salpa-cli

The Salpa node authoring CLI. Scaffolds node packages that match the shipped template contract — the deterministic path to a new node — checks them against that contract, and puts them into the Salpa app on your machine.

python3 -m pip install salpa-cli
salpa new my-analyzer          # interactive
salpa new my-suite -t multi-node-package --yes
salpa dev                      # guided: what is left, and the command for each
salpa add gromacs mdanalysis   # find them, add them, narrow the platforms, solve
salpa validate my_suite        # will it install and register?
salpa smoke my_suite           # does it actually run?
salpa env status               # what is built, and which checks would run
salpa push my_suite            # into your running Salpa app
salpa unpush my_suite          # and back out again

Installing adds the salpa command. If your shell can't find salpa afterward, your Python scripts directory isn't on your PATH — installing into a virtual environment is the simplest fix (it puts salpa on PATH).

Requirements

Python 3.9 or newer. pip install salpa-cli reports a clear error below that.

• macOS — the system python3 is 3.9, which is enough. For a newer one: brew install python, or python.org.
• Linux — most distributions ship 3.9+. Otherwise use your package manager (apt install python3, dnf install python3, …) or pyenv.
• Windows — often has no python3 at all; install from python.org or the Microsoft Store, and tick "Add python.exe to PATH".

pixi — needed only for pixi run test (building a node's environment and running its tests). You do not have to install it separately if you run Salpa: the app keeps a pixi under ~/.bocoflow, and salpa new finds it and prints its path. To have one on your own PATH, install it from pixi.sh. Stdlib-only nodes need no pixi at all.

salpa new creates ./<under_scored_name>/ from a bundled template, with every placeholder substituted and the directory underscored (the dir name becomes the Python package name, so a hyphen would break from .core import ...; meta.toml's name stays kebab-case).

Templates

Templates ship inside the wheel and are the single source of truth for the node contract, so there is no second copy to drift.

Options

salpa new [NAME]
  -t, --template     individual-node | multi-node-package
  -d, --description  one-line description
      --author       author name
      --category     UI category (e.g. "Cheminformatics")
  -o, --output       directory to create the package in (default: .)
      --hashtags     comma-separated discovery hashtags, 5-10
  -y, --yes          non-interactive: accept defaults, no prompts (and no guided setup)

The guided loop

salpa dev [PATH]
      --status       print the checklist and exit — no prompts, no actions
      --port         port the Salpa app is listening on

salpa dev reads your package, shows where it stands, and lets you press a number instead of remembering a verb. salpa new offers to drop you straight into it.

╭─ solvate-box single ───────────────────────────────────────────────────╮
│  DESCRIBE IT                                                       0/2 │
│ ▸   1 ○ Metadata      description, author, category, hashtags          │
│     2 ○ The science   core.py is still the placeholder                 │
│  BUILD IT                                                          0/2 │
│     3 ○ Dependencies  only python + redis-py                           │
│     4 ○ Environment   not built                                        │
│  PROVE IT                                                          1/2 │
│     5 ✓ Checks        no findings (structure only)                     │
│     6 · Does it run?  needs the environment first                      │
│  SHIP IT                                                           0/1 │
│     7 ? Push to Salpa not checked — pushing finds out                  │
╰────────────────────────────────────────────────────────── 1 of 7 done ─╯
  Metadata — walks the fields; every comment in the file survives.
  ↑↓ step · ←→ phase · enter run · 1-7 jump · r refresh · q quit

The checklist is the menu. ↑ ↓ move between steps, ← → between phases, Enter runs the highlighted one, 1-7 runs one directly. The cursor starts on the first outstanding step and moves on after each one lands, so most of a session is Enter. The line under the box says what the highlighted step will do, which is why the rows can stay short. Grouped by phase because the steps are not equal: describing is one-time, prove it is an inner loop you run many times while editing core.py, shipping is the end. Framed because it sits directly under the last step's output, and without a border the two are the same material.

Without a terminal it reads a line instead and the numbers do everything — it never offers keys that will not work. SALPA_NO_KEYS=1 forces that mode if your terminal or multiplexer misbehaves with the arrow keys. Every step is a command you could type yourself; this is a shorter path through them, not a different one.

The panel is erased before each step and drawn again after, so it appears once, at the bottom, always current — while the output you need to read is printed once and kept.

? is not a pass. Whether the node runs, and whether the app has it, cannot be read off the disk — running them is what finds out. They stay ? until you do, and they go back to ? when you edit your code, because "it passed" is a statement about a package that no longer exists.

Step 1 walks the metadata fields (Enter keeps the current value) and writes meta.toml back one line per field, with every comment intact — the comments in these files are the documentation. The whole result is re-parsed and checked before anything is written, so a value that would break the TOML is refused rather than written on a best effort. That step also catches what salpa validate structurally cannot: a fresh scaffold with author = "TODO" passes validation with zero errors, because its placeholder check only matches literal {{BRACES}} and salpa new fails outright if any survive.

--status (or piped input) prints the checklist once and exits 0. It is a report, not a gate — salpa validate --strict is the gate, and two gates drift.

Dependencies and platforms

salpa add PACKAGE... [--path PATH]
      --spec         version for a SINGLE package; with several, put it on the name
      --keep-platforms   report what cannot run there, but leave `platforms` alone
      --dry-run      say what would change; write nothing
      --json         machine-readable report

You name the packages — nothing can infer biopython from import Bio. Salpa finds each one in the channels your manifest already declares (so the conda-forge-only trap cannot bite), works out which of your declared platforms can run all of them, writes them, narrows platforms with a comment saying why, and solves.

$ salpa add gromacs mdanalysis "biopython>=1.85"
  ok    gromacs 2026.3      conda-forge  ·  5 platform(s)
  ok    mdanalysis 2.10.0   conda-forge  ·  6 platform(s)
  ok    biopython 1.88      conda-forge  ·  7 platform(s)

  ok    linux-64
  ok    osx-64
  ok    osx-arm64
  no    win-64   no build for gromacs
  ...
Solved. gromacs, mdanalysis, biopython and everything already here resolve together
on all 3 platform(s).

Four rules worth knowing:

• Several names, one solve. The platform list becomes the intersection — what can run all of them — because that is what pixi lock insists on. A solve answers "do these resolve together", which is only meaningful once all of them are in the file.
• Nothing is written unless every name resolves. One typo out of four and you see the three that were found, a near miss for the fourth, and Nothing was written. The fix is to run the same command again, which is only safe because the failed run changed nothing.
• One reason comment per dropped platform, naming only the packages responsible for that one — a single "removed by salpa add gromacs mdanalysis" becomes a lie the moment two packages have different gaps. A later run replaces its own note rather than stacking a stale one.
• It never widens. When a package also builds somewhere you do not declare, it says so and stops. Adding a platform is a claim about testing, not only about solving.

Quote an inline version — salpa add "numpy>=1.24" — or your shell reads > as a redirect and creates a file called =1.24 before Salpa ever runs.

Full background, including the four places a dependency can be written and the one that installs anything: salpa docs dependencies-and-platforms.

Validating

salpa validate [PATH] checks a package against the contract the app enforces at install and registration time. Every check maps to a real failure — a node missing from [package.nodes], a shared-environment name that drifted between three files, a stray per-node pixi.toml that silently opts a node out of the shared env. It is not a linter: style is not checked.

salpa validate [PATH]           # default: the current directory
      --strict                  # warnings become errors (CI; except H2 — see below)
      --ignore CODE             # suppress one check by code, repeatable
      --json                    # machine-readable report
      --no-import               # skip the import checks
      --python PATH             # override the auto-detected import-check interpreter

Exit codes: 0 clean · 1 at least one error · 2 the path is not a package.

Findings are ERROR (will not install or will not register) or WARN (installs, but is probably not what you meant). Each carries a stable code (E3, S2, …) so it can be quoted, grepped, and suppressed. salpa docs codes lists every one of them with what it means.

Import checks. I1-I3 import your node.py in a subprocess to confirm the app can read its OPTIONS. They need the authoring SDK (bocoflow_core), and the interpreter that has it is found for you — your package already declares the SDK in its own pixi.toml, so its solved test environment is the right one to judge the package against:

salpa env install    # solve the test env — the one your pixi.toml puts the SDK in
salpa validate       # the import checks now run, no flags

They are skipped by default, and the summary says so. Without the SDK there is no way to import your node.py, so salpa validate reports the structural checks only and marks the count:

0 errors, 3 warnings · import checks DID NOT RUN

Worth taking seriously rather than reading as a pass — I1 is the check that catches a node.py the app cannot load at all, which no amount of structural checking will find.

The search order, first hit wins: .pixi/envs/test → .pixi/envs/default → the interpreter running salpa. A candidate that cannot import bocoflow_core is passed over, and the report names the one that answered. pip install bocoflow-core-sdk beside salpa works too, and --python PATH overrides the search entirely (a --python that lacks the SDK skips rather than falling back — naming an interpreter is an instruction, not a hint).

With no interpreter at all, the checks are skipped, and both the status line and the summary say so — No findings (import checks skipped) is a smaller claim than No findings. This package should install and register., and reads as one.

Suppressing a check. Some warnings describe a shape that is occasionally correct — E3 (a per-node pixi.toml) is a deliberate opt-out when a node's dependencies conflict with its package's shared environment. --ignore lets such a package pass --strict without silencing everything:

salpa validate --strict --ignore E3

Suppressed findings are still reported (and appear under suppressed in --json) — they are hidden from the exit code, not from you.

One code --strict does not promote. H2 reports build artifacts (__pycache__, .pytest_cache), and the documented order is pixi run test then salpa validate — so the test run creates exactly what the next command reports. Promoting that would fail a package for having been tested, which it did: 37 of the 39 first-party packages were rejected, 36 of them on H2 alone. It stays a WARN, still reported and still counted. It is right about a published tree and wrong about a working one, so it is enforced at publish time instead.

salpa new runs a validation pass over its own output and prints a one-line summary. That pass covers structure, naming and environment only; the import checks are left to salpa validate.

Smoke-testing

salpa validate proves a package installs and registers, and every check it makes is static — a node whose execute() raises on its own sample input passes it cleanly. salpa smoke [PATH] closes that gap by running the thing:

salpa smoke [PATH]              # default: the current directory
      --python PATH             # interpreter to run the node with (same resolution
                                # as `validate`; name it explicitly in automation)
      --timeout SECONDS         # per node, default 300 — covers all of its runs
      --json                    # machine-readable report

Exit codes: 0 every node behaved · 1 one did not, or the checks could not run · 2 the path is not a package.

Three checks:

• smoke — import node.py, call execute() on a file from demo_data/, and require success with a non-trivial payload.
• negative — run it again with its real input taken away: a path that does not exist for a node that was given a file, no upstream for a node that reads predecessor_data. Given is the operative word — a node that declares a file parameter but is driven by something else (an accession, a SMILES string) never had a file to lose, so the check reports n/a rather than accusing it. A node that still reports success is doing except: success = True, which is the most common way a broken node looks fine. Raising here is a pass; claiming success is a failure. If a node has neither kind of input, the report says the check did not apply — "not checked" must not look like "checked and fine".
• determinism — the same input twice gives the same payload.
• idempotence — running twice into the same output directory gives the same result. Catches the classic self-input leak: a node that scans its working directory for input and writes its output there picks up its own previous output on a re-run.
• path-independence — the same input read from a different directory still works, and still gives the same content. Catches a hardcoded path.

The last three are advisory — reported, never blocking. A node that samples may legitimately be non-deterministic, and a node that appends may legitimately be non-idempotent. They exist to tell you something a static check cannot.

Each node is executed up to three times — once for smoke, once for negative, and once more for determinism (the smoke run is reused as the first sample rather than running a fourth). All three share one --timeout budget, so raise it for an expensive node.

Multi-node packages are not chained. Every node runs on its own demo_data/ and its own declared parameters, because a package manifest gives its membership and never its connections — and guessing at connections produced both false failures and undetectable false passes.

That is also the bar a node should meet: every input is a parameter, and predecessor_data only fills one in when the canvas left it blank. A node that can only be driven by its upstream cannot be tested on its own, which is a gap in the node. Smoke reports such a node as not checked — never a failure, and never a pass.

Telling it how to run your node

node.py carries a DEMO_CONFIG dict — parameter name to value, with demo_data/… strings resolved relative to the node. The scaffolded node.py test builds its flow_vars from the same dict, so the declaration has exactly one source and the test proves it is right.

Without it, a checker can only infer from a parameter's type — which gives its shape and never its value. It cannot know which of two files is the topology, and a node driven by an ID has no file to infer from. Anything undeclared is still inferred, so it is worth adding incrementally.

What it does not prove

That the node is correct. It proves the node runs, returns something, and refuses bad input. A node whose science is wrong — or absent — passes all three as long as it returns a plausible dict. No checker that reads a package in isolation can do better: that needs an expected answer the package does not contain. Green here means "nothing obviously broken", which is a smaller claim than it reads as.

The development environment

Your node's dependencies live in a pixi environment beside the package. salpa env manages it — it never touches the Salpa app's environments, which are shared between packages and removed by reference count.

salpa env status [PATH]      what is built, how big, and which checks would run now
salpa env install [PATH]     build the test environment (`pixi install -e test`)
salpa env clean [PATH]       remove the package's .pixi/ (keeps pixi.lock)

status answers the question that otherwise has no home — not what exists, but what would happen:

$ salpa env status
my_analyzer (single)
  test     412.6 MB  python 3.12.13  SDK 0.1.1
  default  not built
  total 412.6 MB — `salpa env clean` reclaims it

Right now:
  salpa validate  import checks will RUN (.pixi/envs/test/bin/python)
  salpa smoke     will run with .pixi/envs/test/bin/python
  pixi run test   ready

Why test and not the default environment. The default one deliberately excludes the authoring SDK — it would shadow the runtime bocoflow_core the app provides — so a package built that way cannot run its tests, validate's import checks, or salpa smoke. install always builds test, which carries your dependencies as well.

pixi run test builds that environment on demand, so install is a convenience rather than a requirement. clean is not: a trivial package's environment is ~100 MB and one pulling AmberTools or GROMACS is gigabytes, and nothing else reclaims it.

Pushing it into Salpa

salpa push [PATH] makes a package appear in your running Salpa app's Marketplace ▸ Browse, ready to install.

It is not publishing. It targets the app on your machine; nothing is uploaded anywhere, and there is no public hub involved.

salpa push [PATH]
      --copy           copy into the app instead of linking your directory
      --source-id      id for this source in the app
      --source-name    display name for this source
      --port           port the app is listening on (default 18000, or $SALPA_PORT)
      --force          push despite `salpa validate` errors; with --copy, also
                       replace a package copy that this push did not create
  -y, --yes            accept the layout change without asking

Exit codes: 0 ok · 1 refused or failed · 2 the path is not a package.

push validates, arranges your package into the layout a package source uses (registry.json + packages/<pkg>/, asking before it moves anything), generates the catalog file, and asks the app to sync it and put the package on the shelf.

The install itself stays in the app. push stops at the shelf — installing is user-consented in the UI, and runs the same installer every other package goes through. There is no second installer in this CLI, and no assumption anywhere about where the app keeps its files: the app is asked, over the API its own UI uses. Salpa must therefore be running; if it isn't, push says so and stops.

By default your directory is the source — nothing is copied, so re-running after an edit refreshes what the app offers:

salpa push packages/my_suite      # edit, push, repeat

--copy copies the package into the app instead. Self-contained and survives moving your working copy, at the cost of needing another --copy push per edit. It lands in the directory the app keeps its own packages in, so it will replace an earlier copy of yours but refuses to overwrite anything else — your [package].name decides the directory name, and a collision there is someone else's package.

Taking it back out

salpa unpush [PATH] is the reverse: it removes the package from the app's shelf and deregisters the source, so the app stops offering it.

salpa unpush [PATH]
      --source-id      id of the source to remove (default: whichever the push created)
      --port           port the app is listening on (default 18000, or $SALPA_PORT)
      --force          remove it even though the package is installed
  -y, --yes            remove without asking

Exit codes: 0 ok (including nothing to remove) · 1 refused or failed · 2 the path is not a package.

Your own directory is never touched — the packages/ layout and registry.json that push wrote stay put, and salpa push registers them again whenever you want the package back. You do not have to remember whether you used --copy: unpush reads which mode was used off what is actually in the app, and cleans up the copy in the app's catalog when there is one.

Unpushing something that was never pushed reports Nothing to remove and exits 0, so it is safe to run twice.

unpush refuses while the package is installed, and points you at Marketplace ▸ Installed to uninstall it first — removing the source out from under an installed copy would leave it pointing at a source that no longer exists. --force removes it anyway. As with push stopping at the shelf, unpush never uninstalls anything itself; that stays in the app.

Full walkthrough: salpa docs publishing-to-your-app.

Authoring guide

salpa docs lists the bundled guides; salpa docs <name> prints one. They ship inside the wheel, so they match the version you installed:

Full documentation: https://salpa.app/docs/custom-nodes

Reserved

salpa publish / install / run are reserved for later and not implemented yet. The current surface is salpa new, salpa dev, salpa add, salpa validate, salpa smoke, salpa env, salpa push, salpa unpush and salpa docs.