claude-code - 💡(How to fix) Fix [BUG] Routine cloud sessions resolve a different Gmail MCP server registration than interactive sessions on the same account, causing 100% draft-creation failure since ~2026-05-20 23:00 UTC.

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

On a single Anthropic account ([email protected]), two classes of Claude Code on the web sessions are resolving different Gmail MCP server registrations:

Interactive cloud sessions: ToolSearch for create_draft returns mcp__d91189ac-f000-45a9-a9ed-a7425f1e19f0__create_draft. Calls execute normally. Routine / non-interactive cloud sessions: ToolSearch for create_draft returns only mcp__Gmail__create_draft. Every call — including a 4-byte payload — returns: Streamable HTTP error: Error POSTing to endpoint: MCP tool call requires approval immediately, with no inline approval UI rendered to the session. list_drafts against the same named-alias registration fails identically, ruling out OAuth scope, payload, schema, name resolution at the client, and harness allowlist (mcp__Gmail__* is allowed; user has accepted "always allow"). Disconnect/reconnect of the Gmail integration does not change the binding.

Routine sessions previously created drafts successfully against this account for weeks. The regression cutover sits inside a ~5 hour window last night: a routine completing at 2026-05-20 18:46 UTC delivered its draft; the next routine at 2026-05-20 23:39 UTC, and every routine since, returns the error above.

What Should Happen?

This is a server-selection regression, not a client bug. Either the named-alias registration was switched to require per-call human approval (which non-interactive sessions structurally cannot provide), or routine sessions on this account were re-bound from the working UUID-prefixed registration to the named alias. We need either confirmation of what shipped in that window and a rollback, or a documented mechanism to force routine sessions to bind to the UUID-prefixed registration.

Error Messages/Logs

Streamable HTTP error: Error POSTing to endpoint: MCP tool call requires approval


Returned verbatim from every call to `mcp__Gmail__create_draft` in the broken session, including a 4-byte test payload:


to:      ["[email protected]"]
subject: "MCP test 2026-05-21"
body:    "test"


The identical error is returned by `mcp__Gmail__list_drafts` against the same registration. No HTTP status code, no JSON envelope, no stack trace — that single string is the entire response body.

The routine's own catch handler recorded the failure to `out/gmail_draft_id.txt` on the affected branches with this placeholder:


PENDING_APPROVAL — Gmail MCP create_draft was blocked by the harness (tool requires approval, no interactive session). The payload is fully built and committed. Open the draft manually via out/gmail_payload.json or out/gmail_draft_preview.html.


No client-side logs beyond that — the error returns immediately and the session has no further telemetry to attach.

Steps to Reproduce

1. On account [email protected], connect the Gmail integration on claude.ai/code (already connected and confirmed working).

2. Open a fresh interactive cloud session in any repo. In chat, run:
     ToolSearch query: "create_draft" max_results: 10
   Observe: returns exactly one tool named
     mcp__d91189ac-f000-45a9-a9ed-a7425f1e19f0__create_draft

3. In the same interactive session, call that tool with this minimal payload:
     to:      ["[email protected]"]
     subject: "MCP test interactive"
     body:    "test"
   Observe: returns a draft id (e.g. r4277309467001118109). The draft appears
   in Gmail Drafts within seconds.

4. Trigger a routine (non-interactive) cloud session on the same account. For
   our setup this is any of the Alaska.Ai column routines under
   Talonsturgill/linkedin-alaska-ai-weekly — concretely the latest is the
   Anchorage Desk routine that ran at 2026-05-21 13:39 UTC on branch
   claude/linkedin-desk-2026-05-21 (Claude Code session id
   01CJ8vp9fRRNknZ8kc7zuqEv per the PR body / session URL pattern). Any
   non-interactive trigger surface should reproduce — webhook, scheduled
   trigger, GitHub Action triggered job.

5. In that routine session, before the failing call, run:
     ToolSearch query: "create_draft" max_results: 10
   Observe: returns exactly one tool named
     mcp__Gmail__create_draft
   (different from step 2 — no UUID-prefixed variant is exposed in this
   session.)

6. In that routine session, call the tool with the same minimal payload:
     to:      ["[email protected]"]
     subject: "MCP test routine"
     body:    "test"
   Observe: returns the exact string
     Streamable HTTP error: Error POSTing to endpoint: MCP tool call requires approval
   No draft is created. No approval prompt renders.

7. In the same routine session, run:
     mcp__Gmail__list_drafts pageSize: 5
   Observe: returns the identical error string. The named-alias registration
   rejects every call, not just create.

8. Inspect .claude/settings.local.json in the repo root:
     {"permissions":{"allow":["mcp__Gmail__create_draft","mcp__Gmail__search_threads"]}}
   The harness allowlist matches the tool name being called. The error is
   not from the client-side approval gate.

Comparison artifacts for log lookup:

• Worked: Cold Take routine, session completed 2026-05-20 18:46 UTC, PR https://github.com/Talonsturgill/linkedin-alaska-ai-weekly/pull/8, branch claude/linkedin-contrarian-2026-05-20-02. Resulting draft id recorded: r-4765449699230921012.
• First failure: The Stack routine, session completed 2026-05-20 23:39 UTC, PR https://github.com/Talonsturgill/linkedin-alaska-ai-weekly/pull/10, branch claude/linkedin-stack-2026-05-20. out/gmail_draft_id.txt recorded the PENDING_APPROVAL placeholder.
• Still failing: Anchorage Desk routine, session completed 2026-05-21 13:39 UTC, PR https://github.com/Talonsturgill/linkedin-alaska-ai-weekly/pull/14, branch claude/linkedin-desk-2026-05-21. Same PENDING_APPROVAL placeholder.

The repository is public (or at minimum accessible to Anthropic support with the account context above) so the artifacts under each branch's out/ directory — including gmail_payload.json for the failed runs — can be inspected directly to confirm payload validity.

Minimal reproduction without routine triggers: if a routine surface isn't available for testing, the divergence can be observed by running step 2's ToolSearch from any interactive session on the account (returns UUID-prefixed) and comparing against the captured ToolSearch output from any of the three routine session logs above (returns named alias).

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.146 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Repository (public, all artifacts reproducible): https://github.com/Talonsturgill/linkedin-alaska-ai-weekly

Per-routine evidence — out/ artifacts on each failing branch:

• PR #14 Anchorage Desk: out/gmail_payload.json (the fully-built payload the routine generated) and out/gmail_draft_id.txt (the PENDING_APPROVAL placeholder the routine wrote when create_draft rejected it) on branch claude/linkedin-desk-2026-05-21.
• PR #10 The Stack: same two files on branch claude/linkedin-stack-2026-05-20. The gmail_draft_id.txt on this branch has since been updated to a real id (r1050369637301707005) — that id was fired from an interactive session (this one) using the UUID-prefixed registration, which is the exact divergence the bug is about. Commit a52a2f7 is the manual-fix; f... before it is the routine's original placeholder.
• PR #8 Cold Take (control — routine that worked): out/gmail_draft_id.txt recording draft r-4765449699230921012 on branch claude/linkedin-contrarian-2026-05-20-02. Same routine code, same payload shape, ran four hours before the cutover, landed.

Config files relevant to the bug:

• .claude/settings.local.json: {"permissions":{"allow":["mcp__Gmail__create_draft","mcp__Gmail__search_threads"]}} — confirms the failing tool name is on the harness allowlist, ruling out client-side approval gate.
• No .mcp.json, no .claude/mcp.json, no mcpServers block anywhere in the repo — rules out project-local MCP server overrides.

Current workaround in place:

The routines stop at writing out/gmail_payload.json and committing it. A human-in-the-loop interactive session (which resolves the UUID-prefixed registration) fires the draft from that JSON. Round-trip works every time. This is brittle — defeats the point of the routine — but unblocks shipping.

Screenshots that would help if your team wants visual confirmation (I can't capture from this container, but the account-holder can):

1. The chat panel of a failing routine session showing the verbatim Streamable HTTP error... string returned from create_draft.
2. The Claude Code on the web Integrations panel showing the Gmail integration listed once, with its "always allow" toggle set.
3. The Gmail Drafts UI showing the manually-fired drafts (proving the underlying OAuth grant is healthy and the account can receive drafts from MCP) versus the absence of the routine-generated ones.

Cross-reference, if useful: The Anchorage Desk PR (#14) is a "no defensible target this cycle" run — the email body is the validator's drop list rather than a finished post. It's a particularly clean reproduction because the failure is purely about Gmail transport, not about any content the routine was trying to ship. The other two failing branches (#10, ##14 again) failed identically with very different payload shapes (one full HTML post with image, one tabular drop-list). Payload content is not a variable.