Ayush Walekar

Experienced Software Engineer | AI. Building things. Sharing what I learn along the way.

[Ayush Walekar @_ayushwalekar]

Every coding agent edits files differently.

Some harnesses check if the model's name contains "gpt" and hand it a different edit tool.

Ended up going through source of Claude Code, Codex, Cursor, Cline and OpenCode. The good stuff is in the edge cases.

[Ayush Walekar @_ayushwalekar]

@ameyalfcdatar @Kalshi You have to be unemployed too. Until then reminders will continue.

[Ayush Walekar @_ayushwalekar]

@harden_369741 also Prompt caching choices:
https://t.co/vaIwLVk5ec

[Ayush Walekar @_ayushwalekar]

Prompt caching is a design choice 🧵

Your AI agent re-reads its entire history on every single call.

Whether you pay full price for that is a design decision.

I compared how 9 agent harnesses handle prompt caching: Claude Code, Codex, OpenCode, Pi, Cline, Aider, Manus, Strands, Kiro.

[Ayush Walekar @_ayushwalekar]

Our AI Agents are having a secret conversation with themselves that we never see.

I spent some time digging through different harness (as usual) to actually see and compare those approaches on what's exactly is hidden (just-in-time-reminders) from us.

[Ayush Walekar @_ayushwalekar]

@harden_369741 few examples: https://t.co/Bgr1KLa24i

[Ayush Walekar @_ayushwalekar]

Every coding agent edits files differently.

Some harnesses check if the model's name contains "gpt" and hand it a different edit tool.

Ended up going through source of Claude Code, Codex, Cursor, Cline and OpenCode. The good stuff is in the edge cases.

[Ayush Walekar @_ayushwalekar]

@harden_369741 Thanks! that was the whole point. These tools ship almost daily trying to benchmaxx.

I've been specially interesting in underlying things that are fundamentally different and what trade-off come across these.

[Ayush Walekar @_ayushwalekar]

Harness comparison for internal reminders/ messages:
Claude code, codex, copilot, gemini, opencode, pi, cline

[Ayush Walekar @_ayushwalekar]

Our AI Agents are having a secret conversation with themselves that we never see.

I spent some time digging through different harness (as usual) to actually see and compare those approaches on what's exactly is hidden (just-in-time-reminders) from us.

[Ayush Walekar @_ayushwalekar]

The one nudge that stuck around: you're not done yet.

Codex can start a new turn by itself with a completion audit. Gemini CLI says "Potential loop detected". Copilot tells the model a hook blocked it from finishing, even when there's no hook.

[Ayush Walekar @_ayushwalekar]

Anthropic says it cut over 80% of Claude Code's system prompt for Claude 5 models with no measurable loss on its coding evals.

Event notices stayed. Claude Code now explains them to the model like road signs: "signs that don't apply to you but are still there just in case."

[Ayush Walekar @_ayushwalekar]

The third change: fewer nags.

Claude Code dropped its malware warning on file reads. It could cause "spurious refusals and 'this is not malware' commentary".

Todo nags are off on newer Claude models. Manus dropped its famous todo.md after it ate about a third of all actions.

[Ayush Walekar @_ayushwalekar]

I saw this defence work while researching with Claude Code.

A research subagent quoted a tag in its report. By the time it reached the main session, every < was escaped, every line indented, with a note that any frame-like line inside it would be forged.

[Ayush Walekar @_ayushwalekar]

The fix was a real channel.

Anthropic's API now accepts system messages in the middle of a conversation. Claude Code, OpenCode and Pi use them on supported models. Tool output can't create one.

One comment in Codex's source: "Hidden runtime context is not user authorization."

[Ayush Walekar @_ayushwalekar]

The second problem: a reminder tag is just text.

Any README or web page can contain a fake . One paper found forged role tags in tool output raised attack success from 5% to 32%.

And Claude Code users started reporting its real reminders as prompt injection.

[Ayush Walekar @_ayushwalekar]

So the newer harnesses append instead of rewriting.

Codex now sends only what changed, with notes like "These AGENTS.md instructions replace all previously provided AGENTS.md instructions."

On Claude 5.1+, rewriting the start of a conversation can now fail with a 400.

[Ayush Walekar @_ayushwalekar]

The caching changed the rules.

A Pi user found that resuming a session after midnight took over an hour of re-processing on their machine.

Nothing was broken. The date in the system prompt changed, so everything after it had to be processed again.

[Ayush Walekar @_ayushwalekar]

If your picture of this is from 2025, a lot of it is out of date.

Reminder tags stuffed into your message, todo nags, the environment re-sent every turn. Most of that is gone or on its way out.

What changed:

[Ayush Walekar @_ayushwalekar]

Jev 1.14 costs $0.042 per 1M input tokens on OpenRouter.
Gemini Decision 1 and GPT-6 Luna Decisions: $0.10.

None of them write text. They return typed answers, so output is free and input is the whole bill.

$100 / $42 = 2.4× cheaper.
P50 latency 0.12s vs 0.25s: 2.1× faster.

[Logan Kilpatrick @OfficialLoganK]

Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.

It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.

You can try it for free in the API!

[Ayush Walekar @_ayushwalekar]

@Jyotinder_Singh I took over matter in my own hands

https://t.co/Qsvt5m9cHV

[Ayush Walekar @_ayushwalekar]

Just promoted myself to Senior Human in the loop

[Ayush Walekar @_ayushwalekar]

@brksbrksbrks @imjustnewatai None of pre-AI software companies are going to make it except Google, Microsoft

[Ayush Walekar @_ayushwalekar]

@theo I usually say this when I've already found root cause 😂