What

Raises the chat agent's tool-loop step cap from 20 → 100 on the four surfaces the user interacts with directly:

Background/automation paths (orchestrator dispatch, operation registry, execute-orchestrator, etc.) are intentionally left at their existing caps , this only touches direct-user chat.

Why

Investigating why the Slack bot sometimes replies "I did not generate a response." The message is a fallback that fires only when the agent finishes with empty text. Prod telemetry (last 14 days):

• 7 of 153 Slack replies went out empty ({"text":""}, no buttons, no thread replies).
• Every one traced to a Chat Agent (Understudy) generate() that ended on finish_reason: "tool-calls" , i.e. it hit maxSteps: 20 while still mid tool-loop (8k,24k output tokens).

The prompts force multiple silent lookups (find-crm-conversations → get-conversation → write) before any output and only ever emit text inside a final <final_answer> block, so a run that exhausts the step budget has no assistant text anywhere → extractFinalAgentText returns "" → the fallback string. Raising the ceiling gives these runs room to reach their final answer.

The frontend never hit the cap in the measured window (0/238 exhaustions; it also streams, so it wouldn't blank out) , but it's raised too for consistency across direct-user surfaces.

Follow-ups (not in this PR)

This raises the ceiling but does not remove the failure mode , a run that still exhausts 100 steps hits the same blank-message fallback on the headless generate() path. Recommended next steps:

1. Forced-final-answer retry: when a run ends on finish_reason: "tool-calls" / empty text, do one more "wrap up now" generate step instead of giving up.
2. Graceful empty-text fallback + fix the latent responseText = structured?.text ?? rawText (?? preserves "") so a threadReplies-only response isn't blanked.
3. Cost/latency watch: a genuinely stuck loop can now run ~5× longer / ~5× tokens before terminating; the retry above also caps that downside.

🤖 Generated with Claude Code

Greptile Summary

Raises the Mastra agent tool-loop limit from 20 to 100 for the four direct-user chat surfaces so multi-step requests have more opportunity to produce a final response.

• Updates headless Slack, SMS, and iMessage agent generation limits.
• Updates the legacy in-app streaming chat path to the same limit.
• Leaves background and automation agent limits unchanged.

Confidence Score: 5/5

The PR appears safe to merge, with the explicitly acknowledged increase in worst-case agent cost and latency remaining the primary operational tradeoff.

The changes only raise existing step limits on direct-user chat invocations, and the review found no concrete new timeout, retry, delivery, security, or correctness failure.

Important Files Changed

_{Reviews (1): Last reviewed commit: ["fix(chat): raise maxSteps 20→100 on user..."](https://github.com/cedarcopilot/cedar}