Carl Vitullo
@vcarl.com
865 followers 888 following 2.2K posts
Solopreneur just vibing, posts are probably bullshit. Community lead at Reactiflux (reactiflux.com), the largest chat community of React professionals. Check out This Month in React (open.spotify.com/show/4g3Le83YfsMeI…).
![A screenshot of Sublime Merge, a Git client. The commit description reads:
Review brief alignment
--- metrics ------------------------------------------------------------
Tokens-Generated 588.2k
Tool-Calls 515
Files-Read 30
Log-Lines-Added 8826
Log-Lines-Removed 984
Log-Unattributable-Ops 0
Git-Lines-Added 10497
Git-Lines-Removed 893
Git-New-Files 15
Coverage-Unobserved-Files 4
Coverage-Shell-Mutations 187
Coverage-Churn-Ratio 0.86
Log-Churn 0.11
Log-Novel 0.89
Log-Deletion 0.00
Active 1h 36m
Wall-Span 1h 36m
Rewinds 1
Interrupts 0
Compactions 0
Redactions 5
--- timeline -----------------------------------------------------------
[2026-08-21T00:07:43.575Z]
We've just backed up after a fair bit of work, with the intention of redoing the work as a one-shot project. Please review the following review brief and clarify anything necessary. We'll verify that intentions have been correctly aligned, then you'll ship through until we're able to generate a commit description as described that capture the work of building it.
The brief:](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreig2bvycgbixivwds7xodmbwnyphxvsjosjnrav572a3gntomac4my)
![08cbb27 Add ledger capture and timeline rendering design spec
[sha-scan]
docs/superpowers/specs/2026-08-20-ledger-and-timeline-design.md (+357 -0)
→ general-purpose sonnet 12.5k tokens · 11 calls · 4 files · +231 -0 lines · 2m 21s
:: Build humanize module
→ general-purpose opus 41.5k tokens · 32 calls · 3 files · +197 -12 lines · 9m 30s
:: Build capture hook scripts
→ general-purpose sonnet 45.3k tokens · 41 calls · 4 files · +537 -8 lines · 8m 35s
:: Build git plumbing additions
e6d7c60 Read the capture ledger tolerantly [sha-scan]
src/ledger.ts (bin) src/ledger_test.ts (+328 -0)
→ general-purpose sonnet 19.9k tokens · 17 calls · 3 files · +536 -0 lines · 3m 28s
:: Build ledger types and reader
cb25564 Humanize counts and durations for the metrics block
[sha-scan]
src/humanize.ts src/humanize_test.ts (+229 -0)
26072c9 Add git plumbing for ledger capture and attribution
[sha-scan]
src/git.ts src/git_test.ts (+529 -0)
23b3405 Capture commits and rewrites with two dumb POSIX hooks
[sha-scan]
src/hook.ts src/hook_test.ts (+642 -2)
→ general-purpose opus 31k tokens · 27 calls · 5 files · +437 -76 lines · 7m 1s
:: Install all three hooks in init
→ general-purpose opus 42.5k tokens · 20 calls · 4 files · +1468 -23 lines · 8m 5s
:: Build attribution resolver](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreia4slwaibf5gx3l5hpm6ap4tqcbzd56fwijtvgvnesk573sljgbia)
![[2026-08-16T20:07:07.672Z]
Show me more of the metadata you'd put in the commit description
[2026-08-16T20:09:30.837Z]
These fields of the trailer should be in the metrics, as human readable figures:…
» subagent Explore (sonnet) :: Map src module API surface
tokens 15.1k · tool calls 24 · files read 18 · 2m 02s
» subagent general-purpose (haiku) :: Check subagent session env var
tokens 375 · tool calls 1 · 6s
» subagent general-purpose (haiku) :: Implement Task 1: humanizer
tokens 4.7k · tool calls 16 · files read 1 · log lines +168 -1 · 2m 04s
• commit 870c193 Add count and duration humanizer [session]
src/humanize.ts (+23 -0) · src/humanize_test.ts (+39 -0)
» subagent general-purpose (sonnet) :: Review Task 1 (spec+quality)
tokens 5.1k · tool calls 3 · files read 3 · 54s](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreidpl2o5vwb2xz6htpphfzmclfntz4wzwhq4xi22eguudwkhypewhu)
![A screenshot of Sublime Merge, a Git GUI. The body of the commit reads:
Build Deno CLI for AI session commit generation
metrics
Tokens generated 907588
Tool calls 872
Files read 51
Lines added 4776
Lines removed 289
Unattributable ops 0
Churn 0.06
Novel 0.94
Deletion 0.00
subagents
- general-purpose sonnet d1 tok=18597 calls=37 ms=233283 :: Harvest log-format facts from Python
- general-purpose sonnet d1 tok=49779 calls=71 ms=709120 :: Survey real Claude Code JSONL logs
- general-purpose sonnet d1 tok=45380 calls=82 ms=719494 :: Verify human-turn predicate on real logs
- general-purpose sonnet d1 tok=69643 calls=69 ms=785545 :: Write Phase 0 log schema note
- general-purpose opus d1 tok=48061 calls=7 ms=807810 :: Draft the frozen spec
- general-purpose sonnet d1 tok=141396 calls=93 ms=1570348 :: Build pinned fixture corpus
- general-purpose sonnet d1 tok=54547 calls=49 ms=600716 :: Document toolUseResult shapes in ref doc
- general-purpose sonnet d1 tok=205047 calls=142 ms=10240650 :: Implement the Deno CLI
- general-purpose sonnet d1 tok=114406 calls=153 ms=1482998 :: Follow session_id to span split sessions
- general-purpose sonnet d1 tok=45467 calls=79 ms=723485 :: Fix system-reminder exclusion rule
human turns
[2026-08-14T01:24:20.606Z]
# Goal
Build a Deno CLI that generates a git commit whose message is a condensed, …](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreiatamok5vkdihuntr2gn745st5qf4hwz3e5q4nrqixrjo2dtdqxja)
![Carl: Yes. Ah, I alluded earlier to playing a lot with agents. A good internet friend of mine made a very silly AI agent, MMO. So it's like a text based, if you're familiar with like a mud, a multi-user dungeon. It's basically that. It's a text based MMO that you can play through your terminal, and it's only for ais to play. Not everyone is, but, so I've been using that to learn about how to do agents and how to run AIs autonomously for a longer period of time, which has been fascinating. And so this, I've now watched somebody in real time, like I heard about this project. [36:07]
I went, oh, neat. Let me set up an AI to play this game. So I got there on like one, I think after it launched. Somebody else who got there on day one fucking ran with it has been, you know, he's been working professionally on AI agents for a year, he said. He already knew how to do all the harnesses in the orchestration, which I, you know, I had never tried to do either. [36:44]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreifdjk22qcrqczotnk5i4nrlo64m6wdbfr7ae74h7zlh6nrecd665i)
![So I'm learning this from scratch watching him do it with a year of experience and like, holy shit, it is unbelievable how much code he produced. I think the total amount of code is like, now, I dunno if it was 60,000 or 600,000 anymore, but it's like a lot of code and it works like not perfectly. There's a lot of duplication, there's a lot of errors, there's a lot of, you know, like this code path looks like it works, but actually you get into the details and it doesn't quite work. [37:06]
But like, as we use it, we like, so these AI agents are instructed as part of the game, like, "Not everything works. Please submit bug reports as you find them. Then this guy has his own agents that are reading those bug reports, turning it into reproductions, and then writing tests and fixing them, like largely autonomously. [37:34]
Not completely autonomously, but largely, which now has me imagining this, like imagine if you set up, you know, you have an idea for a product, you build a prototype. You put it on a server and then you have on the, on that same server, a little sidecar or you know, running in the same environment. I mean maybe not literally on the same server. [37:55]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreiaofpgjh7n73c7yd6kl56m6oi2lksr5hwt625is3e3pnfo4edk7fm)
![An AI agent that is receiving user feedback. Like, you know how like every major enterprise application will say every so often, like, "we value your feedback. How are you using this product?" And then you fill that out and it goes into the void and nothing ever changes. You never hear about it again. You never get asked for a follow up and the bugs that you reported never get fixed. [38:16]
Like what if that went into an automated system that triaged, reproduced and fixed things in real time autonomously as you submitted reports. Like that now appears to be within the realm of possibility. Like, like not theoretically in the future, but like currently I have seen tech that looks like it might be able to do that, like maybe not fully autonomously right now. [38:38]
Oh God. Like pretty good. Pretty close. So that's fascinating. I think that's really interesting for the idea of like a one person startup, you know, a small team operating at a scale that previously would've required 50 to a hundred people. I think that's interesting, like scratching your own itches and publishing it so that other people can use it, and then soliciting feedback to guide the development now looks really powerful. [39:04]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreicig3yncvy3wcjprmbs736a5oiwbnvsh6ws72rnpxlya4yjdum6bq)
![Carl: As I'm trying to think about how you might solve problems of caching and runtime things without directives. And I'm thinking about, you know, my experiences using like CloudFlare workers or other runtimes where it's not a traditional process scope. And I don't know, I guess like it is a really challenging question of like, how do you signal to a developer that the file they're editing runs in a certain type of context? [00:51:27]
And I guess I understand why directives are an appealing thing to reach for, but man, just the state we're in right now is really a lot. I found the Next.JS docs page with directives and it lists 'use cache', 'use client', 'use server', and then two sub variants of 'use cache'. And that's not all of 'em that, so there's missing ones here and. [00:51:56]
I don't know. When you start getting into like many different contexts, just naming them at the top of the file doesn't really seem like enough for me. I don't know. And then like the implication for that code can import other code that will not have it flagged what context it's running in via a directive at the top of the file. [00:52:20]
And I don't know. It's a tough problem. It's a weird unanswered question in the industry, like not even the ecosystem, but like the industry. How do you flag code that might run on your browser, a real server, serverless environment, and yeah, I don't know. It's a challenge. [00:52:40]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:vj54op7mobr3edx24yn6uax4/bafkreicc76g7w7tqjujjve674rvffratpk222dc46zzipffl2jraje36lm)