@crewhaus/tool-text

Deterministic text tools. Every one is pure: no filesystem, no network, no clock, no randomness. The same input always produces the same bytes.

That property is the point. A harness spends a model call when it needs judgement; it should not spend one to count lines, pull a version string out of a log, or fill in a template. These nineteen tools cover the mechanical half so the model is left with the part that actually needs it.

tools:
  - all-text          # every tool below
  - -renderTemplate   # ...except this one

| Tool | What it does | |---|---| | CompactLog | Collapse a long log to distinct lines with repeat counts, failures first | | CountTokens | Characters, words, lines and an estimated token count | | DiffParse | Read a unified diff back into files, hunks and lines, each line numbered in the new file | | EscapeString | Quote text for a regex, shell word, JSON, URL, HTML, markdown, CSV or SQL LIKE | | ExtractEntities | Pull URLs, emails, IPs, UUIDs, semvers, tickets, amounts and more out of text | | ExtractKeywords | Rank a document's significant terms, stop words excluded | | FuzzyMatch | Rank candidates against a query by edit distance, Jaro-Winkler, trigram or token overlap | | GlossaryReplace | Apply a term mapping, longest term first, whole-word by default | | MarkdownOutline | List a document's headings, or return one section by title or slug | | MarkdownTable | Render records as an aligned GitHub-flavoured table | | NormalizeText | Canonicalize line endings, whitespace, Unicode form, ANSI codes, invisible characters | | RegexExtract | Every match with named groups, offsets and line numbers | | RenderTemplate | Fill {{placeholders}} from structured data, strict about missing keys | | RuleClassify | Label text against operator-written phrase and regex rules | | SortLines | Sort, de-duplicate, key by a delimited field | | TextDiff | Unified diff plus added/removed counts | | TextSimilarity | Score two texts 0..1, lexically | | TruncateToBudget | Cut to a character or estimated-token budget, keeping head, tail or both | | WrapText | Hard-wrap at a column width with an optional prefix |

What is deliberately not here

Anything whose core is judgement rather than mechanism: summarizing, translating, classifying by meaning, drafting prose, ranking by relevance beyond lexical scoring. RuleClassify labels by rules an operator wrote, not by understanding. ExtractEntities finds things that have a syntax, not things that have a meaning — it is regex, explicitly not named-entity recognition. TextSimilarity is lexical, with no embeddings.

Where a deterministic first pass exists, it belongs first and the model is the escalation path.

Layout

src/lib/ holds the pure functions and is where the behaviour is tested; src/index.ts wraps them as tools. A bug in levenshtein reads better as a failing unit than as a failing tool call.

Safety flags

All nineteen are readOnly, non-destructive, scope: "internal", and declare no io capability, because none of them crosses a process or network boundary. packages/tool-text/src/index.test.ts asserts that for every tool, so a future addition that reaches outside has to change the assertion deliberately.