs, while the new cloud-session warning notes that messages from sessions which skip permission prompts may be held and dropped [2].
Risks
The zstd decoding issue is described by Anthropic only as a security issue, with no severity rating, advisory identifier, CVE, affected versions or technical description, so the release notes alone do not let a team judge exposure or confirm whether earlier builds carry it [1].
Two default-behavior changes carry operational weight. Auto mode becomes the starting point for interactive sessions with no permission mode on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, which means fewer prompts before an action runs, alongside a warning that a message from such a session may be held and dropped by a cloud session [2]. The fail-closed default for mcp_tool hooks improves safety but can block an action whenever a hook is flaky, adding cost to any pipeline that depends on MCP tools [2].
Several fixes point to continuing fragility in multi-agent and background execution: subagent model selection, permission modes, background sessions and account switching all needed correction across these releases [1][2][3][4]. The v2.1.328 notes also record a regression in 2.1.327 affecting subagent effort handling, and the available text does not establish whether the separate v2.1.325 subagent-model fix was later regressed [1][2]. Configuration and policy formats remain vendor-defined across managed settings, policy plugins, gateway hint headers and SDK-level Admin API support, so teams building on them carry exposure to changes in those formats [2][3].
Context
Claude Code is described by Anthropic as an agentic coding tool that runs in the terminal, understands a codebase, and executes routine coding tasks, explains complex code and handles git workflows through natural-language commands (vendor_claim) [1][2][3][4]. That description comes from the release pages themselves, and no independent evaluation is included.
The v2.1.325 notes cover four operational areas. First, a security issue in how zstd-compressed responses are decoded was fixed [1]. Second, sessions on many third-party gateways and Microsoft Foundry no longer stop with a context length error on every turn instead of compacting, and subagents with their own model are no longer moved to the plan-mode model under opusplan or haiku settings or sized with the main session's model [1]. Third, token counting on Amazon Bedrock no longer sends the conversation to Bedrock when the allowedProviders managed setting does not allow it, and AWS credential_process helpers and Google Cloud default credentials no longer fail when Claude Code starts without a HOME environment variable [1]. Fourth, several session-state problems were corrected, including background sessions started on a claude.ai login using and billing an API key or token from the shell that started the background service, and more cases of a session moving to another sign-in in the background, which are now refused [1].
Later releases in the set add deployment-facing features. v2.1.326 added Code Mode support alongside agent teams, startup warnings for malformed tool-rule entries and for managed-mcp.json changes made mid-session, a /status row indicating when a provider or environment variable such as DISABLE_TELEMETRY turns features off, and the OTEL_RESOURCE_INCLUDE_HOST setting [4]. v2.1.327 added claude mcp add-json <name> to read a server's JSON from stdin so tokens stay off the command line; a partial rendering of the same notes shows the command as claude mcp add-json -, so the exact form is not settled by the available text [3]. The same release added a /status diagnostic list covering broken agent, skill, command and output style files, keybinding errors, an unavailable sandbox, ignored permission settings and hook matchers that never match [3]. v2.1.328 added auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, with interactive sessions that have no permission mode starting in it, plus "onFailure": "block" for mcp_tool hooks, an OpenTelemetry endpoint notice, and self-hosted runner labels and orchestrator metrics [2].
Known, claimed, inferred
One inference holds across the set: because the fixes cluster around gateways, managed settings, hooks, telemetry and background sessions rather than around new coding capabilities, the direction of travel in these four releases is toward enterprise and self-hosted deployment of agentic coding [1][2][3][4]. The addition of self-hosted runner labelling and queue metrics, the diagnostics surfaces in /status and claude doctor, and the managed-settings checksum sent in gateway hint headers are consistent with that reading [2][3][4].
The two changes that alter default behavior sit on either side of that deployment story. Auto mode now applies to interactive sessions without a permission mode on the major hosted platforms, and mcp_tool hooks fail closed when their call fails [2]. Both reduce or bypass a step that previously interrupted the user, which shifts part of the control burden onto the settings and monitoring around Claude Code rather than onto the operator at the prompt [2].
Limitations & uncertainties
The evidence consists only of vendor-authored Anthropic release notes, so the fixes and their scope are not independently verified, and no metrics, user counts, performance figures, cost data or adoption data are provided for any version [1][2][3][4]. The notes for v2.1.326, v2.1.327 and v2.1.328 are truncated, so their complete change lists are unknown [2][3][4]. No release dates are given for v2.1.325 through v2.1.328, only evidence retrieval dates, so the cadence between releases cannot be established, and no release notes for versions before v2.1.325 are available for comparison [1][2][3][4]. It is not stated whether the v2.1.325 fixes were backported to earlier versions, and there is no separate security advisory or patch announcement [1]. The claude mcp add-json invocation differs between the full and partial renderings of the v2.1.327 notes, and the available text does not settle which form is correct [3].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementClaude Code v2.1.325 fixed a security issue in how zstd-compressed responses are decoded.[1]
- Primary-source or vendor statementSessions on many third-party gateways and Microsoft Foundry no longer stop with a context length error on every turn instead of compacting.[1]
- Primary-source or vendor statementSubagents with their own model are no longer moved to the plan-mode model under opusplan or haiku settings, and subagents on a different model no longer size file reads and MCP results with the main session's model.[1]
- Primary-source or vendor statementv2.1.328 added auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, and interactive sessions with no permission mode start in it.[2]
- Primary-source or vendor statementv2.1.328 added "onFailure": "block" for mcp_tool hooks, so a hook whose call fails blocks the action.[2]
- Primary-source or vendor statementv2.1.326 added support for turning on Code Mode while agent teams are enabled, and v2.1.327 added a /status diagnostic list covering broken agent, skill, command and output style files, keybinding errors, an unavailable sandbox, ignored permission settings and hook matchers that never match.[3]
- HypothesisThe direction of travel across these four releases is toward managed and self-hosted deployment of agentic coding.[2]
- Primary-source or vendor statementClaude Code is an agentic coding tool that runs in the terminal, understands a codebase, and executes routine coding tasks, explains complex code and handles git workflows through natural-language commands.[1]
Sources
- 1anthropics/claude-code v2.1.325●github.comretrieved Nov 18, 2026
- 2anthropics/claude-code v2.1.328●github.comretrieved Nov 19, 2026
- 3anthropics/claude-code v2.1.327●github.comretrieved Nov 18, 2026
- 4anthropics/claude-code v2.1.326●github.comretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.