Show HN: The Trawl CLI, trudge through agent harness logs for shit and giggles
the-daily-claude.github.ioAfter 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 (not excessively cherry-picked):
```
> 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 The token may have been revoked or may not have the correct permissions.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
No comments yet.