Question: about logic

[r0bis]

Sorry for a naive question - where does the logic and the actual words of the code comment come from? I suppose part of the question is - does it incur claude token use as well and if so - roughly how many?

[1270011]

It comes from the actual model you're already talking to. Claude just impersonates your buddy in the last sentence and appends it as an invisible HTML comment (<!-- buddy: ... -->). A Stop hook extracts it and shows it in the status line. No separate API call, no Haiku, no extra model - just one extra sentence from the same conversation.

Token estimate: roughly ~1.5k tokens at startup (MCP tool definitions + instructions + skill) and ~20 tokens per response (the buddy's comment). Over a full 50-turn session that's about 2.5k tokens total — somewhere around 1% of a typical session. Not a completely free meal, but 1% of a 5-hour session window means you lose about 3 minutes worth of tokens.

I'll add proper diagnostics tooling so people can measure it themselves (bun run doctor already shows the setup, planning to add a token counter as well). #19

[r0bis]

That's ingenious. The cost is quite ok. But as for the mechanism - there is then an invisible request to claude to produce that last sentence in HTML comment; something that gets passed along with your instruction for coding, and that something contains the required mix of patience, wisdom and snark parameters so that claude can construct the buddy comment?

[iilonmasc]

@1270011 quick info, if not already noticed: This also relates to #76