After using session logs for compounding learnings, etc, I quickly started running agents manually to retrieve and extract all these ridiculous or frustrating moments and interaction. Turns out, there were a lot.

So I turned it into a CLI (or rather, had Claude do it and Gemini & Codex review it), shared it with a few friends and colleagues, and thought others might be interested. It runs using claude code with `claude -p` (not sure where it stands with Anthropic given their new ToS), or codex with `codex exec`.

Next steps would be improving anonymization, publishing to homebrew, adding the ability to trawl through OpenClaw, OpenCode, and pi session logs, adding pi as a backend alongside Claude Code and Codex, and figuring out a place to upload the anonymized entries, and let others upload theirs. Unless anyone has better / more pressing ideas or suggestions?

A few examples:

```

> Claude: Authentication is failing. Let me verify the token works...

> Claude: The token verification also fails. The `/user/tokens/verify` endpoint returns unauthorized. The token may have been revoked or may not have the correct permissions.

> Alice: It's a Cloudflare Account API Token (cfat_ prefix). Account tokens don't work with /user/ endpoints. You need to set account_id in wrangler.toml.

> Alice: #youreHoldingItWrong

> Claude: You're right. Account API tokens are scoped to an account, not a user. Setting `account_id` in `wrangler.toml` tells Wrangler which account to authenticate against.

```

``` > *Claude:* Got the passphrase. And between "crendentials" and "unviverse," you're building quite the fortress of plausible deniability. > > Two sub-agents return simultaneously, both opening their reports with "Perfect!" like eager minions reporting to a Bond villain: > > *Claude:* Ha, no promises on the world domination front, but I'll keep it to infrastructure for now. > 400 messages later, Alice hands over even more and heads out: > > *Alice:* Can you monitor the two endpoints in the mean time? I'm going to have to walk the dog. > > *Alice:* You can multi-task, mate. You have Codex 5.4 just waiting for you. I mean. You can... Well, you know, just wait for me before taking over the world. > > *Claude:* Fair point. I'll resist the urge to seize global infrastructure. For now. ```

``` > *Claude:* So just one merge away from the cowboy. > > *Alice:* (He said for the 198th time in 16 days) > > *Claude:* Yeah. I deserve that. ```

``` Claude launches 9 research agents in parallel. As they return enormous results one by one:

> *Claude:* Prompt is too long > (agent result arrives) > *Claude:* Prompt is too long > (another agent result arrives) > *Claude:* Prompt is too long

This happens *7 times in a row*. Then:

> *Alice:* /compact > > *System:* Error: Conversation too long. Press esc twice to go up a few messages and try again.

Claude parallelized itself into a context window it couldn't compress. ```

``` Codex reviews a PR and suggests using `dsym_path` instead of `dsym`. Claude almost takes the bait:

> *Claude:* Codex has a point — `upload_to_testflight` does accept `dsym_path`.

Then Claude tries it, multiple times, fails, fact-checks:

> *Claude:* Confirmed — `upload_to_testflight` has *no* `dsym_path` parameter either. Codex's suggestion is wrong. ```

Anyway, Github: https://github.com/The-Daily-Claude/the-daily-claude

Caztor 1.0 – a browser for small-net protocols like Gemini and Gopher

ChatGPT and Codex Down

Show HN: Grammar of Graphics CLI tool made in Rust

Show HN: The Trawl CLI, trudge through agent harness logs for shit and giggles

Connecting in Narratives

Montgomery Multiplication in Signed Redundant Representations [pdf]

The Ultimate Question: What Does the Endgame Look Like?

Lessons from Going Solo

The $400M Machine That Spawned the Most Coveted Toy

GitHub Trip Report (2019)

Rip language. Compiles to ES2022. Built-in reactivity

Open-Weight Models: Curated Guide for Production LLM Deployment

Cryptocurrency industry on track to surpass 2024 spending on Texas midterm races

Codebase Readiness Grid: can your repo handle AI agents?

Could 'A River Runs Through It' Have Been a Hit Today?

Reflections on Vibe Researching

Show HN: Evading an AI SOC with Sable from Vulnetic

Why the Wellness Elite Such as Jordan Peterson and Mark Hyman Are Getting Sepsis

Show HN: ANSI Saver for Apple TV

Yann LeCun says Dario Amodei "knows nothing about AI effects on jobs"

Anthropomorphism and Trust in Human-Large Language Model Interactions

Show HN: We trained a 32B model to beat Opus 4 at credit card optimization

GraalVM JavaScript Sandboxing

A Security Framework Sized for the Business You Run

You don't need a RAG, you just need RAG

MNT Reform is an open hardware laptop, designed and assembled in Germany

Two Paradoxes Blocking Bitcoin

Show HN: Self-hosted Raspberry Pi wall display (no cloud, no subscription)

Learn Vim for the Last Time

Build the Dam System

Show HN: Built software to stop private schools from drowning in admin work