loader Public

Notifications Fork 0 Star 0
4effa19
Branches trunk
1 Branches 0 Tags
Go to file T
Code
espadonne Plan Sprint 13 semantic diff work
4effa19 162 Commits
.docs Plan Sprint 13 semantic diff work
src Extract main turn loop control from conversation runtime
tests Add direct tests for turn loop control
.gitignore Ignore Loader runtime state artifacts
MODELS.md fix: approval bar focus and threading issues
README.md Rewrite README for Loader's actual runtime
build.sh fix: approval bar focus and threading issues
package-lock.json fixes to agent behavior and more i forget
package.json fixes to agent behavior and more i forget
pyproject.toml Stabilize Loader test collection and package imports
uv.lock chore: update lockfile

Loader

Loader is a local-first coding assistant that runs against Ollama and drives a small tool-using agent loop from the terminal or a Textual TUI.

Current state

Loader is still early-stage. The project already has:

  • a working terminal and TUI interface
  • a default tool set for file reads/writes, editing, search, globbing, and shell commands
  • project-context detection
  • runtime safeguards, recovery logic, and heuristic completion checks

It does not yet have the stronger runtime architecture described in .docs/REPORT.md. The current implementation plan lives under .docs/sprints/, and the Sprint 00 runtime baseline is tracked in .docs/PARITY.md.

Requirements

  • Python 3.11+
  • uv
  • a running local Ollama server on http://localhost:11434
  • at least one pulled Ollama model

Setup

uv sync
uv sync --extra dev

Common commands

uv run loader
uv run loader "write a hello world program"
uv run loader --no-tui
uv run loader --select-model

uv run pytest
uv run pytest tests/test_runtime_harness.py -q
uv run ruff check src tests
uv run mypy src

Repository notes

  • Loader source lives under src/loader/
  • tests live under tests/
  • reference repos used for comparison live under refs/ and are gitignored
  • uv run pytest is scoped to Loader's own tests/ directory

Known limitations

  • the agent loop is still monolithic and heuristic-heavy
  • there is a known tool-result contract regression captured by the runtime harness and left red for Sprint 01
  • permissions are still confirmation-based rather than mode-based
  • session persistence, planning artifacts, and verification loops are not implemented yet