Version Packages

[NSchatz]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and publish to npm yourself or setup this action to publish automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@cosyte/cli@0.0.2

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.
Versions and publishing are managed with Changesets;
this file is maintained by hand (Changesets handles the version bump and publish only).

Unreleased

Fixed

• The public surface now states that 0.0.1 is published and uninstallable (ASSETS-P8). 0.0.1
  published on 2026-07-29 with all ten file:vendor/*.tgz dependency specifiers intact. vendor/ is
  not in files and there is no bundledDependencies, so every install route (npm i, npm i -g,
  npx) fails with ENOENT on node_modules/@cosyte/cli/vendor/cosyte-fhir-0.0.0.tgz. Reproduced
  in a clean directory. A published version is immutable (ADR 0001), so 0.0.1 stays broken and the
  fix must ship as a later version.
  • README.md gains a "Known issue" section stating what fails, the exact error, why, and that
    a source checkout is the only workaround. docs-content/installation.md gains the same under
    "Installing fails today". Both copies of the broken npx-based MCP registration snippet are
    annotated (README.md and docs-content/mcp.md).
  • Four false claims corrected. (1) README.md said "not yet published to npm"; it is
    published. (2) It described swapping the vendored sibling deps for real @cosyte/* npm ranges as
    a step still to come "at that flip"; the flip already happened without the swap, which is the
    defect. (3) The same "not yet published" claim appeared in three docs-content/ pages. (4)
    The redact/deid terminal diagnostic, the --help text, three JSDoc blocks that compile into
    dist/*.d.ts, and five docs pages said @cosyte/deid was unpublished, unshipped, or unbuilt;
    it is published at 0.0.2, and the accurate statement is that the CLI does not wire it yet. Every
    count here was re-derived by census of the base tree, not estimated.
  • The "all eight formats" claim was checked and is correct: CosyteFormat and OP_SUPPORT both
    enumerate exactly eight.
  • RELEASING.md records that the documented dependency-swap step was skipped, that a green
    npm publish --dry-run cannot catch this (it packs a tarball but never resolves its deps from a
    registry), and adds a checklist step to install the published version from outside the repo.
• The route to an installable release, recorded in RELEASING.md and verified against the
  registry. @cosyte/hl7 (0.0.3), @cosyte/terminology (0.0.4) and all six breadth parsers
  would swap to real ranges today. @cosyte/fhir is unpublished (FHIR-NPM-NAME, an npm E403
  name-similarity rejection) and @cosyte/transform@0.0.2 fails E404 on its @cosyte/fhir peer,
  so neither can. An installable release is nonetheless reachable before that unblocks, because npm
  tolerates an optionalDependency that fails to resolve (measured). It needs a code change first:
  @cosyte/fhir and @cosyte/transform are imported with a raw await import(), so they must be
  routed through a guarded loader to degrade to CLI_PARSER_UNAVAILABLE (exit 69) rather than
  crash. loadOptional() cannot be reused unchanged: it takes a CosyteFormat, and "transform" is
  not one, and its diagnostic hardcodes the word "parser". Not undertaken here.

Added

• README.md now opens with the shared Cosyte lockup in a <picture> block, above the H1. The
  dark-ground tile (cosyte-lockup-tile-on-dark-1200x300.png) sits behind a
  prefers-color-scheme: dark media query and the light-ground tile
  (cosyte-lockup-tile-on-light-1200x300.png) is the inner <img> fallback, so the mark is read on
  a ground that matches the page it is read on. The # @cosyte/cli heading and the blockquote under
  it are unchanged: the artwork reads "Cosyte" and the heading names the package, so nothing on the
  page is duplicated. Both URLs were re-checked immediately before the push and returned
  200 image/png (10513 bytes dark, 10455 light). The block was copied byte for byte out of
  @cosyte/hl7's README.md rather than retyped and diffed against it, because a transcription
  error in one of these URLs is a broken image on a public package page.

  This replaces the per-package banner added earlier in the same unreleased window, and the reason
  it replaces it is recorded rather than dropped. That banner was a plain markdown image, chosen
  over <img> or <picture> on the stated ground that whether npm's markdown sanitizer preserves a
  <picture> element was unverified. That was an accurate account of what was known when it was
  written, and it has since been measured: GitHub honours the prefers-color-scheme switch
  (observed on @cosyte/astm in dark mode, where the rendered image's currentSrc resolves to the
  on-dark tile and its parent element is PICTURE), and on the npm package page the <img> is
  hoisted out of its <picture> by the anchor wrapper rather than the element being stripped, so
  the light cut renders, which is the correct one there because npmjs.com has no dark mode. Those two
  measurements were taken on astm and reported into this repo rather than re-taken here; what was
  re-checked directly for cli is that both tile URLs return 200 image/png. The failure mode is
  safe either way: a renderer that strips <source> renders the inner <img>, so the worst case is
  a light-ground mark on a dark page, never a missing or broken image.

  Why this is corrected rather than annotated in place. 0.0.1 published on 2026-07-29, a day
  before the banner landed, so no @cosyte/cli tarball has ever carried the banner or the sentence
  announcing it. Annotating an entry no consumer received would publish an addition and its
  replacement as two changes when only one was ever visible, so the banner entry is removed from this
  release and its reasoning is carried here instead of being silently reversed.

  The alt text describes the artwork, a plus mark set in two overlapping rounded squares beside the
  Cosyte wordmark, rather than the package. It is what a screen reader on the npm page reads out and
  what a reader gets when the image fails to load, so repeating the @cosyte/cli heading below it
  would be a wasted line. The wording is the one eight sibling packages already carry, confirmed
  against both rendered PNGs here rather than copied on trust.

• Phase 7: release hardening (the final roadmap phase; the CLI is feature-complete). No new
  runtime command surface: this phase is publish-readiness.

  • Fuzz gate over the two input boundaries. test/fuzz.property.test.ts fuzzes the terminal
    (run, over arbitrary argv vectors + stdin bytes) and the agent surface (dispatchTool, over an
    arbitrary tool name + arguments), asserting neither ever throws an unhandled exception, always
    resolves to a documented exit code, and never leaks a raw stack frame onto a secondary channel. The
    case count scales via CLI_FUZZ_RUNS; a scheduled Fuzz workflow (.github/workflows/fuzz.yml)
    runs it nightly at a high count, and pnpm test:fuzz runs it on demand.
  • Exit-code golden matrix. test/exit-code-matrix.test.ts locks one representative invocation for
    every code in the 0/1/2/65/66/69/70 contract, driven end-to-end through run, so a regression that
    turns an invalid-input exit 1 into a 0 (or renumbers a code) fails CI. The exit-code map and the
    stable CLI_* diagnostic codes are a stability surface: renaming one is a breaking change.
  • Publish dry-run proven. A new smoke gate (scripts/smoke.mjs, wired into verify.sh)
    exercises the built package (the dual ESM/CJS . and ./mcp subpath exports, and both
    cosyte / cosyte-mcp bins under node) and npm publish --dry-run assembles a clean tarball
    (dist + README/LICENSE/CHANGELOG). attw remains a publish gate.
  • Honesty + release docs. docs-content/limitations.md (wraps-not-implements, the non-goals, the
    honest per-(format, operation) support matrix, the PHI-default posture), a man-page-style
    docs-content/reference-commands.md, and RELEASING.md (the one-package-two-bins publish,
    provenance/OIDC, the vendored-file:→npm dep swap, and the two standing founder stops).

• Phase 6: six more formats + streaming + shell completion (ADR 0025). The cosyte CLI now wraps
  all eight cosyte formats, routed through a single lazy per-format adapter registry
  (src/core/parsers.ts) that replaces the old per-command hl7 ? : fhir branches and makes support
  per (format, operation). An unsupported (format, op) is a value-free CLI_FORMAT_UNSUPPORTED,
  never a fake (ADR 0018).

  • New formats and their honest capabilities. x12, astm, ncpdp (SCRIPT) → parse · inspect ·
    fmt · validate; ccda → inspect · fmt (XML re-serialize) · validate (parse deferred, no
    library-blessed JSON model; XML is the canonical form); dicom → inspect · validate (parse/fmt
    deferred. The model is binary); mllp → parse · inspect (a transport container the CLI de-frames
    to its enclosed HL7 message(s)). Content autodetection now covers all eight (conservative + disjoint:
    a leading 0x0B VT byte routes to mllp, ISA→x12, an H-record→astm, <ClinicalDocument>→ccda,
    a <Message> in the NCPDP namespace→ncpdp, DICM@128→dicom); --format accepts mllp.
  • Streaming / multi-message. parse emits NDJSON with per-record isolation for inherently
    multi-record inputs: an MLLP stream (one record per frame) and any input under the new
    --ndjson flag (one record per non-empty line: the FHIR bulk-data convention). A record that
    fails to parse becomes a value-free { record, error } line and the stream continues; the overall
    exit is a data error (65) if any record failed. A single message is unchanged (one pretty, or
    --json compact, envelope + a value-free warning-count note).
  • Shell completion. cosyte completion <bash|zsh|fish> prints a static, value-free completion
    script generated from the command tree.
  • Dependencies. The cap stays 4 (no umbrella edit). The six breadth parsers are vendored
    optionalDependencies, lazy-loaded per format and outside the hard-runtime-dep closure
    (ADR 0025, mirroring the MCP SDK isolation of ADR 0024). An absent optional parser degrades to a
    value-free CLI_PARSER_UNAVAILABLE (exit 69), never a crash. Pinned sibling commits: dicom
    d1ed590, x12 0c60606, ccda 3753216, ncpdp 184eecc, astm 92ac210, mllp aecff75 (all
    v0.0.1 except astm records-layer). Third-party CLI-core runtime deps stay zero.
  • New diagnostic CLI_PARSER_UNAVAILABLE (exit 69); the exit-code contract is otherwise
    unchanged (0/1/2/65/66/69/70). New value-free inspect summaries per format; new programmatic
    exports (OP_SUPPORT, supportsOp, formatsSupporting, parseFormat, inspectFormat, fmtFormat,
    validateFormat, deframeMllp, loadOptional, valueFreeLocator, DETECTABLE_FORMATS,
    completionCommand, and the result/summary types). The public WIRED_FORMATS set is removed in
    favour of the per-op OP_SUPPORT matrix (pre-alpha 0.0.x surface change).

• Phase 5: the cosyte-mcp MCP server (the agent front door). A stdio Model Context Protocol
  server that exposes the shared command core to an LLM/agent as callable tools: the second adapter
  over one core (ADR 0022, 0024). Reachable three ways: the new cosyte-mcp bin, the cosyte mcp
  subcommand, and the @cosyte/cli/mcp subpath export. Tools: parse, validate,
  inspect, and convert, each a thin wrapper that calls the same command handler the terminal
  uses (with --json), so cosyte parse and the MCP parse tool agree by construction; the CLI
  re-implements nothing.

  • PHI posture, inherited and hardened. Every tool runs under the value-free posture. There is
    no --unsafe-show-values door on the agent surface. A tool result carries the requested data
    (the parsed model / converted Bundle: the explicit request); a tool error carries only the value-
    free diagnostic (a stable code + positional context), never an input value. A parsed-but-invalid
    validate verdict is a successful call reporting the verdict, not a tool error; only a hard
    failure (unparseable / no input / usage) sets isError.
  • The MCP SDK is isolated and runtime-optional (ADR 0024). @modelcontextprotocol/sdk (the CLI's
    first and only third-party runtime dependency) is declared in optionalDependencies (pinned
    1.29.0) and imported only in src/mcp/server.ts, reachable solely via the ./mcp boundary
    (the subpath, the cosyte-mcp bin, and a dynamic import() on the cosyte mcp branch). A cosyte parse invocation never loads it; the core works with the SDK absent (--omit=optional). Because it
    is not part of the hard runtime closure, the umbrella verify-policy.json cap on cli runtime
    dependencies stays 4: unchanged. A static isolation test proves no core/commands module
    imports the SDK.
  • New subpath export @cosyte/cli/mcp and new cosyte-mcp bin; new programmatic exports
    (createMcpServer, startStdioServer, dispatchTool, TOOL_DEFS, and the MCP result types) on the
    ./mcp subpath. redact/deid (gated on @cosyte/deid) and map-codes are deliberately not
    exposed as tools yet.

• Phase 4: convert / map-codes (the consumer-of-consumers commands). Two commands that wrap
  the higher-layer libraries; the CLI adds no mapping or terminology logic of its own.

  • convert <file|-> --to fhir [--json] [--quiet]: HL7 v2 → FHIR R4 via
    @cosyte/transform. Parses the input with @cosyte/hl7, hands the parsed message to
    transform.toFhir, and emits the serialized FHIR message Bundle (the library's canonical
    serialization) on stdout: cosyte convert equals transform's programmatic output. The
    conversion's value-free issues (a stable code + a v2-index → FHIRPath locator, never a field value)
    render on stderr (or as a JSON envelope under --json); --quiet suppresses them. The
    load-bearing rule mirrors validate: an error-severity transform issue drives exit 1,
    never 0. --to fhir is required (the only target); a non-HL7 source (e.g. a FHIR document) is
    a value-free CLI_FORMAT_UNSUPPORTED data error (65), never a fake conversion; an unparseable
    HL7 input is CLI_PARSE_FAILED (65).
  • map-codes <conceptmap|-> --code <code> [--system <uri>] [--version] [--display] [--json] [--quiet]: translate a single source coding through a BYO FHIR R4 ConceptMap via
    @cosyte/terminology ($translate). The positional is the ConceptMap document; the source
    coding is named by flags. A ConceptMap and a code are reference data, not PHI, so the
    translation result goes to stdout: a match → the target coding(s) + exit 0; an
    unmapped code → the never-fabricate TERM_TRANSLATE_UNMAPPED signal + exit 1. A map that
    is not valid JSON or not a loadable ConceptMap is the new value-free CLI_MAP_INVALID data
    error (65), surfacing the stable terminology-loader code (e.g. TERM_CONCEPTMAP_MALFORMED),
    never the map's bytes.
  • New CLI_MAP_INVALID diagnostic code. New programmatic exports: convertCommand,
    convertOutcome, mapCodesCommand. New runtime dependencies (ADR 0023): @cosyte/transform
    (e6c4531, v0.0.0) and @cosyte/terminology (e5ed368, v0.0.1) as hard, first-party,
    lazy-loaded deps: vendored as pnpm pack tarballs under vendor/ until PUB-FLIP
    (pnpm vendor:refresh; umbrella ADR 0008). The umbrella verify-policy.json cap on cli runtime
    deps was raised 2 → 4; third-party CLI-core runtime deps stay zero (both siblings are
    lazy-loaded per command, so the parse fast path never loads them).
  • ADR 0023: wire @cosyte/transform + @cosyte/terminology; the deliberate 2 → 4 cap raise
    (amends ADR 0021).

• Phase 3: validate / inspect / fmt. Three commands over the two wired parsers
  (HL7 v2 + FHIR R4), each a thin wrapper that re-implements no library logic.

  • validate <file|-> [--profile] [--json] [--quiet]: parse + run the wrapped parser's own
    validation surface, with the verdict in the exit code: 0 valid, 1 invalid (parseable
    but non-conformant), 65 unparseable, 66 no input, 2 usage. The load-bearing rule: a
    validation failure is never exit 0; "unparseable" (65) is a distinct signal from "parsed
    but invalid" (1). Findings are value-free: a stable code, a severity, and a positional
    locator (a FHIRPath, or an HL7 segment/field index), on stderr by default, or as value-free JSON
    on stdout under --json; --quiet makes the exit code the whole signal. The CLI invents no
    verdict: FHIR validity is @cosyte/fhir's validateResource().valid (plus any error-severity
    read issue); HL7 validity is "parseable" (its warnings are non-fatal by the library's design:
    surfaced, never failing). --profile is gated to an honest CLI_NOT_IMPLEMENTED (exit 69):
    the CLI bundles no profiles yet, so it never fakes or silently drops a profile verdict.
  • inspect <file|-> [--json]: a value-free structural summary: HL7 message type, version,
    per-segment-type counts, and a warning count; FHIR resourceType, Bundle entry counts by type, and
    a read-issue count. Counts and structural type codes only, never a field value.
  • fmt <file|->: canonical re-serialization via the wrapped library's spec-clean
    serializer (Hl7Message.toString() / serializeResource); its stdout is the data channel. An
    unparseable input is a data error (65) with no partial emit.
  • New EXIT.INVALID (1): the validate verdict code (the exit-code contract is now
    0/1/2/65/66/69/70). All four commands share one input + format front door (core/resolveInput)
    and one value-free parser-failure boundary (core/wrap), so the value-free-by-default posture and
    the --unsafe-show-values chokepoint apply uniformly; parse was refactored onto the shared
    helpers (behavior-preserving). New programmatic exports: validateCommand, inspectCommand,
    fmtCommand, resolveInput, parseFailureResult, formatHl7Position, errorResult. No new
    runtime dependencies: stays within the cap of 2.

• Phase 2: PHI posture hardened + redact/deid + --unsafe-show-values.

  • --unsafe-show-values: a global, opt-in, PHI-exposing flag, resolved once and order-
    independently and funnelled through a single chokepoint (core/phi.ts), so the "a value
    reaches a secondary surface iff the flag is set" property holds in one place. Off by default;
    with it set, a CLI_PARSE_FAILED diagnostic appends a bounded, single-line excerpt of the
    offending input. Every other surface stays value-free, and a successful parse still puts values
    only on the stdout data channel.
  • redact / deid (<file|->, --format): the de-identification command, shipped as an
    honest, typed CLI_NOT_IMPLEMENTED (new exit code 69, EX_UNAVAILABLE). It is gated on
    @cosyte/deid (unpublished, DEID-1 in flight), never reads the input, and never emits a
    partial Safe-Harbor scrub presented as de-identified. A built-in redactor is deliberately
    withheld: a partial scrub over only the obvious PHI loci would leave PHI behind and present a
    false-safety impression (the cardinal hazard). It delegates to @cosyte/deid via a documented
    seam (core/deid.ts) when that library ships and is vetted.
  • Never a PHI temp file / never a file log: proven by test (no command creates a file in the
    working directory) and by design (commands return a RunResult; only the thin bin writes to
    process streams).
  • New CLI_NOT_IMPLEMENTED diagnostic code and EXIT.UNAVAILABLE (69); new programmatic exports
    (PhiPosture, VALUE_FREE/SHOW_VALUES, extractPhiPosture, unsafeInputSuffix, deidStatus,
    redactCommand).

• Phase 1: the cosyte parse foundation. Reshaped the scaffold from a library skeleton into a
  bin package: package.json#bin maps cosyte → dist/bin/cosyte.mjs (a shebang entry over a
  testable core), argument-parsed with Node's built-in util.parseArgs + a hand-rolled subcommand
  dispatcher (no third-party CLI framework).

• cosyte parse <file|->: reads a file argument or stdin (-); autodetects the format by
  content (HL7 v2 MSH framing, FHIR JSON resourceType): conservative and fail-safe (a confident
  single match parses; ambiguity/no-match is a data error asking for --format, never a guess); routes
  to the wrapped parser (lazy-loaded per format); emits the parsed model as typed JSON on
  stdout. Flags: --format, --json, --quiet, --no-color.

• The exit-code contract (sysexits.h-grounded, documented, tested): 0 success · 2 usage ·
  65 data/parse error (EX_DATAERR) · 66 no input (EX_NOINPUT) · 70 internal (EX_SOFTWARE).
  The CLI never exits 0 on input it could not handle.

• Value-free diagnostic channel with stable CLI_* codes (CLI_FORMAT_UNDETECTED,
  CLI_FORMAT_AMBIGUOUS, CLI_FORMAT_UNSUPPORTED, CLI_NO_INPUT, CLI_EMPTY_INPUT,
  CLI_PARSE_FAILED, CLI_USAGE, CLI_INTERNAL). stdout is the data channel; every stderr line is
  value-free: code + position only, never a field value. No temp files, no file logging.

• Programmatic core API (the . subpath): detectFormat / classifyCandidates /
  detectionError, EXIT, CLI_CODES / CliError, run, parseCommand, VERSION.

• Runtime dependencies (ADR 0021): @cosyte/hl7 (46d50eb, v0.0.1) and @cosyte/fhir (7a099b2,
  v0.0.0) as hard, first-party deps (an npx bin cannot peer-depend) vendored as pnpm pack
  tarballs under vendor/ until PUB-FLIP (pnpm vendor:refresh; umbrella ADR 0008). Capped at 2
  by the umbrella verify-policy.json; third-party CLI-core runtime deps stay zero.

• ADRs: 0021 (developer-tooling tier is a bin that hard-depends on first-party siblings;
  third-party runtime deps minimized) and 0022 (one repo, two bins: the CLI and the future
  cosyte-mcp MCP server over one core; the web playground is out of scope).

Changed

• The documented per-(format, operation) support matrix now states its negatives in words. In
  docs-content/limitations.md the table used a bare U+2014 as a cell value meaning "not
  supported". The em-dash sweep below rewrote it as punctuation, so dicom fmt, mllp fmt and
  mllp validate rendered as a stray colon: "support absent" silently became "support unstated",
  on the page whose whole job is honest capability disclosure, in a form a reader would read as a
  rendering artifact rather than as a claim. Caught by the conformance refuter, not by CI, and nothing
  in this repo's CI could have caught it (test/docs-content.test.ts only executes runnable ts
  blocks, and Prettier's glob does not cover docs-content/). Those three cells now read
  not supported. The support facts are unchanged; the page now says them. Before sweeping any
  repo for this character, find the places it is a value rather than punctuation and convert those by
  hand, to a word, first.

• No cosyte surface in this repo uses an em dash any more, and a CI gate keeps it that way. The
  brand rule (knowledgebase/06-brand/voice-and-tone.md, "No em dashes. Ever.") bans U+2014
  outright and names commit messages explicitly. Measured byte-level over all 124 tracked files,
  not over markdown alone: 659 occurrences across 87 files, all as the literal character and none
  in an encoded form. 61 of the 87 are not markdown (26 are), and one is package.json, whose
  description is published to npm and rendered on the package page. Every occurrence was rewritten
  in place with a period, a colon, a comma, or parentheses, never re-encoded. Consumer-visible text
  that changed wording: the npm description, README.md, docs-content/, cosyte --help, the
  generated bash/zsh/fish completion scripts, and the redact/deid and reserved --profile
  unavailability messages. No stable code, exit code, flag name, or JSON field changed, so nothing
  branching on an exit code or parsing --json is affected. Commit subjects already on main are
  left alone: the message half runs on pull_request only, and history is not rewritten.

  • The gate. scripts/check-no-emdash.sh (pnpm check:no-emdash) plus
    .github/workflows/no-emdash.yml scan both halves the rule covers: every tracked file, and the
    pull request title, body, and commit messages, on the non-default edited trigger so retitling a
    pull request re-checks it. The new no-emdash check-run context is required by the
    ci-required-checks ruleset.
  • Binary handling, which is why this repo does not run the text-only variant its siblings run.
    cli tracks eleven binaries (the ten vendor/cosyte-*.tgz packed sibling dependencies and
    test/__fixtures__/sample.dcm), so the script partitions on an explicit NUL-byte rule rather than
    on grep -I's heuristic, which would silently skip a genuine text file with a broken encoding.
    The reason is measured rather than precautionary: vendor/cosyte-hl7-0.0.0.tgz already
    contains the byte sequence E2 80 94 by compression coincidence, so a text-only scan reds on it
    today, naming a compressed byte stream nobody wrote and offering no way to fix it.
  • Disclosed limits, in the script header rather than left implicit. A tracked text file
    holding a NUL byte is exempt and its em dash would be missed (there is none today; the excluded
    count on the OK line is the tell). Encoded-form matching is literal, so lowercase %e2%80%94, a
    capital-X —, and a semicolon-less &#x2014 pass. The scan reads file contents, never
    file names.

• The published type declarations no longer carry internal project bookkeeping. dist/index.d.ts
  and dist/mcp.d.ts (and their .d.cts twins) are compiled from the JSDoc on every exported symbol,
  and that JSDoc was citing item identifiers (CLI-6), ADR numbers (ADR 0018, 0021, 0022,
  0025), the meta-repo roadmap in prose and by path (cli roadmap §7,
  operations/roadmaps/cli.md), and phase language (§Phase 4, Phase-5, "in a later phase"). All
  of it rendered on hover for anyone who installed the package. Measured on 62fba77 with the rule
  set that ships alongside: 61 occurrences across 23 tracked src/ files, producing 56 across the
  three declaration files a consumer receives (43 in dist/index.d.ts, 12 in dist/mcp.d.ts and 1
  in the shared dist/io-*.d.ts chunk both entries import), each mirrored in its .d.cts twin. Now
  0 on every one of them. (The shared chunk is easy to miss and was missed once here: a count that
  scans only the named entry points under-reports.) Every removal is a cut, not a rewrite: the
  surrounding guarantees (value-free stderr, the never-a-fake CLI_FORMAT_UNSUPPORTED, the gated
  redact, the exit-code contract) are worded exactly as strongly as before. Internal traceability
  stays where the convention puts it: this file, the changesets, the commits and the roadmap.

• A dependency-budget figure left src/core/deid.ts too. The module said wiring @cosyte/deid
  "would breach the CLI's runtime-dep cap (2)". Only the number was stale: the cap has been 4 since
  convert/map-codes landed, and the constraint itself still holds, because the package declares
  exactly 4 hard runtime dependencies and a fifth would indeed breach it. The clause went because a
  dependency-budget figure is internal bookkeeping a consumer cannot act on, not because the
  constraint lapsed. The reasons that actually ground the refusal to ship a built-in redactor are
  untouched: @cosyte/deid is unpublished, the wrapped parsers expose no de-identification API, and
  a partial scrub would present a false-safety impression.

• CosyteFormat's documentation no longer understates autodetection. It read as though content
  detection recognised only HL7 v2 and FHIR, with the other six accepted by --format but "not yet
  wired", which has not been true since all eight formats gained signatures. The stale sentence is
  removed rather than restated.

• A gate now enforces the public-surface rule, which is why the class stops regrowing.
  scripts/check-no-internal-refs.sh (pnpm check:no-internal-refs, on the verify.sh cli ladder)
  plus .github/workflows/no-internal-refs.yml port the shape of hl7's gate
  (hl7#62, hl7#64)
  and ncpdp's (ncpdp#36) rather than the file. Four
  passes: the public markdown surface line by line and paragraph-joined, the npm metadata, src/
  doc comments, and src/ string literals (the pass that would have caught the two identifiers this
  package printed to a user's terminal). Seven rules. Re-derived for this repo: the scan surface, a
  standards-designation exclusion list covering all eight formats (this is the package where
  HL7-V2, FHIR-R4, DICOM-SR, NCPDP-SCRIPT, X12-837P, CCDA-R2.1, MSH-2, NM1-03,
  439-E4 and ICD-10-CM are live at once, so the WORD-N trap is at its widest), and a seventh
  rule no sibling has, a prose roadmap citation, which was 30 of the 61 and which hl7's
  path-keyed rule cannot see. Both self-test halves run on every invocation: positive samples prove
  each rule still matches, negative samples prove none has been widened into the WORD-N shape that
  would delete the reference material the CLI's docs exist to provide. CHANGELOG.md is excluded on
  purpose, as it is in hl7 and ncpdp: it ships inside the npm tarball, yet the convention names it
  as one of the places identifiers belong. That contradiction is ecosystem-wide, and it is recorded
  here rather than settled by one repo.

• redact/deid no longer names an internal tracking identifier on any consumer surface. The
  CLI_NOT_IMPLEMENTED text printed when de-identification is unavailable carried an internal work
  item that means nothing to anyone running the command, and the same identifier reached the
  published type declarations (dist/index.d.ts / dist/index.d.cts), where it surfaces in editor
  tooltips. Both now state only the consumer-observable fact: the command delegates to
  @cosyte/deid, which is unpublished, and the CLI ships no built-in redactor because a partial
  scrub would present a false-safety impression. The stable CLI_NOT_IMPLEMENTED code and the exit
  69 are unchanged.

• CLI_PARSER_UNAVAILABLE's message no longer cites an internal decision record. The error
  raised when an optional per-format parser is not installed pointed at an ADR number, which is
  meaningless to a caller; it now just names the package to install. The stable
  CLI_PARSER_UNAVAILABLE code and the exit 69 are unchanged. The same sweep removed the
  remaining roadmap-phase language from README.md and docs-content/troubleshooting.md.

• The CI checks are now binding on main. ci / verify (22, ubuntu-latest),
  ci / verify (24, ubuntu-latest), ci / actionlint and codeql / analyze (javascript-typescript)
  are required status checks, each pinned to the github-actions app, alongside branch deletion and
  force-push protection. They were advisory before: a red check could not stop a merge. Dependabot
  now watches the npm and GitHub Actions dependency surfaces weekly, which nothing did previously.

• Reshaped the package from the parser-library scaffold to a bin package. Removed the archetype
  stubs (parseCli, WARNING_CODES, FATAL_CODES); replaced the library src/index.ts and the
  round-trip property test with the command tree, the programmatic core API, and command-contract /
  autodetection / PHI-leak / equivalence tests. Rewrote docs-content/ and README.md for the CLI.

Deprecated

Removed

Fixed

• README + guides now describe the shipped Phase-3 command surface. The README.md and
  docs-content/guides-overview.md "Status" blurbs read as a Phase-1-forward roadmap ("Phase 1 ships
  parse…"); they now state the current surface directly (parse / validate / inspect / fmt
  and the gated redact/deid) over the two wired parsers (HL7 v2 + FHIR R4). The pre-alpha,
  not-yet-published-to-npm status is unchanged (accurate), and the npx/npm install -g examples now
  carry a "not on npm yet" caveat (docs-only; README-ORG-SWEEP).
• phi-scan now scans the real fixture directory. The scanner's fixture root pointed at a
  nonexistent test/fixtures/; it now walks test/__fixtures__/ (and the same path in the staged
  filter), so the PHI commit-gate actually covers the CLI's synthetic fixtures.

Security