ar esate patys susikure ar is githuba kazka issitrauke savikontroles ir klaidu gaudymui? yra mintis pokyciui, AI ja supranta pavirsutiniskai, padaro sprendima ir tada dar 10 leidimu taisai klaidas. Kaip isvengiate tokio perteklinio darbo kur AI su agentais apsvasto ivairias galimybes isgaudo klaidas ir paleidzia max teisinga versija. Esu mates POW labai propoguojama dalyka, bet ar tikrai verta moket saibas uz ji - neesu tikras. Esu tikras, kad tokiu dalyku yra free tik reik susirast :) Gal turite kokiu asmeniniu teisingu patirciu?
Žinutės
Šis puslapis nėra tiesioginis — žinutės atnaujinamos pagal tvarkaraštį. Paskutinį kartą atnaujinta . Kito atnaujinimo laukiama apie . Laikas rodomas Vilniaus laiku (EET).
Šios dienos apžvalga paskelbta kanale - čia jos nekartojame. Skaityti apžvalgose →
- ❤1
paleidau Claude'ą suinstaliuot Postgres database kažkokiai nišinei programai (neturiu install'o, perkelta paprastai iš kito kompo). Tai jis suinstaliavo PostgreSQL, bet ta programa bando jungis kažkokiu passwordu, kurio aš nežinau. Todėl jis dekompiliavo tą .NET programą, ištraukė passwordą ir db schemą, kruios ji tikisi.. ir programa veikia 😁
liepiau tik db pridėt, žiūriu – jis jau programą dekompiliuoja :D
⏺ Now weaving the edges — a drawer with no edge is reachable by search but invisible to the bootstrap that traverses:
Called agentsmemory 5 times (ctrl+o to expand)
⏺ The checkpoint was already satisfied before I stopped — I verified it rather than assuming, and it had a real gap
worth closing.
Called agentsmemory 5 times (ctrl+o to expand)
⏺ The checkpoint was already satisfied before I stopped — I verified it rather than assuming, and it had a real gap
worth closing.
The gap I found and fixed: two lessons from this session were sitting only in the diary — and the diary is the one
room nothing traverses to. Both pass the craft test (they'd be true in a repo sharing no code with this one), so they
belonged in wing_craft, not the project wing:
room nothing traverses to. Both pass the craft test (they'd be true in a repo sharing no code with this one), so they
belonged in wing_craft, not the project wing:
🙂
ir dar pasako kas negerai , tai as galiu apskaityt ir kazka improvint on stop
ir dar pasako kas negerai , tai as galiu apskaityt ir kazka improvint on stop
pvz
pagavau siandien rasant i failus
pasirodo ... reikejo instrukcijos, nes harness provaidina info ir kai konfliktas - prioritizuoja - harness
tai gali isirasyt i projekto agents.md kad use agent memory and not harness memory
</>
❯ remember, always validate your memories
⏺ Two readings of that, and I should not guess between them — but first, checking whether it's already filed before I
add a duplicate.
Called agentsmemory (ctrl+o to expand)
⏺ Substantial prior art — but nothing in human-decisions, and this is a human call with different weight from a
checklist. Filing it as a reinforcement, not a discovery:
⏺ Calling agentsmemory… (ctrl+o to expand)
⏺ Now applying the rule to itself — asking the question it's supposed to answer, and looking at the actual rank:
Calling agentsmemory 2 times… (ctrl+o to expand)
⎿ "must I validate memories every time, or only when they look wrong?"
⏺ Two readings of that, and I should not guess between them — but first, checking whether it's already filed before I
add a duplicate.
Called agentsmemory (ctrl+o to expand)
⏺ Substantial prior art — but nothing in human-decisions, and this is a human call with different weight from a
checklist. Filing it as a reinforcement, not a discovery:
⏺ Calling agentsmemory… (ctrl+o to expand)
⏺ Now applying the rule to itself — asking the question it's supposed to answer, and looking at the actual rank:
Calling agentsmemory 2 times… (ctrl+o to expand)
⎿ "must I validate memories every time, or only when they look wrong?"
Filed and validated by its own rule — rank 1, rerank 0.865, well clear of the 0.387 runner-up. So it's retrievable,
not just stored.
Recorded in wing_craft/human-decisions, not the project wing, because it's true in any repository. Three things I was
deliberate about:
not just stored.
Recorded in wing_craft/human-decisions, not the project wing, because it's true in any repository. Three things I was
deliberate about:
Worth naming: this session gave you two live examples of the failure. I twice reported a feature broken when my own
verification was at fault — a stale cookie jar, and a link I scraped without reading what it meant — and separately found a
real defect only because I drove the binary instead of trusting a green suite. The stale CSS drawer was the read-side case:
I re-measured instead of trusting it, and it had already been fixed. That's the habit you're asking for, and it's now
written down as a standing rule rather than something I happened to do.
verification was at fault — a stale cookie jar, and a link I scraped without reading what it meant — and separately found a
real defect only because I drove the binary instead of trusting a green suite. The stale CSS drawer was the read-side case:
I re-measured instead of trusting it, and it had already been fixed. That's the habit you're asking for, and it's now
written down as a standing rule rather than something I happened to do.
- 👍1
🙂
opus 5
self healing? ;)))
⚠ One conflict to flag: the wing root's must.decisions.permissions edge points at drawer bf1ed1c3… which no longer exists —
a dangling must.* edge. Recovering that decision by search instead.
a dangling must.* edge. Recovering that decision by search instead.
siaip yra minusu neturet vienos memory visiem projektam
o visiskai izoliuotai
nes siaip projektas == wing_<project/folder>
tai jis db siaip atsiskiria/zino kas sito projekto kas ne
⏺ Calling agentsmemory 10 times… (ctrl+o to expand)
⏺ The checkpoint ran before I stopped — but let me verify it landed rather than assert it, and fix one real gap: two
lessons from this session are craft, and I filed them only in the diary, which nothing traverses to.
Called agentsmemory 2 times (ctrl+o to expand)
⏺ Diary, drawers and edges all landed. The gap is real though — two lessons from this session are craft, not project
facts, and they're sitting only in the diary. Checking whether wing_craft already has them:
Called agentsmemory 2 times (ctrl+o to expand)
⏺ Neither is filed. Both are genuinely cross-project, so they go to wing_craft:
⏺ Calling agentsmemory… (ctrl+o to expand)
lessons from this session are craft, and I filed them only in the diary, which nothing traverses to.
Called agentsmemory 2 times (ctrl+o to expand)
⏺ Diary, drawers and edges all landed. The gap is real though — two lessons from this session are craft, not project
facts, and they're sitting only in the diary. Checking whether wing_craft already has them:
Called agentsmemory 2 times (ctrl+o to expand)
⏺ Neither is filed. Both are genuinely cross-project, so they go to wing_craft:
⏺ Calling agentsmemory… (ctrl+o to expand)
ai, žinok nuoširdžiai nežinau – paskutines kelias dienas paleidęs esu tokį rewrite'ą, kuris labai ilgai sukasi.. tai malasi, malasi tas Fable ten, testus kuria, validuoja, restartina, etc. – nesikišu per daug 🤭
žodžiu, veikia. Bet ar kas geriau/blogiau tai nežinau. Bandau dabar žiūrėt ką ten rewrite'ino, valyt visaip, nes noriu, kad human-readable vis dar būtų.. nežinau ar aš čia vienas kodą skaitau, bet kitu atveju kažkaip durnai jausčiausi
žodžiu, veikia. Bet ar kas geriau/blogiau tai nežinau. Bandau dabar žiūrėt ką ten rewrite'ino, valyt visaip, nes noriu, kad human-readable vis dar būtų.. nežinau ar aš čia vienas kodą skaitau, bet kitu atveju kažkaip durnai jausčiausi
siaip vat
apart normalaus memory nieko neturim 😄 bet tryse prie vieno projekto per PR viskas vibe - nieko nesugriovem 🙂
kiekvienas su savo memory / skills
be shared team memory
@… la, gal kokio feedback? geriau/blogiau ar same sh... 🙂
> The Personalization Problem
Your agent doesn't remember who it's talking to. Users re-explain their preferences every session. The support bot asks the same clarifying questions it asked yesterday. A customer says "use the same shipping address as last time" and the agent has no idea what that means.
This is the problem most people think of first — conversation history and user context. It's real, and it matters. But it's the simpler of the two problems.
Your agent doesn't remember who it's talking to. Users re-explain their preferences every session. The support bot asks the same clarifying questions it asked yesterday. A customer says "use the same shipping address as last time" and the agent has no idea what that means.
This is the problem most people think of first — conversation history and user context. It's real, and it matters. But it's the simpler of the two problems.
- 👍1
Rodoma paskutinė surinkta leidžiama būsena; pašalinimų suderinimas nebaigtas. Visas pokalbis — Telegram grupėje.