Aleks Petrov

Claude Ambassador • Harness AI engineer

Pilot

[Aleks Petrov @aleks_maker]

@PreslavMihaylov Thanks 👋

[Aleks Petrov @aleks_maker]

Mods are great in Claude Code.

Added responsiveness and a few more features to the Navigator plugin mod.

https://t.co/ZTp72hyh9G

[Aleks Petrov @aleks_maker]

Above the prompt input, one line tells you where you are.

Ask about something unrelated twice and it turns orange "off route — park it or go back".

Park turns the detour into a task doc for later; back drops it.

Open source, MIT.

Works with Claude Code 2.1.287+

Repo:
https://t.co/JtFcmlchOg

Docs (temp)
https://t.co/SxQ9xj4daq

[Aleks Petrov @aleks_maker]

Navigator 8.0.1 released 🎉

This update changes what you see while you work. Navigator now runs as a Claude Code mod — and it keeps you on track.

v8 runs Navigator's ops inside Claude Code (2.1.287+) instead of a Python process per hook event. Same ops, byte-for-byte parity on ~8k generated cases.

* Python stays as the fallback on older versions *

Type "/nav" — a pane opens next to the chat:

[Aleks Petrov @aleks_maker]

What's on the panel

Next
The step you're on, as a button. Press n and it becomes your next prompt,
with a time estimate from how fast you've been going.

Context
How full the AI's memory is, and whether it's time to clear it.

Judge
Navigator classifies every prompt before the AI sees it (small fix? big task? unclear?) and shows its call, so you know why it asked for a plan or didn't.

Memories
Three past decisions relevant to your task; pin one into your next prompt.

[Aleks Petrov @aleks_maker]

New @gitlab UI made by drunk vibecoder. Guys, did you try to create a token? Spent and hour, cool, no try to update it's preferences, oh, you can't, there is no edit, do that again.

[Aleks Petrov @aleks_maker]

Pilot harness + Sonnet 5.5 = deployment machine-gun

pure gold 🤌

[Aleks Petrov @aleks_maker]

Navigator v7.7.0: the keyword matchers got a judge.

Why it exists

Every prompt-time decision in Navigator's hook runtime was a keyword match. Fourteen phrases decide Loop Mode. A weighted keyword sum decides Task Mode. Ambiguity is a base of 0.5 minus credits for file paths and numbers. Cheap, deterministic, and wrong in ways you spot instantly.

Last week a subagent report came back with a section header that read "F. Loop mode". The matcher saw the phrase and put the session into Loop Mode. Nobody asked for autonomous iteration. A header did.

How it works

With judge.enabled, each prompt goes once to TypeSafe's Jev — a model that answers typed questions with probabilities instead of text. Eight questions in one request:

→ is this a task, or a question / a reply / a pasted log
→ does it ask for unattended iteration until done
→ complexity on a four-level rubric
→ ambiguity on three
→ are scope, limits, approach, verification stated

Each axis overrides its keyword heuristic only when the answer is decisive. Undecided, timeout, no key, any error → the keywords answer, byte for byte. The heuristic is the floor; nothing was deleted.

That "Loop mode" header scores 0.14 on the loop question. Silent.

Numbers, 60 labeled prompts including the false-fires:

• tier (direct / task / loop): 38 → 53 of 60
• task-shaped: 43 → 54 of 60
• ~0.7 s per prompt, ~700 input tokens, ~$0.00003

Thresholds came from a sweep over recorded responses, replayable offline. The cookbook defaults lost by one or two prompts per axis.

Why it's different from "just ask the model"

• Typed answers, not parsed prose — a probability per question, nothing to regex
• Decisive-or-fallback per axis — the judge can't make a decision worse than the keywords; it can only replace one it's sure about
• One request, eight questions, parallel — one network call per prompt, cached for both ops that need it
• Privacy is explicit — ships OFF because the prompt leaves the machine; secret-shaped tokens are redacted before it does; tool output and files never go

Try out

/plugin marketplace add alekspetrov/navigator
/plugin install navigator
export TYPESAFE_API_KEY=...
"enable judge"

Open source, MIT. Eval fixture and recorded responses ship in the repo.

https://t.co/uhUopgSlwJ

[Aleks Petrov @aleks_maker]

How many lines do we ship on the daily?
Say, me say many, many, many, many.

Hire us:
https://t.co/Jzm89K0L0p

[Aleks Petrov @aleks_maker]

@neeraj7105 Claude is down

[Aleks Petrov @aleks_maker]

No rest, no sleep — just ship 🚀

What’s your longest Claude Code streak?

[Aleks Petrov @aleks_maker]

@maruf_ix AWS

[Aleks Petrov @aleks_maker]

@aslambeg84 None?

[Aleks Petrov @aleks_maker]

No excuses for slow web web in 2026!

We ship blazing fast products for real users.
Built on AWS + Next.js, tuned by paranoia.

Work in progress, already 100% performance, 3/3 on agentic search optimization 🤌

Hire us:
https://t.co/eWeC6pTO5F

[Aleks Petrov @aleks_maker]

@i_mika_el Exactly the reason it was created.

[Aleks Petrov @aleks_maker]

Grafana/Prometheus, but in your Terminal.

Grom turns metrics into a live TUI dashboard.
Open-source, fast, written with Go by Quant Flow Studio.

Repo: https://t.co/gAOujUqU9s

#devtools #metrics #analytics #terminal

[Aleks Petrov @aleks_maker]

Fable 5 orchestrated Pilot agent to ship a self-hosted AI analyst in just 3 days 🤯

Secured by design.

You chat with your production DBs, PDFs, External APIs, etc. Numbers compile deterministically from versioned metric definitions.

Made by:
https://t.co/eS6TdyHSdd

[Aleks Petrov @aleks_maker]

Moved our autonomous dev daemon from a MacBook to EC2.

10 parallel autonomous executions on one t3.xlarge. ~$0.65/task.

Let's go 🚀

[Aleks Petrov @aleks_maker]

Pilot just shipped its own refactor, on schedule.

All 10 integrations — GitHub, Linear, Jira, GitLab, Slack, Telegram… — now run on one shared SDK. The duplicated in-tree adapter code had to go.

No human wrote the code.
https://t.co/eS6TdyHSdd

[Aleks Petrov @aleks_maker]

This oven is hot 🔥

Pilot - ai coding agent pushing it's own queue.

https://t.co/eS6TdyHSdd