Skip to content

Wingbox 0.14.0

Choose a tag to compare

@it-baron it-baron released this 19 Nov 11:11
728321a

Two changes need your attention when you upgrade.

  • The query commands (impact, usages, advice, align, place, compare) print S-expressions by default. --human prints the text they printed before, --json prints JSON for code, and --sexp prints exactly what 0.13.0 printed with it.
  • MCP clients: three tools have new names, build is graph, module is modules, query is facts. The wingbox command keeps its command names.

Changes

  • The block that wingbox install writes into the assistant's instructions file says more. It now
    tells the assistant to ask the graph before grep, rg, find or opening files and to read
    only the files and lines the graph names, with text search as the fallback for a keyword that is
    not a symbol or a graph older than the files; and that tests follow the impact analysis: the
    dependents wingbox impact names decide which tests to write and run, and for every changed
    function a test that reaches it is shown (wingbox diff_functions, wingbox usages). Run
    wingbox install or wingbox update once to get the new block.
  • Three MCP tools are renamed, so that all eleven are nouns named after their answer: build is
    graph, module is modules, query is facts. An MCP client that calls one of the three by
    its old name must use the new one
    ; the other eight tools and every wingbox command keep their
    names.
  • The description of every MCP tool is rewritten for an agent that reads nothing else: what the tool
    does and what comes back, when to use it, which tool to use for the neighbouring question, that a
    long list stops at 200 items unless all is true, and what an unknown name returns. definitions
    now says that name may be a pattern (parse_.*).
  • The default output of the commands that answer a question changed. wingbox impact, usages,
    advice, align, place and compare printed text for people; without an option they now
    print S-expressions, the compact form a coding agent reads. --human gives the old text, byte
    for byte as 0.13.0 printed it, with its list limits (--all lifts them). --json prints the
    same answer as one JSON object for code. The other query commands (analyze, quick, module,
    deps, rdeps, layers, boundaries, entity, search, the find_* commands and the rest)
    printed fact lines before and print the same lines now; they take --human and --json as
    well, and their default answer is complete: analyze and quick no longer stop at 200
    dependents unless --human is given. Two form options at once are an invalid argument.
  • impact takes several files in one call: wingbox impact a.py b.py 2 (a last argument that is a
    whole number is the depth). The answer opens with (graph :built <UTC time> :files <n> :stale <n>), which says when the graph was written, how many files it holds and how many changed
    files it is known to miss (0 after the refresh every query starts; the number of changed files
    with WINGBOX_AUTO_REFRESH=0), then one answer per file in the order given. A file that is not
    in the graph gives (error no-match <file> ...) in its place, the other files are answered, and
    the exit status is 1.
  • The default form of impact names each thing once: (impact <file> :direct N :total M (1 <file> ...) (2 (:through <file> <file> ...)) (read-first <file> ...)), with file paths, the rows
    grouped by depth and by the file they are reached through, and the reason of each row (:via
    for a relation that is not an import, :more N, the name lists of the edge). The default form
    of usages is (usages <Name> <kind> :at <file>:<line> :uses U :files M (<file> <kind> <line>:<col> ...) ...). Both are complete: nothing is cut and --all changes nothing there.
    impact --json prints {"graph": {...}, "impact": [{"file", "direct", "total", "rows", "read_first", "advice"}, ...]} and usages --json {"usages": {..., "rows": [...]}}; every
    other command prints {"facts": [...]}, one object per fact line.
  • Code for Claude Code's Code Mode: the skill carries references/code-mode.md, one JavaScript
    helper (wingbox(...args), the command with --json through $`...`) and a recipe per
    question: what depends on several files, every use of a name, where a feature lives, what
    changed since a commit with the tests that use it, the architecture rules. The block that
    wingbox install writes names the call and the file.
  • A --json call that fails as a whole says so in one place: the document starts with "error": {"code", "message"} on standard output, for a folder without a graph (no-graph), an argument
    that is not valid (invalid-argument), a name nothing matches (no-match) and a git reference
    that does not exist (git-diff-failed). Code checks answer.error and nothing else. A file of
    an impact call that is not in the graph carries its own error and the other files are
    answered.
  • --sexp stays where 0.13 had it and prints exactly what 0.13.0 printed with it: for impact
    and usages the flat facts, one per line (impact-root, (impact <root> <dependent> :lang .. :via .. :depth N ...), impact-summary, read-first; usages-target, usage,
    usages-summary), with module IDs, the list limits unless --all, and exit status 0 for a file
    that is not in the graph; for advice, align, place and compare their facts. wingbox module <id>, wingbox boundaries and wingbox boundaries --strict print the same lines with
    the same exit statuses. A script written for 0.13 that passes --sexp (the pull request check
    does) reads what it read; a script that parsed the text of impact or usages without an
    option adds --human.
  • MCP: the impact tool takes target (one file) or targets (a list) and answers in the
    default S-expression form, complete, with the graph fact first when several files are asked;
    it answered in text before, cut at 200 dependents unless all. all is still accepted there
    and changes nothing. The other ten tools answer as before.
  • The skill has a short section "Use from code" (one call for several files, read as JSON), and
    the always-on block of wingbox install says the same in one line; wingbox update refreshes
    both. The skill no longer tells an agent to pass --sexp to a command that answers in facts
    by default.

Fixes

  • wingbox query "(help)", and so the MCP tool facts with the expression (help), lists the
    functions of the query language, as the tool's description says. It answered "unknown function:
    help" outside the interactive wingbox repl.

Upgrade: wingbox update, or pipx upgrade wingbox / uv tool upgrade wingbox / pip install -U wingbox.