invoiceninja-cli/GOALS.md

Goals

What this crate is trying to be, what it is deliberately not, and what is still open. Same purpose as invoiceninjars-sdk's own GOALS.md: several choices here look like oversights until you know the reasoning, and that reasoning should live here rather than only in commit messages.

What this is for

A command-line way to drive the Invoice Ninja v5 API through invoiceninja-sdk, for the cases Rust code is the wrong tool for: a quick "what does this endpoint actually return", a one-off fix from a terminal, or a shell script that doesn't want to be a Rust program. The SDK's own GOALS.md names this exact need — its examples/ cover ad-hoc probing today, and this crate is the standalone version of that.

It is a client over the SDK, not a second implementation of anything the SDK already does. If a command needs logic beyond "turn flags and JSON into one SDK call", that is usually a sign the logic belongs in the SDK instead.

Goals

1. Every command is a thin translation, with one documented exception

list, get, create, and delete do nothing but turn CLI flags into a ListParams/model/id and call the matching SDK method. update is the one exception, and it stays an exception rather than a pattern to extend casually: it fetches the record, overlays the caller's JSON on top, and writes the merged result back, because Invoice Ninja's PUT is a full replace and the SDK's models default an omitted field rather than leaving it alone. See the comment in src/commands/clients.rs and the matching section in CLAUDE.md. Any other deviation from "thin translation" should be written down the same way.

2. --output json is always the full, lossless record

The table view is a convenience subset — whatever columns a resource's print_table picks — and it is allowed to leave things out. --output json is not: it is always the model's own Serialize output, including extra, so it is a drop-in input to --data/--file for a later update or create and a reliable thing to pipe into jq. This is the CLI-level version of the SDK's "never lose data in a round trip" goal.

3. One pattern across resources

Adding a resource should be mechanical, not a new design each time: a Command variant, a src/commands/<resource>.rs with the same run/list/get/create/update/delete/print_table shape as clients.rs, and a registration line in two places. CLAUDE.md's "Adding a resource" section is the checklist; if a resource needs something the template doesn't cover (bulk, uploads, entity actions, non-CRUD endpoints), that is scope for a deliberate addition to the template, not a one-off.

4. Scriptable by default

Errors go to stderr and a failure exits non-zero; nothing is printed to stdout on an error path that a script could mistake for a result. Secrets (--token, --api-password) never appear in --help or in any echoed default — see hide_env_values on each in src/cli.rs. The SDK already keeps both out of its own Debug output; this crate's job is not to undo that by printing them itself.

5. Inherit the SDK's safety defaults rather than re-deciding them

This crate does not expose its own retry/timeout policy or add a write path the SDK doesn't already consider safe. If the SDK decides POST isn't replayed on a 5xx because it might double-create an invoice, that reasoning applies here unchanged — the CLI is not a place to relax it for convenience.

6. CI mirrors the SDK's, job for job

.github/workflows/ci.yml has the same three jobs as invoiceninjars-sdk's own — pinned-toolchain test, build-only msrv on the declared floor, and an advisory stable that reports red but stays out of required status checks — rather than inventing a different shape for this crate. See CLAUDE.md's "Rust version and CI" for the per-job reasoning; a deviation from this shape should be written down there the same way update's exception is written down in goal 1.

Non-goals

• No Cargo workspace with the SDK. Explained in CLAUDE.md: the SDK is its own git repo with its own CI and crates.io release, mirrored read-only to GitHub. Now that the SDK is published (0.2.0, 2026-10-25), the dependency is a plain crates.io version requirement — there isn't even a path dependency linking the two build graphs anymore, let alone a workspace. This doesn't mean co-development across both crates is gone: .cargo/config.toml.example documents a per-developer paths override back onto a local ../invoiceninjars-sdk checkout, gitignored so it never becomes the default for anyone else or for CI. See CLAUDE.md.
• No interactive mode or TUI. Every command should work non-interactively from a script; a prompt-driven mode would be a different tool.
• No local state, cache, or config file — yet. Credentials and the base URL come from flags or environment variables only, resolved fresh on every invocation. See the open question below on multi-account profiles before assuming this is permanent.
• No stability promise before 1.0. Same posture as the SDK: flags are a compatibility surface too, and pre-1.0 they can change. Record breaking changes somewhere legible (a CHANGELOG.md, once there's enough history to need one) rather than letting them be discovered at the command line.

Open questions

These are genuinely undecided; deciding one is a change to this file.

• Bulk actions. The SDK exposes bulk_<resource> (archive/restore/delete plus resource-specific actions) on every resource with a /bulk route. Nothing here calls it yet. Worth having, but a bulk delete from a CLI invites a worse mistake than a one-record delete — whether that needs a confirmation prompt, a --yes flag, or a dry-run preview is unresolved.
• --dry-run for create/update/delete. Would print what the merged body or the target of a delete would be without sending the request. Cheap for create/delete; for update it would mean exposing the fetch-merge result for inspection before the write half runs.
• Multi-account profiles. Flags and env vars cover one account per invocation. A config file keyed by profile name (base URL + token) would help anyone working against more than one instance regularly, at the cost of being the first place this crate stores anything on disk.
• How closely should list flags track ListParams? Right now clients list exposes exactly what ListParams has typed setters for, plus --param key=value for the rest. A per-resource convenience filter (e.g. a --unpaid shorthand for client_status=unpaid) would read nicer but means hand-picking which filters earn a dedicated flag per resource.
• Whether the Nix packaging goes further than this repo's own flake. flake.nix now has packages.default/devShells.default (built with a pinned rust-overlay toolchain, for parity between nix build and nix develop) and overlays.default (for another flake to pull invoiceninja-cli into its own pkgs, built with that nixpkgs's own rustPlatform instead of the pin). The overlay path means this crate's rust-version = "1.88" has to be met by whatever rustc the consuming channel ships — there's no guard here if it isn't, just a build failure over there. Whether that's worth submitting to nixpkgs proper (and how an MSRV bump would then need to be communicated beyond this repo) is unresolved.