Summary

• The Claude chat harness started every turn with ~194k of a 200k context window taken by tool schemas, so any turn with a few tool calls hit the SDK's "Autocompact is thrashing" guard and showed users "Agent Interrupted". Two causes: ToolSearch was missing from options.tools (the SDK restricts built-ins to the names listed, so tool-search deferral was silently off and all MCP schemas loaded every turn), and alwaysLoad: true on every user MCP server pinned their schemas into context even with deferral on.
• Fix: name ToolSearch in the allowlist (new pure, tested buildAllowedTools) and drop alwaysLoad from user MCP server configs. Tools are now discovered on demand.
• Because a user MCP server still pending at the start of a turn is not in the deferred list, the harness system prompt now carries one extra line naming the user's configured servers and telling the model to use ToolSearch for them (sorted and deduplicated so the prompt is byte-stable for caching, names over 64 chars dropped, capped at 20, no line at all when there are no servers).
• Adds per-turn context telemetry (below) so the baseline size, missing deferral, not-ready servers and any compaction are visible in Axiom, plus Sherlock knowledge entry #47.

Test plan

• Gates: check-brand-assets, oxlint --deny-warnings, pnpm deps:check (no violations), pnpm run types (apps/server) all clean. Not run: full pnpm run autofix:local (it ends in an unscoped vitest run src/ and the whole mail Jest suite, which this repo's hygiene rules forbid).
• Scoped vitest: src/mastra/workflows/chat (14 files, 169 tests) passes after merging staging. New/changed tests: context-telemetry.test.ts (summarizers and the observe closure incl. never-throws), delegate-tool-scoping.test.ts (buildAllowedTools includes ToolSearch; the connected-servers line: sorted, deduplicated, byte-identical across input orders, cap at 20, long/reserved names dropped, single line, unchanged prompt for no servers), user-mcp-servers.test.ts (alwaysLoad absent on every server config).
• Battle-tested live on the real code (local chat-service, real model, persistence off, read-only prompts), A/B against the current config:
  • /context baseline for a Notion-connected user: 193.9k (97%) -> 23.5k (12%); a no-external-MCP control 131k -> 25k.
  • A turn shaped like the failing prod one (find-crm-conversations -> fetch-conversation -> Read -> Grep): current config reproduced the exact Autocompact is thrashing error (3 compactions, 190s); with this change it completes, 0 compactions, first request ~42k tokens (re-verified on the final commit).
  • On a natural prompt ("check my Notion for the most recently edited page") the model goes straight to ToolSearch then the Notion tool; a server reported pending at init was still discovered in the same turn and was listed as configured from the prompt line.
• Thermo review: 3 fresh-agent rounds on the main change plus a fresh review of the prompt-line commit; every finding fixed or explicitly justified.

Verify in prod

Log lines from run-chat-agent-sdk.ts (emitted once per Claude-mode chat turn):

• Axiom dataset: cedar-prod-logs (also cedar-staging-logs) , query:

  ['cedar-prod-logs'] | where _time > ago(1h) and message == "[run-chat-agent-sdk] first request context"
  | summarize p50=percentile(todouble(context_tokens), 50), p95=percentile(todouble(context_tokens), 95), max=max(todouble(context_tokens)) by bin(_time, 10m)
  

  Other lines: [run-chat-agent-sdk] tool baseline (tool_search_available, available_tool_count, available_mcp_tool_count, mcp_servers name+status; counts include deferred tools), [run-chat-agent-sdk] context compacted (trigger, pre_tokens), warns [run-chat-agent-sdk] tool search unavailable and [run-chat-agent-sdk] user MCP server not ready at init (servers name+status), [run-chat-agent-sdk] context telemetry failed.
• CloudWatch: `/aws/ecs/aws-pr