T3 Code is significantly slower than Codex for the same coding task (20+ min vs ~3m24s)

Summary

T3 Code appears to be significantly slower than Codex for the same repository investigation task, using the same model configuration.

With the prompt below, Codex completed the task in 4m35s, while T3 Code took 15+ minutes.

This report is about end-to-end task completion time, not answer quality.

Benchmark prompt

Inspect this repository and answer these questions without editing anything:

1. Where is authentication implemented?
2. Which files handle login/session/token validation?
3. What middleware, guard, or server-side check protects authenticated routes?
4. Where are authorization rules enforced?
5. If a user loses access to a team/organization/project, what code path would prevent them from seeing protected data?

Requirements:
- Read-only investigation only
- Do not edit files
- Do not propose a refactor unless necessary
- Cite the exact files and relevant functions/classes
- End with a concise summary of the authentication flow

But I also tried several other prompts, and this does not seem to be specific to this exact prompt.

Settings

• GPT 5.4
• High
• Fast: On

Results

• Codex: 4m35s
• T3 Code: 15+ minutes

Expected behavior

For the same prompt and comparable visible model settings, T3 Code should be much closer to Codex in total completion time, or at least not several times slower.

Notes

This task does not require editing, building, or running the project. It mainly involves repository search, file reading, code-path tracing, and synthesizing a final answer.

Because of that, the slowdown seems more likely related to orchestration/tooling/app overhead than to the model itself.

Environment

• T3 Code version: 0.0.9 (4b6a0c433cb8)
• Hardware: Apple M1 Pro, 16 GB RAM
• macOS: Tahoe 26.2

If useful, I can provide exact timestamps, a screen recording, repeated comparisons, or rerun the benchmark with diagnostics enabled.

Reproduce the benchmark in T3 Code 0.0.9 on the stated macOS and hardware, comparing it with Codex using the supplied read-only repository-investigation prompt. Collect diagnostics, exact timestamps, and repeated runs to isolate orchestration, tooling, or app overhead; done means identifying the bottleneck and substantially reducing end-to-end completion time.