# CLAUDE.md Guidance for working on this repository. It records decisions made when the project was started (August 2026) that are not obvious from the code. ## What this is Small single-file Python utilities for **Conan 1.x, targeting 1.66**. Not Conan 2. The first (and so far only) tool is `conandeps.py`; see `README.md` for what it does and how it is used. ## Commands (all inside the container) ```bash ./dev.sh build # first time, and after editing Containerfile ./dev.sh python -m pytest # tests (starts a throw-away conan_server) ./dev.sh ruff check . && ./dev.sh ruff format . shellcheck dev.sh # after editing the wrapper (runs on host) ./dev.sh uv lock # after changing pyproject.toml ``` Users install/upgrade on their own machine with `uv tool install git+https://kode.naiv.no/olemd/conan-1.6-utilities.git` / `uv tool upgrade conan-utils`; a pushed commit is immediately installable. ## Hard constraints - **Conan 1.66 only.** Use the `conans.client.conan_api.ConanAPIV1` API and Conan 1 graph internals (`Node.binary`, `Node.dependencies`, `conanfile.requires`). Do not introduce Conan 2 APIs or a Conan 2 code path. - **Python 3.8+.** Conan 1.66 runs on new Pythons too (it falls back to `importlib` where `imp` is gone) – verified on 3.14. It does emit SyntaxWarnings from its own modules there, which `conandeps.py` silences before importing `conans`; keep `conans` imports deferred so that filter is in place first. `requires-python` in `pyproject.toml` and the PEP 723 header must stay in sync. - **All development runs inside the container.** Use `./dev.sh ` for every python/pytest/ruff/uv invocation. Never run `conan` against the host's `~/.conan`; never `pip install` on the host. Rebuild the image with `./dev.sh build` after touching `Containerfile`. - **Never depend on the private remote.** The user's real remote only exists on the machine where the tool is used and must not be named in the repository – use a placeholder such as `myremote` in docs and help text. Tests must use the local `conan_server` fixture in `tests/test_conandeps.py`. ## Design decisions (and why) - **Binary availability is computed via Conan's graph, not `conan search`.** The graph is built once per build type through `ConanAPIV1.info()` and each node's `binary` status is read. This honours `package_id()` overrides, options and `default_package_id_mode` exactly like `conan install`. Do not replace this with settings-dict matching against `conan search` output. - **"Missing" means missing for the exact profile.** Not "no binary with that build_type at all". - **Dependency table shows the shortest path, not a level number.** `_paths()` does a breadth-first walk from the root; `Dep.path` is the list of package names on the *shortest* path (empty = direct, so a package that is both direct and transitive is direct). `Dep.paths` holds one path per direct parent (shortest path to that parent + parent) – bounded, unlike "all paths". The `via` column renders them one per line as `A -> B` (`→`/`
` in HTML); `_table()` supports multi-line cells. Rows sort on (depth, path, host before build, name). The tree section keeps graph order. - **Tree tags are typed.** `_tree_rows()` yields `(head, [(kind, text)], tail)` with kinds path/build-require/override/range/missing; `_tree_lines()` (text, ANSI via a `fmt_tag` callback) and the HTML renderer (``) colour by kind from that one source. Add new tree annotations there, never by string-replacing rendered lines. - **Only the remote counts, and the source is always shown.** Conan says `Cache` for a locally cached binary without asking the remote, so `_RemoteIndex` verifies cache hits with `search_packages` on the remote. Cache-only binaries are reported as `not available` and count as a problem (exit 1). Every available cell names the remote. The user rejected an opt-in flag for this – it is the tool's core question. - **Overrides are parsed from Conan's WARN output.** Conan 1 does not record overrides on the graph; `Requirements.update()` mutates `req.ref` in place and only emits `": requirement overridden by to "`. `conandeps` captures the output stream with a non-coloured `ConanOutput` and matches `_OVERRIDE_RE`. `test_overrides` pins this format – if a Conan patch release changes the wording, that test is the alarm. - **Both explicit and implicit overrides are reported.** `override=True` requirements are collected from `conanfile.requires` to label an override as explicit; everything else parsed from the WARN lines is implicit. - **`range_ref` is not a version range after an override.** Conan reuses `Requirement.range_ref` to hold the pre-override reference, so only treat it as a range when `req.version_range` is truthy. - **Single file, stdlib + Conan only.** No third-party deps beyond Conan. Packaging is hatchling with `only-include = ["conandeps.py"]`; the PEP 723 header makes the file usable via `uv run conandeps.py` without a checkout. - **Exit codes:** `0` fine, `1` missing binaries, `2` Conan error. CI relies on this. - **Tables and colour, consistently.** Dependencies, missing binaries and overrides are all rendered through the same `_table()` helper (text) and `_html_table()` (HTML). Colour is an aid only – every state is also spelled out in words – and `_table()` aligns on *visible* width so ANSI codes never break columns. `--color auto` is the default (terminal + no `NO_COLOR`). - **stdout is the report, stderr is progress.** Graph resolution against a remote takes a while; progress lines (`_progress`) and `--verbose` live Conan output go to stderr only, so stdout can be redirected to a file. ## Workflow - Lint/format: `./dev.sh ruff check .` and `./dev.sh ruff format .` (config in `pyproject.toml`, target py38, line length 100). Run `shellcheck dev.sh` after editing the wrapper. - Tests: `./dev.sh python -m pytest`. The fixture spins up `conan_server` on a free port with a pre-written `server.conf` (the port cannot be given on the command line), creates fake packages whose `build()` does nothing, uploads them, then wipes the cache so binaries can only come from the remote. Fake recipes need `--build=missing` at create time because dependencies' binaries are deliberately incomplete. - **Assert on table output with regexes, not fixed-width strings.** Column widths depend on the longest ref in the fixture and package_ids change whenever a fixture recipe's requirements change (`semver_direct_mode`). Use `re.search(r"libbar/1\.1\s+libfoo\s+host", text)` and `[0-9a-f]{40}` for ids. Also: `_conan()` in the tests raises with Conan's full output on failure – read it before guessing. - **Use the `conan_env` fixture only for graph/remote behaviour.** Anything about rendering (`_via`, `_table`, `_tree_rows`, HTML) should be a fixture-free unit test on hand-built `Dep` objects, like `test_via_lists_every_parent_path` – instant, and no `conan_server`. - Dependencies: `uv.lock` is committed; regenerate with `./dev.sh uv lock` after changing `pyproject.toml`. - Commits: atomic, no `--amend`; run ruff + pytest + shellcheck before committing. Add a memory of anything a future session could not derive from the repository. - Releases: bump `version` in `pyproject.toml` (must match the tag), commit, then `git tag -a vX.Y.Z && git push origin vX.Y.Z`, build with `./dev.sh uv build` (writes `dist/`, gitignored), write `SHA256SUMS`, and `fj release create --tag vX.Y.Z --attach dist/ --attach dist/ --attach dist/SHA256SUMS --body "..." "vX.Y.Z"` from inside the repo – `fj release` does **not** accept `--repo`. Then update the "Latest:" line in the README's Releases section and delete `dist/`. ## Adding another tool Follow the `conandeps.py` pattern: one executable file with a module docstring explaining *how* and *why*, a PEP 723 header, a `main(argv)` that returns an exit code, a `[project.scripts]` entry, an `only-include` entry in `pyproject.toml`, tests under `tests/` using the shared `conan_env` fixture, and a section in `README.md`.