A local learning harness for using Codex as a practical coding coach instead of an answer generator.
The initial scope has two tracks.
systems: systems programming, concurrency, memory, files/processes, and performance practice with C/C++/Rustsecurity: defensive security engineering for large-scale software, including authorization boundaries, secure defaults, auditability, abuse resistance, and incident-ready design
Existing algorithm and C++ study repositories are integrated under imported/ as Git submodules.
git submodule update --init --recursive
python3 -m coach db init
python3 -m coach list
python3 -m coach list --track security
python3 -m coach show systems-bounded-queue-cpp-001
python3 -m coach new --task systems-bounded-queue-cpp-001
python3 -m coach new --track security
python3 -m coach test submissions/systems-bounded-queue-cpp-001/<attempt-id>
python3 -m coach hint systems-bounded-queue-cpp-001 --level 1
python3 -m coach review submissions/systems-bounded-queue-cpp-001/<attempt-id>
python3 -m coach request-problem --track security --level 2The new command copies starter code under submissions/. Implement the attempt there, run test for local checks, then use the prompt generated by review to ask Codex for a focused code review.
The request-problem command creates a Codex-ready prompt under problem_requests/. It uses the local DB plus submodule reference signals to ask for one new curated task. Codex should create the actual task JSON and starter fixture files; the harness then runs and tracks that task locally.
coach/ # dependency-free Python CLI
tracks.json # editable track catalog
tasks/ # JSON task definitions
fixtures/ # starter code and tests
submissions/ # local attempt outputs, gitignored by default
reviews/ # review prompts/results, gitignored by default
problem_requests/ # generated prompts for Codex problem generation
imported/ # legacy study repositories mounted as Git submodules
docs/ # design notes
tools/ # helper scripts
This project is currently a DB-backed harness MVP, not a fully autonomous agent. It can index local tasks and submodule reference files, track attempts, and generate Codex problem-generation prompts. Codex is still responsible for creating new curated task files from those prompts.
All first-class harness content should be written in English: task specs, fixture instructions, CLI messages, review prompts, docs, and future agent prompts.
Technical interview feedback should also be written in English. It should be direct, concrete, and interview-oriented: call out correctness issues, tradeoffs, edge cases, testing gaps, and how to explain the solution under interview conditions.
Files under imported/ are Git submodules that preserve the original legacy study repositories and may contain original comments or notes in other languages until they are promoted into curated harness tasks.