What was wrong

Jacob Doty (Aspire) reported his "Zeke the Sequencer" outbound agent was still using em dashes and an inconsistent email signature despite already having explicit "no em dashes" and "sign off as Jacob" rules in its subagent doc.

Root cause traced via Sherlock (Postgres chat_messages for the Slack thread slack:T02511TSJ:C0C0XC2FAR4:1789410187.953389, 2026-09-14):

1. Jacob asked the chat agent to build a sequence. The agent correctly called spawn-subagent(type="orchestrator", agentId="5a688ce6-c542-44ba-8a27-581317d3122a") , Zeke's real subagent.
2. That call 400'd against the Understudy gateway:

   "x-understudy-tags key 'userId' must match /^[a-z0-9][a-z0-9-]{0,63}$/"
   

   orchestrator-agent.ts passes the raw Cedar userId straight into x-understudy-tags (WHJSsQgW9nP9xKHmjPkZtYgPwSbow2US , mixed case, no hyphens, per Better Auth's id format), which violates Understudy's required tag pattern. This breaks the call for every Cedar user, not just Jacob, whenever the cedar-understudy-orchestrator flag routes a spawn-subagent(type="orchestrator") call through the gateway.
3. With Zeke's real invocation dead, the chat agent silently fell back to a bare run-code-executor call with a freehand task description , which doesn't inherit any parent subagent's prompt, so none of Zeke's style rules (no em dash, sign off as Jacob) were ever in scope.
4. The resulting document had 55 em dashes and an inconsistent signature , not because the config was ignored, but because Zeke never ran.

What changed

understudyHeaders() (apps/server/src/services/llm/understudy.ts) now sanitizes every x-understudy-tags value , lowercases it, strips characters outside [a-z0-9-], strips leading hyphens, and truncates to 64 chars , before building the header. This is centralized so no future caller needs to separately learn Understudy's tag format.

Added test coverage in resolve-model.test.ts covering the exact mixed-case userId that broke this, plus an arbitrary tag with disallowed characters.

How to test

• pnpm --filter @zero/server exec vitest run src/services/llm/__tests__/resolve-model.test.ts , 12/12 passing, including the two new sanitization cases.
• pnpm --filter @zero/server run types , clean.
• In prod: watch for spawn-subagent(type="orchestrator") calls in agent_tool_calls for users with cedar-understudy-orchestrator enabled , they should no longer show a TOOL_EXECUTION_FAILED / x-understudy-tags 400 in the result, and should not be followed by a run-code-executor fallback for the same task.

Separately (not in this PR , a live document edit, not code): Jacob's two sequencer subagent docs (zeke-the-sequencer.md, event-outreach-sequencer.md) were patched via writeDocumentGuarded to change the sign-off rule from "Jacob" (no dash) to "-Jacob" per his request, and verified via the playbook health + Yjs sync checks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UPnwxwkTP5oh7k6MhZn97b

Greptile Summary

This PR centralizes normalization of Understudy tag values so Cedar identifiers satisfy the gateway’s lowercase, character, leading-character, and length constraints.

• Sanitizes every tag value before serializing x-understudy-tags.
• Preserves environment tagging while preventing invalid values from rejecting requests.
• Adds regression coverage for the reported mixed-case user ID and disallowed punctuation.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete correctness, security, or repository-rule violations identified.

The sanitizer handles the documented character, first-character, empty-value, and maximum-length requirements, and the regression tests exercise the production failure that motivated the change.

Important Files Changed