Issue #3
Agent Ops Brief — Issue #3
Theme: Nested rules finally bind. One bad MCP schema shouldn't blank the server.
Scheduled: Wed Oct 28, 2026 (archive preview) (America/New_York) — web archive (weekday email still held until distribution clear)
Stack coverage: Cursor · Claude Code · MCP · agent CI
Five operator signals for Wednesday. Free readers get the brief. Operators get the paste-ready drop at the bottom.
Issue #2 covered allowlist≠denylist and Grep leaving the workspace. This issue is the next seam: package-local rules that finally apply mid-edit, and MCP servers that fail one tool without going silent.
1. Cursor CLI 2026.10.22: nested AGENTS.md finally applies when the agent edits a file
What happened. Cursor CLI v2026.10.22 (Oct 22) fixed a rules gap operators kept hitting: when the agent edits or creates a file in a subdirectory, it now sees that directory's AGENTS.md and nested always-apply rules — even if it never read a file there first. Each rule is shown once per conversation; glob-scoped and root rules behave as before.
Why operators care. Monorepo conventions under packages/foo/AGENTS.md were easy to ignore mid-edit. Package owners who thought "local rules = local enforcement" were wrong until this build.
Action. Put package-local conventions in nested AGENTS.md (or .cursor/rules with globs). Upgrade CLI: agent update. Spot-check by having the agent edit a file under a package that has its own AGENTS.md and confirm the rule appears.
Sources: Cursor CLI changelog — 2026.10.22
2. Same release: one invalid MCP tool schema no longer hides the rest of the server
What happened. In 2026.10.22, a single invalid tool (or a tool schema without a type) no longer hides the rest of an MCP server's tools. A server that fails to list its tools now reports an error to the agent instead of failing closed into "no tools."
Why operators care. One bad tool definition used to look like the whole connector died. Operators wasted hours "debugging auth" when the real issue was a schema typo.
Action. After agent update, reopen /mcp (or agent mcp list) on any server that previously showed empty. Fix or remove the bad tool; keep the rest.
Sources: Cursor CLI changelog — 2026.10.22
3. Auto-review of MCP calls can finally see JSON-string input schemas
What happened. Auto-review of MCP tool calls now sees the tool's input schema when the server sends it as a JSON string (same 2026.10.22 CLI notes).
Why operators care. Classifier steering in permissions.json is only as good as the schema the reviewer can see. String-encoded schemas used to make MCP approvals a coin flip.
Action. If you steer MCP with Auto-review block_instructions / allow_instructions, re-test one send/publish tool after upgrade. Prefer servers that expose real JSON Schema objects.
Sources: Cursor CLI changelog — 2026.10.22
4. Custom subagents no longer hard-fail when their pinned model is blocked
What happened. When a custom subagent is pinned to a model your team's admin settings block, that requires Max Mode, or isn't available, it now runs on the parent agent's model instead of erroring out. The result notes which model it ran on.
Why operators care. One admin model-policy change used to brick every Task/subagent overnight. Soft fallback keeps fleets moving; you still need to notice the pin was skipped.
Action. Search agent results for "ran on" / model-skip notes after a policy change. Pin only models your team actually allows, or inherit parent deliberately.
Sources: Cursor CLI changelog — 2026.10.22
5. Claude Code 2.1.304: /auto-mode-setup teaches auto mode from your project
What happened. Claude Code 2.1.304 (Oct 19) added /auto-mode-setup: a guided setup that reads your project and recent sessions to teach auto mode about your environment (also available on Bedrock, Vertex, Foundry, Claude apps gateway, and telemetry-disabled sessions). Same build added mod tool.check answer decider: "person" so only a human (or host callback) can approve — refused where nobody can be asked.
Why operators care. Auto mode without a project-taught baseline is vibes. /auto-mode-setup is the operator onboarding path; decider: "person" is the fail-closed gate for headless.
Action. Run /auto-mode-setup on one canary repo. For mods that must never auto-approve in CI, set decider: "person".
Sources: Claude Code changelog — 2.1.304
Operator · Config Drop
Paste-ready file — founding unlock
This week's file: Nested AGENTS.md monorepo pack + MCP schema smoke check. Free readers get the signals above. Operators get the full paste-ready drop every issue.
# packages/api
- Prefer the shared `lib/db` client; do not open ad-hoc Postgres pools.
- Migrations live under `packages/api/migrations` only.
- Never edit `packages/web` from an api-scoped task — stop and ask.
# scripts/mcp-schema-smoke.sh
# fail if a configured server lists zero tools after connect
set -euo pipefail
echo "Implement against your mcp.json inventory — empty tools[] after connect = fail."
…Full file locked — founding Operators unlock every weekday Config Drop.
Already paid? Enter the email from your Stripe receipt to unlock on this browser. New founding Operators — $9/mo for the first 50, then $15/mo. Cancel anytime in Stripe.
Same email as Stripe checkout. No magic link yet — cookie unlock on this browser. Fallback: email receipt to agentopsbrief@gmail.com.
Issue #3 — Agent Ops Brief. Free weekday signals; founding Operator unlocks the week's paste-ready configs (receipt→unlock until webhook ships).