[Bug]: Model reasoning in loop burning all tokens

[pat3dx]

Description

There is a big issue on how the model, whatever the model, is reasoning. For a simple blog post, it took 11 mn .........and cost me $2 for a simple blog post using the new Qwen 3.6 Flash, witch cost nothing. this is just totally crazy.
The AI make the blog post in 30s, then try to fix issue in loop, and never find the issue.
I tried different model Mimo v2.5, Mimo v2.5 Pro, Qwen 3.6 flash, all same issue....agent try to fix issue for 10+ mn for a simple task, and burn a lot of credits for nothing.
You really need to make the agent smarter.
Also, i noticed, in the AI reasoning in the chat, a lot of time it try to read the error, but AI said, it don't know what is the issue, because the lag parse nothing.
So the AI, try to fin by itself and test everything.

The biggest issue, is that AI can't reaf the log because the log is empty.
Need an urgent fix, i had to stop testing the app, because it whould have spend me like $10 to genrerate a simple blog article LOL...

Steps to reproduce

step to reproduce: everything, genrate blog post, landin,g page, everything.

Expected behavior

for a blog post, the AI should genrate it in less than 1mn, not 10mn with a cost of $2+.....and even after 10mn of reseasoning, i had to stop it.....it can't even find and fix the issue in the blog post.

Actual behavior

Agent is stuck in a loop of trying to find and read error log, but the log is empty, so Agent don't know what issue there is, and try everything by itself, so it enter in an infinie loop of reasoning - editing - reasoning - editing.....and burn all credits.

open-codesign version

0.1.4

Platform

Windows

OS version

win 11

Model provider (if relevant)

OpenRouter

Error code

no error code

Diagnostics bundle

No response

Logs

[github-actions]

Thanks for reporting this. I checked the current 0.1.4 code path, and there is evidence that the kind of long self-check/edit loop you describe can be reinforced by the app's current agent/verifier flow, not only by the upstream model.

In 0.1.4, desktop generation goes through the multi-turn agent path (apps/desktop/src/main/index.ts:575), and artifact verification is driven by the done tool (packages/core/src/agent.ts:576, packages/core/src/tools/done.ts:281). In the generateViaAgent() path I checked, I did not find a code-level max-round guard on that loop (packages/core/src/agent.ts:845). I also found two different stop instructions in the current prompt/tool text: "Stop after 3 rounds" in packages/core/src/agent.ts:581 and "Stop calling once status is "ok" or after 5 rounds" in packages/core/src/tools/done.ts:287.

I also verified that the verifier is narrower than a full log reader. It only combines static artifact checks (packages/core/src/tools/done.ts:304) with console warnings/errors and load failures collected from a hidden BrowserWindow during a short runtime check (apps/desktop/src/main/done-verify.ts:21, apps/desktop/src/main/done-verify.ts:66, apps/desktop/src/main/done-verify.ts:113). So if the bad state does not surface as a first-paint console warning/error or load failure, the agent can end up with little or no actionable error data. The current agent prompt explicitly documents that limitation for non-initial UI states (packages/core/src/agent.ts:583).

I searched related history as well. #199 is also about overly long generations, but that report is about transport timeouts against local endpoints, not this verify/edit loop. #216 is about connection-test/runtime parity, so it looks related to diagnostics quality, but not like a duplicate of this report.

Relevant code:

• apps/desktop/src/main/index.ts:575 — generation is routed through the agentic multi-turn path
• packages/core/src/agent.ts:581 — prompt text says to stop done retries after 3 rounds
• packages/core/src/tools/done.ts:287 — tool description says to stop after 5 rounds
• packages/core/src/agent.ts:845 — agent construction path I checked does not show a code-level max-round cap
• apps/desktop/src/main/done-verify.ts:21 — runtime verifier timeout window starts at 3s
• apps/desktop/src/main/done-verify.ts:66 — verifier collects console warning/error events
• apps/desktop/src/main/done-verify.ts:113 — verifier resolves after load/settle or timeout
• apps/desktop/src/main/logger.ts:13 — Windows log location is %APPDATA%/open-codesign/logs/main.log

Need more info:

• Please attach a Diagnostics bundle from Settings → Diagnostics for one affected run.
• Please include the exact OpenRouter model slug that was selected when the loop happened.
• If possible, paste the agent.tool_start, agent.tool_end, and agent.turn_end lines from main.log around one bad run.

---

open-codesign Bot