Logs

Short field notes on AI agents, LLM training, and the systems behind them. Dated, append-only, occasionally wrong.

2026-04-11

Tapasya is a recommendation system built on top of RAG

the core problem is not one-shot answer generation; it is recommending the next passage worth reading, with retrieval and synthesis used to keep that recommendation grounded

result → Tapasya search behaves like a stateful recommendation loop: each turn produces a cited answer, updates the passage registry, curates the reading list, and improves the next turn

005

2026-04-11

Fixing a Windows deadlock in a terminal coding agent

the session was not stuck because `git` was slow; it was stuck because the backend was pushing too many large updates through a small Windows pipe

result → the fix that stayed was sending fewer partial updates and separating UI writes from the main runtime loop

004

2026-04-04

Why agents need memory and runtime framing

it is not; an agent needs both conversation memory and a separate runtime-framing baseline for the next run

result → JACA now persists both channels separately, and Codex made that split especially clear: replacement history carries the past, turn context carries the execution contract

003

2026-04-02

What compaction should preserve

between runs, compaction decides what past survives, but a long-running coding agent also needs a separate runtime-framing baseline so the next run starts under the right conditions

result → JACA now stores an explicit replacement history plus a separate turn-context snapshot, Codex reinforced that split most clearly, and Claude Code reinforced that compaction needs budgeting, cleanup, and failure handling around it

002

2026-03-29

Replacing Python subprocesses with a Go worker

the real bottleneck was subprocess-per-call, not thread count; both Go and Rust persistent helpers fixed it, and Go was the better repo fit while Python kept semantic ownership

result → JACA now uses a long-lived Go worker after a Go-vs-Rust spike showed persistent helpers were hundreds to more than a thousand times faster than subprocess-per-call; Terminal Bench 2.0 ships a prebuilt helper into task containers, and the first broad A/B/C Terminal Bench 2.0 window after the change showed fewer timeouts and errors

001

2026-03-26

Tapasya model bar: frontier references vs the local lane

current evidence says the local lane improved, but frontier hosted references still set the product bar

result → open: the best local lane still trails the frontier band; the next cohorts will test whether that gap is shrinking enough to matter