# AGENTS.md — detz (taicpp) workflow

Source of truth: `../tai.py` (untouched) + `../test_tai.py`, `../tai_test.py`.
Reuse: moloapp (`/home/retoor/projects/moloapp`) — clients/SSE, llm providers, app patterns. Vendor/link only, never modify.

## Workflow (per milestone, in order)
1. Spec delta: update `docs/<module>.md` + parity table vs tai.py symbols.
2. Code: headers `.h` + impl `.cpp`, one class per file pair, `-std=c++23 -Wall -Wextra`, warning-free.
3. Tests: `tests/*_test.cpp` mirroring python cases; `make test` green.
4. Changelog: prepend `CHANGELOG.md` entry.

## Commands
- `make` — build `build/detz`
- `make test` — build + run all tests
- `make lint` — warnings + clang-format check
- `make version-bump PART=patch|minor|major`
- `make changelog MSG="..."`
- `make tag` — annotated tag `detz-v$(cat VERSION)`
- `make release PART=x MSG="..."` — bump+changelog+commit+tag (never push unless asked)

## Rules
- No outputs outside `build/`.
- `src/cli/version.hpp` generated from `VERSION` at build.
- Max depth/layering, maximally OO. argparse tree must stay molo-subcommand-ready.
- Safe C++ is mandatory, per [docs/cpp-safety.md](docs/cpp-safety.md): any code that
  parses untrusted bytes uses overflow-safe size checks (`len <= size - pos`, never
  `pos + len`), reads lengths unsigned and range-checked before any allocation or slice,
  guards every shift/loop bound, and lets no exception escape a top-level dispatch loop.
  A new parser is unfinished without a hostile-input test and a sanitizer run
  (`tools/san-whatsapp.sh` is the pattern). We do not accept regressions here.