[Bug] Per-message token usage spikes unexpectedly without configuration changes

Bug Description
Summary Per-message token/context usage in Claude Code spikes suddenly, with no configuration change on my end, and recurs after being stable for days. Environment - Claude Code version: 2.1.217 - Platform: macOS (Darwin 25.5.0) - Plan: Max (5x) Pattern - Usage is normal for several days of ordinary work (no new installs, no settings changes). - Then, with no action taken by me, per-prompt token/context usage jumps noticeably and stays high. - This has now happened twice. The first time, disabling local plugins seemed to coincide with it settling down — but this second time, the same plugins were already disabled (confirmed enabledPlugins: false for all 11 in ~/.claude/settings.json) and the spike still happened. - Since local plugin state wasn't the variable, I checked claude.ai account-level Connectors (Canva, Gmail, Google Calendar, Google Drive, Vercel MCP) and found them connected. I disconnected all five (kept GitHub Integration, which I actively use). I don't have before/after telemetry to confirm this was the actual cause — I'm not certain it is. What I've ruled out / checked - Local plugins (~/.claude/settings.json → enabledPlugins): all false at time of second spike. - Account Connectors: five were connected (Canva, Gmail, Calendar, Drive, Vercel), now disconnected. - No Claude Code version change at time of spike (already on latest, 2.1.217). - No changes to ~/.claude/CLAUDE.md, MCP server config (~/.claude.json, only clairvoyance_terminal configured locally), or desktop app MCP config that I made deliberately around either spike. What I'd like - Visibility into what actually changed underneath a session's context/token usage — ideally a per-session breakdown of what's contributing to prompt size (system prompt, skills, MCP tool defs, connector instructions, memory/CLAUDE.md), since right now diagnosing this requires manually auditing every subsystem with no single source of truth. - Confirmation on whether claude.ai account Connectors can silently reconnect or re-surface without user action, since that's the one variable I couldn't fully rule out from the client side. Impact This is disruptive for work I can't afford to have interrupted (client-facing analysis, job applications) and I have no way to predict or prevent it recurring.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.217
• Feedback ID: 8bcb411f-aa11-4f30-8f55-d207f3e03004

Errors

[]