/doctor prompt-audit, and ~50 Operational FixesAnthropic tagged Claude Code v2.1.283 stable at 2026-09-25 21:50 UTC. The release is the largest in the v2.1.27x series by item count — 50+ documented changes spanning enterprise model governance, LLM-gateway observability, AWS Bedrock coverage, plugin validation, MCP session lifecycle, prompt-audit tooling, Windows shell safety, and a long tail of UX, latency, and stability fixes. Yesterday's v2.1.282 was a focused managed-settings hardening; v2.1.283 is the broad operational release the hardening chain was pointing at.
Five changes are first-class for builder / operator / founder readers. The other ~45 are operational polish worth a single sentence each.
1. availableModelsMatch managed setting. Added with "exact" mode: when an availableModels entry is set to "exact", the entry allows only the model version it names, so new releases stay blocked until listed in managed settings. Practically: teams that pin models for SOC 2 / HIPAA / FedRAMP / regulated workloads can now have the gate enforced by Claude Code itself, rather than by a wrapper or by reminding engineers. 2. deniedModels managed setting. A separate explicit-deny list, applied even when availableModels allows a model. Two-axis governance: allow-list and deny-list, both managed-policy-driven. This is the first time Anthropic has shipped two-axis model gating at managed-settings level. 3. x-claude-code-prompt-id gateway hint header. Added to outbound LLM-gateway requests, opt-in with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1. Groups all model calls that serve a single user prompt, so gateways (Cloudflare AI Gateway, Portkey, LiteLLM, Helicone, OpenRouter) can bill per prompt, rate-limit per prompt, or audit per prompt — rather than per HTTP request. v2.1.273 added the gateway hint header scaffold (x-claude-code-session-id); v2.1.283 layers the prompt-granularity identifier on top. 4. Amazon Bedrock Mantle upstream provider. The Claude apps gateway now supports mantle as an upstream, joining the existing bedrock, vertex, foundry, and WIF providers. Bedrock Mantle is the AWS-distributed inference tier with separate routing and quota semantics — relevant for AWS-resident agent stacks that need to keep inference inside the AWS VPC boundary. 5. /doctor prompt-audit (also /checkup prompt-audit). A new command that audits CLAUDE.md files, skills, agents, and commands for prompting patterns written for older Claude models. The release notes flag stale paths, stale commands, and contradicting instruction files as leading the report, and explicitly keep thinking keywords that current Claude Code documents. For teams carrying prompt conventions forward across Claude 3 → 4 → 5 → Fable 5.1 → Mythos 5.1 → Opus 5.5, this is a workflow audit tool.
The other 45+ items worth surfacing because they will surface in user reports:
/context now counts MCP server instructions as their own row and toward the total./mcp tool list shows more tools at once, scrolls with the page keys and mouse, and marks tools your organization blocked with a warning icon. MCP tool results: images returned by MCP tools are now also saved to a file, so Bash, Read, and other tools can open them.claude plugin validate now refuses names that Claude Code cannot actually install, and refuses plugins whose outputStyles, themes, monitors, or lspServers paths are missing or point outside the plugin directory — closing a class of plugin install bypasses that v2.1.275 and v2.1.282 were tightening. installed_plugins.json no longer loses records when it holds a record this version cannot read; claude plugin commands now name the record and say how to recover. claude plugin uninstall no longer removes the other of two installed plugins whose ids differ only in case.claude mcp add, add-json, and remove now report failure when the config file could not be written (for example, inside a sandbox). MCP sign-in for a server with no valid URL no longer fails with an opaque SDK error; /mcp does not offer Authenticate for such servers.claude -p and Claude Code Remote no longer load the interactive UI; the auto-mode classifier's rules and the Artifact tool load on first use instead of at launch.DISABLE_PROMPT_CACHING_HAIKU is now honored when Haiku is the session's main model./usage fix. The weekly Fable limit now appears in /usage and the VS Code usage meters when telemetry is disabled./model picker cleanup. Drops "(1M context)" where Opus already has a 1M context window (the window is unchanged); stops showing a hardcoded Haiku version and price when ANTHROPIC_DEFAULT_HAIKU_MODEL pins a different model; stops accepting Sonnet 4.6 or Sonnet 5 with [1m] when the id carried a date or -v1:0 suffix, in the cases where the plain id was refused.V then p now lands on the first non-blank; J joins lines with the right spacing; 3J or Visual-mode J on the last line moves the cursor as Vim does.keybindings.json warns on a misspelled modifier such as ctl+k; footer hints no longer say "Enter to view" after footer:openSelected was rebound or unbound.cmd /c rd, rmdir, del, or erase delete drive roots, the home folder, or other folders that Remove-Item would refuse.git no longer asks credential helpers to store the sandbox proxy's login; managed sandbox settings are no longer ignored entirely when one nested value is invalid — the invalid value now fails closed and the rest of the block still applies.permissions.defaultMode still overrides it.tool.output OpenTelemetry span event when OTEL_LOG_TOOL_CONTENT=1. This is a meaningful new content-logging surface — see the cost/risk/limitations section below.load_test_mode block to the Claude apps gateway config: requests are built and signed but not sent upstream, and clients get a canned reply, so a deployment can be load-tested./ultrareview disclosure. The launch dialog now says that reviewing a local branch may upload uncommitted changes to tracked files.For enterprise governance teams, the availableModelsMatch: "exact" + deniedModels pair is the single most consequential change in the release. Until today, an availableModels: ["claude-opus-5"] entry would silently permit claude-opus-5-5 once 5.5 was GA — that is exactly the failure mode that prompted the v2.1.280 Opus 5.5 default-flip coverage. With "exact" mode, a managed-settings upgrade policy can require that every model upgrade be explicitly approved and listed, and deniedModels gives a hard-deny carve-out for models you do not want even when the allow-list allows them.
For LLM-gateway operators, the prompt-id header closes a long-standing attribution gap. Today, gateway-side cost attribution is per-request, but a single user prompt can fan out to 5–50 model calls (tool calls, subagent spawns, retries, long-context streams). Per-prompt grouping means you can finally answer the question "what does this conversation cost?" at the user level rather than approximating from request volume.
For AWS teams, Bedrock Mantle coverage means the Claude apps gateway can route to AWS-native inference without an external Anthropic API credential — relevant for VPC-bound deployments and for AWS procurement that prefers staying inside the AWS billing perimeter.
For Windows developers, the PowerShell tool fix closes an accidental-destruction guardrail gap: cmd /c rd, rmdir, del, and erase could delete drive roots and the home folder — paths that Remove-Item would refuse. This is the class of fix that matters when an agent runs destructive commands on Windows.
For teams carrying prompt conventions forward across Claude model generations, /doctor prompt-audit is the new workflow audit tool. CLAUDE.md files written for Sonnet 4.6 or Opus 4.7 carry patterns that don't transfer cleanly to Fable 5.1, Mythos 5.1, or Opus 5.5 — prompt-audit flags them and points at the fix.
All evidence is documentation-surfaced from the official anthropics/claude-code CHANGELOG.md and the v2.1.283 GitHub release notes (atom feed entry id tag:github.com,2008:Repository/937253475/v2.1.283, updated 2026-09-25T21:50:12Z). The CHANGELOG.md release notes were verified verbatim at fetch 2026-09-26 14:09 UTC.
Verification level: documentation comparison + CHANGELOG.md verbatim quotes. No firsthand API call run for this article. The article does not claim to have run v2.1.283 against a Claude Code install; every feature, fix, and PR-number claim is documentation-derived from the cited primary sources.
Independent corroboration: availableModelsMatch and deniedModels settings are documented in the v2.1.283 CHANGELOG.md release notes section "What's changed." The x-claude-code-prompt-id header description is documented in the same section. Bedrock Mantle support is documented in the same section. The /doctor prompt-audit command description is documented in the same section. Each item has a single primary source; for a release of this breadth, that is the standard desk-evidence threshold.
Cost. $0 to read these changes; no model calls were made for this article. The OpenTelemetry tool.output content-logging toggle (OTEL_LOG_TOOL_CONTENT=1) adds no direct cost beyond the OTLP span size — but the tool output can include credentials, secrets, or PII depending on what your tools returned, so the operational cost is on the collector-side redaction pipeline.
Risk. Three settings carry deployment risk if flipped without testing:
CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 adds a custom HTTP header to outbound LLM-gateway requests; some proxies or WAFs may need allowlist updates before the header is accepted.OTEL_LOG_TOOL_CONTENT=1 writes MCP, WebFetch, and WebSearch tool outputs to your OTLP collector. Treat the OpenTelemetry content-logging toggle as PII-grade and configure collector-side redaction before turning it on in production.availableModelsMatch: "exact" on a managed-settings block that already has availableModels will silently block any new Claude model release from being used until you also add the new model ID to the list. Plan the rollout against your model-upgrade cadence.Limitations. The article does not claim to have run v2.1.283 against a Claude Code install. The new /doctor prompt-audit output for any specific CLAUDE.md corpus is not empirically tested here. The Bedrock Mantle endpoint name and routing semantics are taken from the CHANGELOG.md entry and not from a Bedrock-side fetch. The deniedModels setting semantics — how it interacts with availableModels when both are present, whether it supports glob patterns, and whether managed settings can scope it per-project — are not exercised here and would benefit from a follow-up.
This is the operational release the v2.1.27x through v2.1.282 chain has been pointing at. The previous releases closed attack chains, hardened managed-settings plumbing, and added /doctor; v2.1.283 turns /doctor into a prompt-audit tool, adds the two-axis model gating that enterprise admins have been asking for, ships the prompt-id gateway header that LLM-gateway operators have been asking for, and lands the Windows PowerShell guardrail. For regulated teams, the availableModelsMatch: "exact" flag is the single most consequential change in the release — pin your managed settings today.
1. Today (regulators / enterprise admins). Review your managed settings. If you maintain an availableModels list, decide whether "exact" mode is the right setting; add deniedModels if you have specific model families you want blocked even when allow-listed. Plan the rollout against your model-upgrade cadence. 2. Today (LLM-gateway operators). Plan to set CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 on test traffic and validate that your gateway sees the prompt-id header on every request for a single user prompt. Update WAF / proxy allowlists if needed. 3. Today (AWS teams running Bedrock). Confirm whether Bedrock Mantle fits your routing / quota profile. If yes, configure mantle as an upstream in your Claude apps gateway config. 4. Today (Windows developers). Upgrade to v2.1.283 before running the next session that uses cmd /c rd, rmdir, del, or erase through PowerShell. 5. This week. Run /doctor prompt-audit on your CLAUDE.md, skills, agents, and commands; treat leading findings (stale paths, stale commands, contradicting instruction files) as priority fixes. 6. Skip if not in scope. If you don't operate a regulated Claude Code install, an LLM gateway, or a Bedrock-routed deployment, this release is mostly operational polish. Update when convenient.
tag:github.com,2008:Repository/937253475/v2.1.283, updated 2026-09-25T21:50:12Z).