Regression: low-quality Agent output, heavy token usage, and Misconfigs tab overload on every interaction

[ramzimalhas]

Summary

Regression affecting Agent interactions: output quality is consistently low, token usage appears abnormally high, and the Misconfigs tab appears overloaded, creating a critical workflow blocker.

Problem

The user reports that this behavior happens every time across terminal/agent/UI/settings surfaces. The combined effect is degraded response usefulness, heavy token consumption, and persistent configuration-noise overload in Misconfigs.

Expected behavior

Agent responses remain high quality and concise for the same prompts, token usage stays within normal bounds, and the Misconfigs surface remains manageable and signal-focused.

Actual behavior

On every interaction, the user observes low-quality outputs, heavy token usage, and Misconfigs-tab overload. This is reported as a critical blocker.

Reproduction steps or desired workflow

1. Open Warp and use normal daily workflows involving Agent/Oz, terminal interactions, and settings review.
2. Submit typical prompts/requests.
3. Observe output quality, token usage patterns, and Misconfigs-tab state.
4. Compare against prior baseline behavior where output quality and config signal were acceptable.

Artifacts

None attached.

Warp version

v0.2026.04.27.15.32.stable_03

Operating system

macOS 26.4.1 (build 25E253)

[oz-for-oss]

@ramzimalhas

Here's what I found while triaging this issue:

This report is missing the artifacts needed to distinguish an Agent backend/model issue from a client UI/settings or account-specific configuration issue. A representative debugging ID and visual evidence of the Misconfigs overload would materially improve triage confidence.

I have a few follow-up questions before I can narrow this down:

1. Can you attach a screenshot or short screen recording of the overloaded Misconfigs tab during or immediately after a failing Agent interaction, with sensitive details redacted?
2. Can you share the AI debugging ID for one representative interaction that showed both low-quality output and unexpectedly high token usage?

Reply in-thread with those details and the triage workflow will automatically re-evaluate the issue and update the diagnosis, labels, and next steps.

Maintainer details

I concluded that regression report for Agent response quality, abnormal token usage, and a noisy Misconfigs/settings surface, but it lacks artifacts needed to isolate client, account/backend, or configuration-specific causes..

Triage summary

This looks like a bug report for a possible Agent regression on macOS Stable involving three related symptoms: lower-quality Agent responses, higher-than-expected token/credit usage, and an overloaded Misconfigs/settings surface. The report is not yet actionable enough to isolate a specific root cause because it does not include a representative conversation/debugging ID, visual evidence of the Misconfigs overload, logs, or exact prompts/usage deltas.

Reproducibility

repro:unknown — the report says the behavior occurs on every interaction, but the reproduction path is broad (normal daily workflows) and lacks concrete artifacts that maintainers can replay or compare against a known-good baseline.

Likely area

• area:agent for Agent/Oz response quality and token usage behavior.
• area:settings-keybindings for the reported Misconfigs/settings surface overload.
• os:mac because the report is for macOS 26.4.1 on Stable v0.2026.04.27.15.32.stable_03.

Code/context notes

The inspected client code primarily displays and persists server-provided usage metadata rather than computing Agent token usage locally:

• app/src/ai/blocklist/controller.rs consumes stream-finished usage metadata and request cost.
• app/src/ai/agent/conversation.rs accumulates request cost, token usage, context-window usage, and tool usage metadata for a conversation.
• app/src/ai/blocklist/usage/conversation_usage_view.rs renders per-conversation usage details.
• crates/graphql/src/api/queries/get_conversation_usage.rs and app/src/settings_view/billing_and_usage/usage_history_model.rs fetch and convert server-side usage history.
• app/src/settings_view/ai_page.rs contains model/profile/context-window controls that can affect usage, but the current issue does not identify a specific setting change.

A repository search did not find a literal Misconfigs tab/surface in the public checkout, so that part may be behind a different name, generated/internal UI, or account/backend state.

Duplicate check

I enumerated all open issues with the GitHub API and found related reports about Agent token consumption or answer quality, but not two open issues matching this report's combined symptoms, especially the Misconfigs overload. Leaving duplicate_of empty for now.

Question reasoning

1. Can you attach a screenshot or short screen recording of the overloaded Misconfigs tab during or immediately after a failing Agent interaction, with sensitive details redacted? — The report includes a visual/UI symptom but no media, and the public checkout does not expose a literal Misconfigs surface to inspect directly.
2. Can you share the AI debugging ID for one representative interaction that showed both low-quality output and unexpectedly high token usage? — A debugging ID is the most direct way for maintainers to correlate the reported quality and usage regression with backend conversation metadata without asking for private account details.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz