[BUG] claude.ai chat unusable for entire account: server-side ~4.8M-token blob injected into prompt assembly (all models, all surfaces, 1+ week) — follow-up to #62175

Preflight Checklist

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

What's Wrong?

Filing here because claude.ai has no public bug tracker and this error class is already documented in #62175 (closed as not planned). Anthropic support has not responded for 4+ days (support conversation ID 215475070708461). This report is precise enough for engineering to root-cause from logs.

Since ~2026-07-10, nearly every message in Claude chat fails for my account with:

This request exceeds Claude's context limit by about N tokens. Try turning off some tools or shortening your message.

• Happens on brand-new, empty conversations with a 4-character prompt ("test")
• All models (Haiku 4.5, Opus 4.5 verified via API; all models via UI)
• All surfaces: claude.ai web, Claude Desktop, iOS app, incognito browser sessions
• N drifts between 4,625,481 and 5,863,234 tokens over the week
• Claude Code and Cowork on the same account work normally

Diagnosis via API bisection (2026-07-17, 09:20-10:00 UTC), all on fresh conversations:

• Overage is identical across models (Haiku 4.5 vs Opus 4.5: delta of 10 tokens) -> a single fixed blob of ~4.84M tokens (~19 MB) is injected server-side during prompt assembly
• Fully independent of the client payload: fails identically with tools: [] on an API-created conversation
• Ruled out while failing: all 11 account connectors disconnected, memory deleted + both memory toggles off, skills/plugins/styles < 15 KB total, local desktop config removed (also fails on mobile and incognito)
• The blob intermittently disappears for minutes and returns: directly observed one ~7-minute success window (09:40-09:47 UTC) in which identical requests suddenly succeeded, then failed again at 09:55 with a new overage value (4,625,481, three identical runs). Matches the past week's pattern of rare single messages succeeding
• Side effect: account memory NEVER rebuilds after deletion (memory endpoint stays empty for days) - consistent with background memory/summarization jobs dying on the same oversized context

SSE error detail:

event: error
data: {"type":"error","error":{"type":"invalid_request_error","message":"This request exceeds Claude's context limit by about 4,641,642 tokens. Try turning off some tools or shortening your message.","details":{"error_visibility":"user_facing"}},"request_id":"req_011Cd7Qwe45kCCijPaknN2wb"}

Request IDs for log correlation (2026-07-17): failing req_011Cd7Qwe45kCCijPaknN2wb (~09:25 UTC, tools:[], Haiku 4.5, overage 4,641,642) vs succeeding chatcompl_011Cd7RQ5v4VstraeuxQsV6p (~09:40 UTC, same account, near-identical payload). Diffing what prompt assembly injected between these two requests should directly reveal the ~19 MB object.

This is critical: the primary paid product (chat) is completely non-functional for the account, and the cause is invisible to every user-facing setting and API.

What Should Happen?

A brand-new empty conversation with the prompt "test" and zero tools must never exceed the context limit. Whatever per-account object prompt assembly injects (index/summary/cache) must be size-bounded, and the error message should say WHAT is consuming the context - a week of methodical user-side debugging (connectors, memory, skills, plugins, styles, models, devices) could not reveal the source because it is invisible to every user-facing setting.

Ask to Anthropic: escalate to engineering, diff the two referenced requests in prompt-assembly logs, identify and clear/bound the injected object for org 73ffd8f5-75cc-4c02-a683-598bbc758ba2, and check the periodic rebuild job that makes the blob vanish for minutes and return (likely the same job preventing memory from rebuilding).

Error Messages/Logs

event: error
data: {"type":"error","error":{"type":"invalid_request_error","message":"This request exceeds Claude's context limit by about 4,641,642 tokens. Try turning off some tools or shortening your message.","details":{"error_visibility":"user_facing"}},"request_id":"req_011Cd7Qwe45kCCijPaknN2wb"}

Observed overage values: 5,863,234 (desktop, Jul 16) / 4,668,050 (web, Jul 16) / 4,660,121-122 (web UI, Jul 17) / 4,657,180 (API replay) / 4,641,642 (tools:[], Haiku 4.5) / 4,641,652 (tools:[], Opus 4.5) / 4,625,481 (3 identical runs, Jul 17 09:55 UTC)

Steps to Reproduce

1. Log into the affected claude.ai account (account 825ad336-6e6d-4dfa-92d9-7dad025a8070, org 73ffd8f5-75cc-4c02-a683-598bbc758ba2 - Anthropic staff can verify server-side; this is not reproducible on arbitrary accounts, which is the point: it is corrupted per-account server state)
2. Start a brand-new chat on claude.ai (web, desktop or iOS; incognito works too)
3. Send the message "test"
4. Error appears immediately: "This request exceeds Claude's context limit by about ~4.6M tokens"
5. Equivalent API repro: POST /api/organizations/{org}/chat_conversations/{uuid}/completion with {prompt:"test", tools:[], attachments:[], files:[], model:"claude-haiku-4-5-20251001", rendering_mode:"messages", parent_message_uuid:"00000000-0000-4000-8000-000000000000", timezone:"Europe/Berlin"} on a freshly created conversation -> same error
6. Repeat for minutes/hours: occasionally (rarely) a request succeeds for a few minutes, then the failure returns with a slightly different overage value

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

n/a - claude.ai chat product; worked normally until ~2026-07-10

Claude Code Version

n/a - bug is in the claude.ai chat backend (prompt assembly); Claude Code itself is unaffected on this account

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Related: #62175 (closed as not planned) documents the same error string caused by connector schemas; my case is account-level and persists with ZERO connectors - I posted the full bisection as a comment there before filing this. Support conversation ID: 215475070708461 (Fin escalated to human team 4 days ago, no response). Public thread with screenshots: https://x.com/iret77/status/2077906079732715951

Happy to provide the full markdown report, additional request IDs, or run further diagnostics on request - the account is available for server-side inspection.