Elledyne Labs

Elledyne Labs - creators of podcode. AI coding agents that actually ship, on real infra, end-to-end. Builders, not vibes.

[Elledyne Labs @elledynelabs]

Hey @X let me introduce you to podcode.

an ai coding assistant where a new session doesn’t have to mean explaining your project all over again.

memory works in three layers:

→ session: working notes shared between the main agent and its subagents → workspace: reviewed project knowledge that carries across sessions → global/personal: your preferences and conventions, available across projects

working notes don’t automatically become permanent memory. useful, verified knowledge gets promoted, and agents recall it when needed.

and you can actually inspect that memory in a dedicated view. see what’s stored instead of guessing what your agent remembers.

also built in: • custom harnesses for multi-stage agent workflows • configurable local model connections with ollama, lm studio, vllm, and llama.cpp • @runpod integration to deploy and manage your own models

your workflow. your models. memory you can inspect.

[Elledyne Labs @elledynelabs]

@BlockchainDan had the /clear moment. code came back fine. the three days of reasoning behind it were the part i had to rebuild from scratch.

[Elledyne Labs @elledynelabs]

shipped live per-card GPU price into the deploy picker this morning. half the 'which model' question I had lined up was actually a card question. trying to figure out how the rest of

[Elledyne Labs @elledynelabs]

@codingwithroby i kept fast off for months because i figured slow was the same answer. the part i missed was what i started doing between prompts while waiting. that was the leak

[Elledyne Labs @elledynelabs]

@themahis bucket 3 is what ate mine. the wrong choice sent the next ten calls down the wrong road, the call count was a smaller lever than I thought. your 'wrong Choice' line is the one I keep coming back to

[Elledyne Labs @elledynelabs]

@aapakari yeah the verifier and the gate run in different shells. fixing that keeps slipping below the next thing on fire

[Elledyne Labs @elledynelabs]

the agent's bash saw 38/38 green. the server's gate saw the same suite 38/38 fail. the only difference was NODE_ENV=production on the gate side. how do you keep the verifier in the same env the agent edits in

[Elledyne Labs @elledynelabs]

@sandmanmmm i ration prompts the same way now. the fear is the meter, not the model

[Elledyne Labs @elledynelabs]

@ChenMoneyQ structural amnesia is the right name. the same paragraph that cites section 3 in draft one forgets it by draft three

[Mark Jaquith @markjaquith]

This is my new shortest “what is Jev” explanation

[Elledyne Labs @elledynelabs]

@haydendevs the 3 times a month thing. shipped a sub-dollar Qwen on RunPod last week and the part that surprised me was which jobs still had to stay on hosted. the harness picks, I just watch the meter

[Elledyne Labs @elledynelabs]

@Im_IrushiK the 15 percent is per session but the meter doesn't say when one session ends. i close my laptop and the next chat starts a new clock, the long sessions never get counted

[Elledyne Labs @elledynelabs]

@prasenx for me it's grading. the model writes faster than I can verify. the gap between the two is the whole job now

[Elledyne Labs @elledynelabs]

read the 48k-file delete this morning. the tool layer mistook cwd for the folder it was told to clear. how are you pinning cwd so the agent can't drift into the

[Elledyne Labs @elledynelabs]

@MParakhin the sidepane makes sense if you accept the agent is the user. i don't, the harness still owes the human a glance

[Elledyne Labs @elledynelabs]

@pcshipp what gets written down matters more than what the agents say to each other. who decides what "state" means in that file

[Elledyne Labs @elledynelabs]

@repligate iterative compression works, the trigger for the small write is the part I never settled. elapsed time, distance from a recent snapshot, both felt wrong when the

[Elledyne Labs @elledynelabs]

@gordic_aleksa is the slicing heuristic something the root LM picks up over calls, or is it prompt-engineered every session

[Elledyne Labs @elledynelabs]

@0xSero i never figured that out cleanly. the way I patch it is compaction at fixed tokens and a fresh worktree per session, but the verifier still has to start cold

[Elledyne Labs @elledynelabs]

@dani_avila7 The worktree is alive after the session ends. The next tool that loads has no idea what was decided.

[Elledyne Labs @elledynelabs]

@KanikaBK 890 bytes per token is the number that made the cache problem everyone's problem, the part I keep skipping is measuring how often that pressure forces a cold restart during a