TAYI Project State

Last updated: 2026-08-24

Summary

TAYI is a working local research prototype for a consumer-GPU coding agent. It currently has:

a Rust CLI for indexing, retrieval, context construction, solving, patch validation, and test execution;
a Python research package for evaluation, distillation, teacher repair, capability analysis, and adapter training flows;
local gold fixtures across Python, JavaScript, Rust, Go, and Java;
scripts for workspace-local environment setup, llama.cpp serving, open-teacher serving, WSL training, and cleanup;
documented v3 context/curriculum and v5 system-optimization research tracks.

The repo is ready to be saved to GitHub as source and documentation. Large generated assets stay local and are restored separately.

Status

Current status is research prototype, not production model release.

Working:

CLI command surface builds and exposes index, analyze, retrieve, context, solve, apply-patch, test, doctor, and local-check.
Workspace-local Cargo target/cache policy is configured through .cargo/config.toml.
Python tests cover the research modules.
Local model and teacher paths are standardized under model/.
Large files are excluded by .gitignore.
README and restore docs explain how to hand off the project.

Known limits:

v5.1 semantic adapter was not promoted because target3 stayed at 2/3.
Current verified semantic repair data is too small for a meaningful capability jump.
Full broad-language claims require more fixtures and verified repairs.
Large local assets are not in Git and must be re-created or preserved outside the repository.

Measured Capability

Known local readout from the existing project notes:

System	Eval slice	Result
Base 7B HF student	local gold-20	`13/20`, pass@1 `0.65`
Early TAYI LoRA v0	local gold-20	`14/20`, pass@1 `0.70`
Runtime v4.4	target3	`2/3`, pass@1 `0.6667`
v5 system adapter	target3	`2/3`, pass@1 `0.6667`
v5.1 semantic adapter	target3	`2/3`, pass@1 `0.6667`, not promoted

Interpretation:

TAYI has a real repo-level loop and measurable progress.
It should be described as a promising local coding-agent system, not as a finished frontier-class model.
The next bottleneck is verified semantic repair data scale and harder fixtures.

Hardware Policy

Primary target:

RTX 5060 Ti 16GB or similar consumer GPU.

Policy:

hard VRAM ceiling: 16GB;
preferred active band: 8-12GB for routine work;
training and serving should reduce context, batch size, rank, or offload before exceeding the ceiling;
use retrieval and context compression instead of brute-force full-repo context.

Git Tracking Policy

Track:

source code;
tests;
scripts;
configs;
docs;
brand assets;
small example fixtures;
schemas and small task definitions.

Do not track:

model/;
artifacts/;
.cache/;
third_party/;
downloads/;
.wsl/;
.aicoder/ and .tayi/ runtime traces;
checkpoint, GGUF, safetensors, or temporary training output files.

Next Mainline

Next milestone:

verified_semantic_repair_pack_v1

Definition:

build 20-50 verified semantic-repair samples;
include source context, paired specs/tests, failed candidates, logs, and passing patches;
expand fixture coverage beyond the current small local set;
run teacher repair only through isolated verification;
train v5.2 with more than an 8-step smoke only after the verified pack exists;
promote only through fixed quality gates.

Recommended Next Commands

. .\scripts\tayi_env.ps1
cargo test
$env:PYTHONPATH="src_py"; python -m unittest discover -s tests_py

Then generate or refresh the next training/eval artifacts:

$env:PYTHONPATH="src_py"; python -m aicoder_research.system_optimizer --help
$env:PYTHONPATH="src_py"; python -m aicoder_research.gold_eval --help
$env:PYTHONPATH="src_py"; python -m aicoder_research.failure_refinery --help