Skip to content

[Bug]: Unrelated Claude MCP deny rules strip read-only hints from Datadog graphs, causing repeated approvals #32185

Open

Description

@hhharryyyyy

Before submitting

  • Searched existing issues and discussions; no duplicate of the cross-server read-only downgrade was found.
  • Included observed behavior, source references, and a reproduction against installed functions.

Area

apps/server (Claude MCP Apps adapter), visible in web/desktop app confirmations.

Problem

A deny rule for one MCP server causes T3's Claude adapter to remove the read-only hint from another server's MCP App tools. Native Datadog graphs then repeatedly show Allow datadog to run _dd_guc?, including in a Full access thread. Approving a call does not authorize the next one.

Datadog's live tools/list correctly supplies annotations.readOnlyHint: true for _dd_guc, _dd_gff, and _dd_wfd. The problem is the host's normalization, not missing annotations from Datadog.

Steps to reproduce

  1. Configure a Claude project with an unrelated MCP deny rule, for example:

    { "permissions": { "deny": ["mcp__buildkite__create_build"] } }
  2. Connect Datadog's MCP server, or another server exposing an MCP App whose background tools are annotated read-only. The unrelated server need not be the app's server.

  3. Use a Claude thread in Full access mode and generate a native Datadog graph.

  4. Let the card load or reload. Its read-only helper call triggers the confirmation dialog.

Expected behavior

An unrelated Buildkite rule should not downgrade Datadog's read-only tools. Correctly annotated read-only graph calls should retain their normal prompt-free behavior. Rules that actually cover the requested server/tool must still be enforced. Removing project safety rules or disabling native graphs is not an acceptable fix.

Actual behavior and evidence

In claudeRulesMayCoverMcpTools, any active ask/deny rule whose tool name starts with mcp__ can match every server. claudeMcpToolForApps then removes the app tool's read-only hint. McpAppFrame confirms every call for which info.value.readOnly is false.

I extracted claudeRulesMayCoverMcpTools and claudeMcpToolForApps unchanged from the installed desktop bundle and invoked them with:

const tool = { name: "_dd_guc", annotations: { readOnly: true } };
const unrelated = [{ behavior: "deny", rule: "mcp__buildkite__create_build" }];
const withoutRules = claudeMcpToolForApps(tool, []).annotations.readOnlyHint;
const withUnrelatedDeny = claudeMcpToolForApps(tool, unrelated).annotations.readOnlyHint;
console.log({ withoutRules, withUnrelatedDeny });
// { withoutRules: true, withUnrelatedDeny: undefined }

The native dialog was observed during real Datadog graph usage. Live tool discovery and the extracted-function reproduction were verified. A fresh-project UI A/B run with and without the rule has not been performed, so that end-to-end control is not claimed.

Impact

Major degradation: background graph loading produces repeated blocking permission dialogs despite read-only annotations and Full access mode.

Version and environment

  • T3 Code Nightly 0.1.8-nightly.20261106.4275, macOS arm64 desktop.
  • Claude provider, Full access thread.
  • The rule check arrived with Claude MCP Apps support in feat(server): render MCP Apps from Claude tool calls #29803 (merged November 4). The November 4 .4151 source lacks it; November 6 .4247, .4262, and .4275 contain it. Local app logs show updates/restarts overnight and this morning. This supports the regression timing, but the previous desktop build was not run for comparison.

Settings and workaround

The app confirmation path does not consult thread permission mode or remembered per-server grants. No applicable MCP App approval setting was found in the shipped settings schema. A Codex MCP approval override does not affect this Claude app path.

Filtering out the native graph tools prevents cards from being generated, but removes wanted functionality, so that workaround was reverted. Native graphs remain enabled.

Related, distinct request

Discussion #32084 proposes inline/batched approvals and remembered grants for app tools that need approval. This report concerns read-only hints incorrectly stripped by an unrelated server's permission rules, independently of that UX proposal.

Activity

  1. jeanpierre121 commented on Nov 14, 2026

    @jeanpierre121

    Same bug on stable 0.1.9 (macOS, Claude provider, Full access thread), this time with Superhuman Mail's draft card. The card calls get_draft_send_status and each call opens "Allow superhuman-mail to run get_draft_send_status?". My Claude settings have 13 deny rules, all for other servers (Grain, Granola, Gmail, Notion, Attio) and none for superhuman-mail.

    One thing the report above doesn't cover: Confirm doesn't help, because the server refuses the call after I approve it, so the card retries and the dialog comes back. My server.trace.ndjson from 13 Nov has 24 ClaudeAdapterV2.callMcpTool spans for this tool and all 24 failed with:

    ProviderAdapterProtocolError: claudeAgent provider protocol error: A Claude deny rule may cover mcp__superhuman-mail__get_draft_send_status, so it was not called.
    

    #32230 looks like it fixes my case. For the paths that keep the conservative guard (older CLIs, tools without a read-only annotation), the client still asks me to approve a call the server will always refuse - could toolInfo report those as not callable so the dialog is skipped?

    I haven't run the PR, and I haven't checked if Superhuman annotates this tool as read-only.

  2. andywillis101 commented on Nov 14, 2026

    @andywillis101

    Another case of this, with the GitHub MCP server (remote, api.githubcopilot.com/mcp/) on Claude threads.

    What happens. After the agent calls update_pull_request, T3 renders the server's pr-edit view. The view loads the PR with callTool("pull_request_read", { method: "get", owner, repo, pullNumber }) (source). T3 shows "Allow github to run pull_request_read?" with only Cancel and Confirm, on every PR edit, in a thread with bypass permissions. Confirm doesn't help: the server-side call is then refused with "A Claude deny rule may cover pull_request_read, so it was not called." The view stays empty either way.

    Why. My project settings have two mcp__ ask rules and seven mcp__ deny rules, all for writes:

    • ask: mcp__github__merge_pull_request, mcp__github__merge_*
    • deny: mcp__github__push_files, mcp__github__push_*, mcp__github__create_or_update_file, mcp__github__delete_file, mcp__github__delete_*, mcp__github__create_branch, mcp__github__update_pull_request_branch

    mcp__github__pull_request_read is explicitly in my allow list. claudeRulesMayCoverMcpTools treats any mcp__ ask or deny rule as covering every MCP tool. So claudeMcpToolForApps drops readOnlyHint, McpAppFrame asks before the call, and the mcp_call path then refuses it on the deny rules. I understand the broad match is there for server aliases T3 can't list. The result is that a merge gate and a few write denies on one server block every read made by that server's views.

    Why #32230 won't fix this case. Its native check_permissions path applies only in Full Access with no ask rule that may cover the tool. My merge ask rule is the one I need to keep, so I stay on the conservative path.

    Suggestions:

    1. When the CLI supports mcp_call_check_permissions_v1, use it whenever a read-only tool's call is checked, whatever ask rules exist. If Claude's own evaluator returns "ask", show the card then. That covers aliases without guessing from rule names.
    2. Treat an explicit allow rule for the exact tool as permission for the app's call.
    3. Add "Always allow this session" to this dialog, as fix(server): Claude threads can allow an MCP app for the session #33707 did for MCP elicitation cards.

    Until then, the only workaround I have is pressing Cancel. The agent's own pull_request_read calls are unaffected, so only the view is lost.

  3. ascrazy commented on Nov 19, 2026

    @ascrazy

    Additional evidence for #32185 on stable 0.1.10, with the repeated-dialog behavior described in #38950.

    What happened

    Switching conversations through the sidebar appears to trigger roughly 5–10 successive Allow datadog to run _dd_gff? dialogs, then they stop. The screenshot shows {"telemetry":{"intent":"Fetch feature flags"}}. The user reports the navigation association as tentative.

    Environment

    Triage was started from macOS T3 Code Desktop 0.1.10, connected to a Linux x64 systemd server running 0.1.10, Node 26.13.1. Recent conversations containing Datadog widgets use Claude with Full access. The exact conversation shown in the screenshot has not been identified.

    Diagnosis and evidence

    • The project has 16 deny rules for specific Datadog write/feedback tools, including mcp__datadog__delete_datadog_dashboard. None names _dd_gff.

    • Running the unchanged v0.1.10 claudeRulesMayCoverMcpTools and claudeMcpToolForApps functions with the real project deny rules and a synthetic _dd_gff tool annotated readOnly: true produces:

      readOnlyHint with no rules: true
      readOnlyHint with project rules: removed
      deny guard with project rules: true
      
    • The Claude adapter treats any active MCP deny rule as potentially covering every tool. This also affects different tools on the same server.

    • McpAppFrame prompts per call without consulting Full access or requiring a widget click. The global confirmation queue displays pending dialogs successively.

    • Between 14:20:15.102 and 14:20:33.875 UTC on November 19, immediately before the screenshot, the trace records 39 successful McpAppRequests.toolInfo spans and 39 ClaudeAdapterV2.listMcpTools spans. No app tool-execution spans appear in that interval. These counts are metadata lookups, not counts of dialogs.

    • Two recent Claude Full access conversations contain 8 and 12 persisted Datadog app widgets. The user suspected two conversations created from Linear issues containing Datadog links; reading these conversations confirmed that their opening messages contain Linear issue context and Datadog links. The agents subsequently called Datadog tools that produced the embedded widgets. Loading multiple widgets is consistent with the observed bursts. Private issue identifiers and URLs are omitted.

    Reproduction limits and workaround

    The source-level permission reproduction is verified; switching conversations was not replayed in the UI. Trace spans do not identify the requested tools, and live _dd_gff annotations were not fetched. No configuration was changed. Full access is already enabled, and removing the intended write-deny rules would weaken the user's restrictions.

    #32230 was closed without merging; #38962 remains open and addresses background widget calls opening dialogs. This evidence belongs on the existing report rather than a new duplicate.

    Investigated by GPT-6 Astra through the Codex harness in T3 Code.

  4. andreaseger commented on Nov 19, 2026

    @andreaseger

    Still reproducing on 0.1.11-nightly.20261117.4825 (macOS desktop, Claude provider, Claude Code CLI 2.1.328). Two details aren't covered above.

    Only ask rules, all on a different server. The project has no deny rules. Its only MCP rules are exact-tool ask rules for write tools on another server, such as mcp__atlassian__createJiraIssue and mcp__atlassian__editJiraIssue, with no wildcards. User-level settings have no MCP rules. That alone is enough for claudeRulesMayCoverMcpTools(rules, ["ask", "deny"]) to strip the read-only hint from every Datadog graph helper. The title says deny rules, but an unrelated ask rule triggers it too.

    Live annotations confirmed. I sent control requests directly to the Claude CLI T3 launches, from the same project directory:

    • mcp_status reports Datadog's graph helpers as read-only app tools:
      _dd_gff  annotations={readOnly: true}  _meta.ui={resourceUri: "ui://apps/dataviz", visibility: ["app"]}
      _dd_guc  annotations={readOnly: true}  _meta.ui={resourceUri: "ui://apps/dataviz", visibility: ["app"]}
      _dd_wfd  annotations={readOnly: true}  _meta.ui={resourceUri: "ui://apps/dataviz", visibility: ["app"]}
      
    • list_permission_rules returns no errors, so the rules are readable and the rules === null fallback isn't involved. The only ask or deny rows starting with mcp__ are the other-server ask rules above.

    Impact. Datadog calls that render a graph card are followed by Allow claude.ai Datadog to run _dd_gff? with {"telemetry":{"intent":"Fetch feature flags"}}. When several cards mount at once, up to about 10 dialogs queue up (as in #38950). No choice is remembered, so they come back on reload.

    #38962 would stop dialogs for calls made without a click. It wouldn't restore the read-only hint, though, so a click inside a Datadog graph would still prompt for a read-only call. The CLI advertises mcp_call_check_permissions_v1. Asking Claude's own evaluator about the specific call, as suggested in an earlier comment, avoids guessing from rule names. It would also keep a real ask rule on an Atlassian write tool from affecting Datadog's read-only graph calls.

    Investigated with Claude Code (Opus 5.5) in T3 Code.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions