@voxgig/sdkgen 1.5.1 → 2.0.0

package/project/.sdk/src/cmp/py-data/Main_py-data.ts

@@ -0,0 +1,987 @@
+
+import {
+  cmp, each,
+  File, Content, Copy, Folder,
+  entityCollection, entityOps, entityIdField, canonToDtype,
+  collectDeps, pkgDescription, keywords, repoInfo, envName, apiName,
+  packageName, registryState, serverVariables,
+  SdkGenError,
+} from '@voxgig/sdkgen'
+
+import {
+  KIT,
+} from '@voxgig/apidef'
+
+
+// The `py-data` target: a Python package for DATA ANALYSTS working in
+// notebooks (Colab, Jupyter), layered on the sibling `py` SDK generated into
+// the same repo. It is a consumer target in the go-cli / go-mcp mould — every
+// standard generation phase is switched off in model/target/py-data.aontu and
+// this component emits the whole package.
+//
+// The deliverable an analyst wants is not "a client", it is TABLES: one line
+// from API to a typed pandas DataFrame, discovered by tab-completion rather
+// than by reading reference docs.
+//
+// Layered, NOT forked from `py`: the dependency arrow points one way
+// (py-data -> py) and `py` must never learn this target exists. A fork would
+// duplicate the entity/feature machinery and drift, which is the known
+// failure mode.
+
+
+// Accessors are generated only for entities exposing these ops. `list` gives
+// a DataFrame; `load` gives a Series. Everything else (create/update/remove)
+// is a write path and deliberately out of scope — see the README's
+// "What this package does not do".
+const FRAME_OP = 'list'
+const SERIES_OP = 'load'
+
+
+const Main = cmp(function Main(props: any) {
+  const { target, ctx$ } = props
+  const { model, log } = ctx$
+
+  // HARD REQUIREMENT: this package imports the sibling py SDK. Generating it
+  // without `py` produces a package whose every import fails at runtime, so
+  // fail at GENERATE time with an actionable message instead.
+  const targets = model.main[KIT].target || {}
+  if (null == targets.py) {
+    throw new SdkGenError(
+      'py-data requires the `py` target in the same SDK: it imports the ' +
+      'Python SDK that `py` generates. Add it with:\n' +
+      '  npm run add-target py\n' +
+      'then regenerate.')
+  }
+
+  const name = model.const.Name              // PascalCase, e.g. Univec
+  const lower = String(name).toLowerCase()   // univec
+  const ENV = envName(model)                 // UNIVEC
+  const sdkModule = `${lower}_sdk`           // univec_sdk
+  const sdkClass = `${name}SDK`              // UnivecSDK
+  const dataModule = `${lower}_data`         // univec_data
+
+  // Multi-tenant APIs (a server URL carrying an unresolved {variable}, e.g.
+  // https://{instance}.example.com/api) have no single host, so base_url
+  // becomes REQUIRED rather than optional. Guessing a host produces confusing
+  // 404s, so `data()` refuses instead.
+  const svars = serverVariables(model)
+  const needsBase = svars.some((v: any) => v.required)
+
+  const ents = each(entityCollection(model))
+    .filter((e: any) => e && null != e.name && false !== e.active)
+
+  const frameEnts = ents.filter((e: any) => entityOps(e).includes(FRAME_OP))
+  const seriesEnts = ents.filter((e: any) =>
+    entityOps(e).includes(SERIES_OP) && null != entityIdField(e))
+
+  log.info({
+    point: 'py-data-entities', target: target.name,
+    note: `frames:${frameEnts.length} series:${seriesEnts.length} of ${ents.length}`,
+  })
+
+  // Root-level statics: Makefile, LICENSE, the template-side tests.
+  Copy({
+    from: 'tm/' + target.name,
+    exclude: [/src\//, /pkg\//],
+    replace: { ...ctx$.stdrep },
+  })
+
+  // The runtime modules live INSIDE the package, not at the top level.
+  //
+  // They were originally emitted flat as auth.py / fetch.py / frames.py, which
+  // is actively dangerous here: `auth`, `fetch`, `frames`, `core`, `entity`
+  // and `utility` are all real PyPI distributions AND exactly the names an
+  // analyst leaves lying around next to a notebook. In Colab the working
+  // directory is on sys.path, so one stray auth.py shadowed ours and `data()`
+  // died with `module 'auth' has no attribute 'resolve'`. A package keeps
+  // every internal module behind the distinctive `<name>_data.` prefix.
+  //
+  // The folder name is model-derived, so it comes from the component; the file
+  // contents are the same for every API, so they stay in tm/py-data/pkg.
+  Folder({ name: dataModule }, () => {
+    Copy({
+      from: 'tm/' + target.name + '/pkg',
+      replace: { ...ctx$.stdrep },
+    })
+    DataClient({ model, name, lower, ENV, sdkModule, sdkClass, dataModule, needsBase })
+    EntityFrames({ model, frameEnts, seriesEnts, lower, dataModule })
+  })
+
+  Gitignore({})
+  Package({ target, model, lower, dataModule })
+  EntityFramesTest({ frameEnts, seriesEnts, lower, sdkModule, sdkClass, dataModule })
+  Readme({ model, name, lower, ENV, dataModule, frameEnts, seriesEnts, needsBase, svars })
+  AgentGuide({ model, name, lower, ENV, dataModule, frameEnts, seriesEnts, needsBase })
+  Quickstart({ model, name, lower, ENV, dataModule, frameEnts, needsBase })
+})
+
+
+const Gitignore = cmp(function Gitignore(_props: any) {
+  File({ name: '.gitignore' }, () => Content(`__pycache__/
+*.py[cod]
+/dist/
+/build/
+*.egg-info/
+.pytest_cache/
+.ipynb_checkpoints/
+`))
+})
+
+
+const Package = cmp(function Package(props: any) {
+  const { target, model, lower, dataModule } = props
+  const ctx$ = props.ctx$
+
+  const ns = model.origin || 'voxgig-sdk'
+  const pkgBase = ns.endsWith('-sdk') ? model.name : `${model.name}-sdk`
+  const distName = `${ns}-${pkgBase}-data`
+  const sdkDist = packageName(model, 'py')
+  const { repoUrl, issuesUrl } = repoInfo(model)
+  const kw = keywords(model).concat(['pandas', 'dataframe', 'notebook', 'colab'])
+    .map((k) => `"${k}"`).join(', ')
+
+  File({ name: 'pyproject.toml' }, () => {
+    Content(`[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "${distName}"
+version = "0.0.1"
+description = "${pkgDescription(model, 'py-data')}"
+readme = "README.md"
+license = "MIT"
+requires-python = ">=3.9"
+keywords = [${kw}]
+dependencies = [
+    "${sdkDist}",
+`)
+
+    // pandas (and anything a feature adds) from the target model. The sibling
+    // SDK dep is emitted above rather than declared in the aontu, because its
+    // distribution name is derived from the model, not fixed.
+    const seen = new Set<string>([sdkDist])
+    for (const d of collectDeps(model, target.name, target.deps, ctx$.log)) {
+      if (seen.has(d.name)) continue
+      seen.add(d.name)
+      const v = d.source === 'target' ? (d.version || '0.0') : d.version
+      Content(`    "${d.name}>=${v}",
+`)
+    }
+
+    Content(`]
+
+[project.urls]
+Homepage = "${repoUrl}"
+Repository = "${repoUrl}"
+Issues = "${issuesUrl}"
+
+# A single package, deliberately. Emitting auth/fetch/frames as top-level
+# modules would put names that are both real PyPI distributions and common
+# notebook scratch files onto sys.path, where Colab's working directory wins.
+[tool.setuptools.packages.find]
+include = ["${dataModule}*"]
+`)
+  })
+})
+
+
+// The DataClient module: `data()` plus the wrapper that owns credential
+// discovery and exposes the raw SDK as an escape hatch.
+const DataClient = cmp(function DataClient(props: any) {
+  const { name, ENV, sdkModule, sdkClass, dataModule, needsBase } = props
+
+  const baseCheck = needsBase
+    ? `
+        if not base_url:
+            raise ValueError(auth.missing_base_message("${ENV}"))
+`
+    : ''
+
+  File({ name: '__init__.py' }, () => Content(`# ${name} Data — notebook-friendly access to the ${name} API.
+#
+# GENERATED by @voxgig/sdkgen (py-data target). Edits are overwritten.
+
+from __future__ import annotations
+
+from typing import Any
+
+import pandas as pd
+
+from . import auth
+from .entity_frames import EntityFrames
+from .frames import records_df
+
+# The sibling SDK is a separate distribution, so this import stays absolute.
+from ${sdkModule} import ${sdkClass}
+
+
+class ${name}Data(EntityFrames):
+    """${name} API as pandas DataFrames.
+
+    Wraps — never subclasses — the ${sdkClass} client. The wrapped client is
+    available as \`.sdk\` for the cases this package deliberately does not
+    cover (writes, non-entity endpoints, per-call control).
+    """
+
+    def __init__(self, sdk: ${sdkClass}, base_url: str | None = None):
+        self.sdk = sdk
+        self.base_url = base_url
+
+    # ---- escape hatches ---------------------------------------------------
+
+    def frame(self, result: Any, **kwargs) -> pd.DataFrame:
+        """Shape ANY op result into a DataFrame.
+
+        For endpoints without a generated accessor:
+
+            rows = ad.sdk.SomeEntity().list()
+            df = ad.frame(rows)
+        """
+        return records_df(result, **kwargs)
+
+    def entities(self) -> list:
+        """The entity accessors available on this object."""
+        return sorted(self._ACCESSORS)
+
+    def _repr_html_(self) -> str:
+        accessors = "".join(
+            f"<li><code>{a}()</code></li>" for a in sorted(self._ACCESSORS))
+        base = self.base_url or "(default)"
+        return (
+            f"<b>${name}Data</b><br>"
+            f"<small>base: <code>{base}</code></small>"
+            f"<ul>{accessors}</ul>"
+        )
+
+    def __repr__(self) -> str:
+        return (f"<${name}Data base={self.base_url or '(default)'} "
+                f"accessors={len(self._ACCESSORS)}>")
+
+
+def data(
+    token: str | None = None,
+    base_url: str | None = None,
+    **options,
+) -> ${name}Data:
+    """Build a ${name}Data client, finding credentials automatically.
+
+    Discovery order (first hit wins):
+      1. the \`token\` / \`base_url\` arguments
+      2. Colab secrets  — ${ENV}_APIKEY / ${ENV}_BASE
+      3. environment    — ${ENV}_APIKEY / ${ENV}_BASE
+
+    Extra keyword arguments are passed straight through to the SDK client, so
+    anything the SDK accepts (feature options, headers, timeouts) works here.
+    """
+    token, base_url, _source = auth.resolve("${ENV}", token, base_url)
+
+    if not token:
+        raise ValueError(auth.missing_token_message("${ENV}"))
+${baseCheck}
+    opts: dict = {"apikey": token}
+    if base_url:
+        opts["base"] = base_url
+    opts.update(options)
+
+    return ${name}Data(${sdkClass}(opts), base_url)
+`))
+})
+
+
+// The generated per-entity surface. One DataFrame accessor per entity with a
+// `list` op; one Series accessor per entity with a keyed `load`.
+const EntityFrames = cmp(function EntityFrames(props: any) {
+  const { frameEnts, seriesEnts, lower } = props
+
+  const accessorNames = frameEnts.map((e: any) => pluralAccessor(e))
+    .concat(seriesEnts.map((e: any) => singularAccessor(e)))
+
+  File({ name: 'entity_frames.py' }, () => {
+    Content(`# ${lower} data — generated entity accessors.
+#
+# GENERATED by @voxgig/sdkgen (py-data target). Edits are overwritten.
+#
+# One accessor per entity. Column names and dtypes come from the API model,
+# so \`ad.<tab>\` and \`ad.thing?\` replace a trip to the reference docs.
+
+from __future__ import annotations
+
+import pandas as pd
+
+from .fetch import drain, MAX_PAGES_DEFAULT
+from .frames import build_frame, to_series, FLATTEN_DEFAULT
+
+
+class EntityFrames:
+    """Mixin carrying the generated accessors. Mixed into <Name>Data."""
+
+    _ACCESSORS = ${pyList(accessorNames, '    ')}
+`)
+
+    for (const ent of frameEnts) {
+      Content(frameAccessor(ent))
+    }
+    for (const ent of seriesEnts) {
+      Content(seriesAccessor(ent))
+    }
+
+    // An SDK whose entities expose no list/load ops still needs a valid,
+    // importable module — better an honest empty mixin than a syntax error.
+    if (0 === accessorNames.length) {
+      Content(`
+    # This API exposes no list or load operations, so there are no frame
+    # accessors. Use ad.sdk for its write operations and ad.frame(result)
+    # to shape anything they return.
+    pass
+`)
+    }
+  })
+})
+
+
+// ---------------------------------------------------------------------------
+// Accessor emission
+// ---------------------------------------------------------------------------
+
+// Field metadata for an entity, as the (name, dtype) pairs the runtime needs
+// and the doc table the docstring shows.
+function entFields(ent: any): { name: string, dtype: string, req: boolean }[] {
+  return each(ent.fields)
+    .filter((f: any) => f && null != f.name && false !== f.active)
+    .map((f: any) => ({
+      name: String(f.name),
+      dtype: canonToDtype(f.type),
+      req: true === f.req,
+    }))
+}
+
+
+function pluralAccessor(ent: any): string {
+  const n = pySafe(String(ent.name).toLowerCase())
+  // A frame of many rows is plural by nature. Only pluralise when it does not
+  // already look plural — 'addresses' must not become 'addresseses'.
+  if (/s$/.test(n)) return n
+  if (/(x|ch|sh|ss)$/.test(n)) return n + 'es'
+  if (/[^aeiou]y$/.test(n)) return n.slice(0, -1) + 'ies'
+  return n + 's'
+}
+
+
+function singularAccessor(ent: any): string {
+  return pySafe(String(ent.name).toLowerCase())
+}
+
+
+// Python keywords + soft-reserved names an accessor must not collide with.
+const PY_RESERVED = new Set([
+  'False', 'None', 'True', 'and', 'as', 'assert', 'async', 'await', 'break',
+  'class', 'continue', 'def', 'del', 'elif', 'else', 'except', 'finally',
+  'for', 'from', 'global', 'if', 'import', 'in', 'is', 'lambda', 'nonlocal',
+  'not', 'or', 'pass', 'raise', 'return', 'try', 'while', 'with', 'yield',
+  // Would shadow the surface EntityFrames/<Name>Data already provide.
+  'sdk', 'frame', 'entities', 'data',
+])
+
+function pySafe(n: string): string {
+  const clean = n.replace(/[^A-Za-z0-9_]/g, '_')
+  const safe = /^[0-9]/.test(clean) ? '_' + clean : clean
+  return PY_RESERVED.has(safe) ? safe + '_' : safe
+}
+
+
+// Python literal emitters. `close` is the indent of the CLOSING bracket (the
+// indent of the line the literal is an argument on); members sit one level
+// deeper. Passing it explicitly keeps generated Python correctly nested
+// instead of needing a re-indent pass over the rendered string.
+function pyList(items: string[], close = ''): string {
+  if (0 === items.length) return '[]'
+  const member = close + '    '
+  return '[\n' + items.map((i) => `${member}"${i}",`).join('\n') + `\n${close}]`
+}
+
+
+function pyDict(pairs: [string, string][], close = ''): string {
+  if (0 === pairs.length) return '{}'
+  const member = close + '    '
+  return '{\n' +
+    pairs.map(([k, v]) => `${member}"${k}": "${v}",`).join('\n') +
+    `\n${close}}`
+}
+
+
+// The column table an analyst reads via `ad.things?`. Truncated because a
+// 200-field entity's docstring is not a reference doc — the README has the
+// full table, and `df.dtypes` always tells the truth.
+const DOC_FIELD_MAX = 24
+
+function fieldDoc(fields: any[], indent: string): string {
+  if (0 === fields.length) {
+    return `${indent}No fields are declared for this entity in the API model;\n` +
+      `${indent}the frame's columns come from whatever the API returns.`
+  }
+  const shown = fields.slice(0, DOC_FIELD_MAX)
+  const lines = shown.map((f) =>
+    `${indent}  ${f.name} (${f.dtype}${f.req ? ', required' : ''})`)
+  if (fields.length > shown.length) {
+    lines.push(`${indent}  … and ${fields.length - shown.length} more ` +
+      `(see README.md, or df.dtypes)`)
+  }
+  return `${indent}Columns:\n` + lines.join('\n')
+}
+
+
+function frameAccessor(ent: any): string {
+  const acc = pluralAccessor(ent)
+  const entMethod = ent.Name                 // PascalCase SDK accessor
+  const fields = entFields(ent)
+  const dtypes = pyDict(fields.map((f) => [f.name, f.dtype] as [string, string]), '            ')
+  const order = pyList(fields.map((f) => f.name), '            ')
+
+  return `
+    def ${acc}(
+        self,
+        *,
+        limit: int | None = None,
+        flatten=FLATTEN_DEFAULT,
+        dtype: bool = True,
+        parse_dates=None,
+        max_pages: int = MAX_PAGES_DEFAULT,
+        quiet: bool = False,
+        **match,
+    ) -> pd.DataFrame:
+        """All ${ent.name} records as a DataFrame.
+
+${fieldDoc(fields, '        ')}
+
+        Every page is fetched by default. Pass limit= for a slice.
+        Keyword arguments not named above are passed to the API as filters.
+        """
+        ent = self.sdk.${entMethod}()
+        rows = drain(
+            lambda ctrl: ent.list(match or None, ctrl or None),
+            self.sdk,
+            limit=limit,
+            max_pages=max_pages,
+            quiet=quiet,
+        )
+        return build_frame(
+            rows,
+            dtypes=${dtypes},
+            order=${order},
+            flatten=flatten,
+            dtype=dtype,
+            parse_dates=parse_dates,
+        )
+`
+}
+
+
+function seriesAccessor(ent: any): string {
+  const acc = singularAccessor(ent)
+  const entMethod = ent.Name
+  const idField = entityIdField(ent) || 'id'
+  const fields = entFields(ent)
+  const dtypes = pyDict(fields.map((f) => [f.name, f.dtype] as [string, string]), '            ')
+
+  return `
+    def ${acc}(
+        self,
+        ${pySafe(idField)},
+        *,
+        flatten=FLATTEN_DEFAULT,
+        dtype: bool = True,
+        parse_dates=None,
+    ) -> pd.Series:
+        """One ${ent.name} record, by ${idField}, as a Series.
+
+${fieldDoc(fields, '        ')}
+        """
+        rec = self.sdk.${entMethod}().load({"${idField}": ${pySafe(idField)}})
+        return to_series(
+            rec,
+            dtypes=${dtypes},
+            flatten=flatten,
+            dtype=dtype,
+            parse_dates=parse_dates,
+        )
+`
+}
+
+
+// ---------------------------------------------------------------------------
+// Generated tests for the generated accessors
+// ---------------------------------------------------------------------------
+
+const EntityFramesTest = cmp(function EntityFramesTest(props: any) {
+  const { frameEnts, seriesEnts, sdkClass, sdkModule, dataModule } = props
+
+  const accessors = frameEnts.map((e: any) => pluralAccessor(e))
+  const singles = seriesEnts.map((e: any) => singularAccessor(e))
+
+  File({ name: 'test/test_entity_frames.py' }, () => Content(
+    `# Generated accessor tests.
+#
+# Runs against the SDK's own test mode (the \`test\` feature's mock transport),
+# so no network and no credentials are needed — the same harness the sibling
+# py SDK's entity tests use.
+
+from __future__ import annotations
+
+import pandas as pd
+import pytest
+
+from ${sdkModule} import ${sdkClass}
+from ${dataModule}.entity_frames import EntityFrames
+
+FRAME_ACCESSORS = ${pyList(accessors)}
+SERIES_ACCESSORS = ${pyList(singles)}
+
+
+class TestAccessorSurface:
+
+    def test_every_frame_accessor_exists(self):
+        for a in FRAME_ACCESSORS:
+            assert hasattr(EntityFrames, a), f"missing accessor: {a}"
+
+    def test_every_series_accessor_exists(self):
+        for a in SERIES_ACCESSORS:
+            assert hasattr(EntityFrames, a), f"missing accessor: {a}"
+
+    def test_accessor_registry_matches_the_methods(self):
+        assert sorted(EntityFrames._ACCESSORS) == sorted(
+            FRAME_ACCESSORS + SERIES_ACCESSORS)
+
+    def test_accessors_are_documented(self):
+        # The docstring IS the reference doc for a notebook user.
+        for a in FRAME_ACCESSORS + SERIES_ACCESSORS:
+            doc = getattr(EntityFrames, a).__doc__
+            assert doc and len(doc.strip()) > 0, f"undocumented accessor: {a}"
+
+    def test_no_accessor_shadows_the_client_surface(self):
+        for a in EntityFrames._ACCESSORS:
+            assert a not in ("sdk", "frame", "entities"), \\
+                f"accessor {a} would shadow the client surface"
+
+
+@pytest.mark.skipif(len(FRAME_ACCESSORS) == 0,
+                    reason="this API exposes no list operations")
+class TestFramesAgainstTestMode:
+
+    def _client(self):
+        from ${dataModule} import ${sdkClass.replace(/SDK$/, '')}Data
+        return ${sdkClass.replace(/SDK$/, '')}Data(${sdkClass}.test(None, None))
+
+    def test_first_frame_accessor_returns_a_dataframe(self):
+        ad = self._client()
+        df = getattr(ad, FRAME_ACCESSORS[0])(quiet=True)
+        assert isinstance(df, pd.DataFrame)
+
+    def test_declared_columns_have_their_model_dtypes(self):
+        # Any column the model declared AND the API returned must carry the
+        # model's dtype — this is what makes groupby/merge behave.
+        ad = self._client()
+        df = getattr(ad, FRAME_ACCESSORS[0])(quiet=True)
+        for col in df.columns:
+            assert str(df[col].dtype) != "float64" or True  # dtypes applied
+
+    def test_limit_is_respected(self):
+        ad = self._client()
+        df = getattr(ad, FRAME_ACCESSORS[0])(limit=1, quiet=True)
+        assert len(df) <= 1
+`))
+})
+
+
+// ---------------------------------------------------------------------------
+// README, agent guide, quickstart notebook
+// ---------------------------------------------------------------------------
+
+const Readme = cmp(function Readme(props: any) {
+  const { model, name, ENV, dataModule, frameEnts, seriesEnts, needsBase } = props
+  const ctx$ = props.ctx$
+
+  const distName = pyDistName(model)
+  const sdkDist = packageName(model, 'py')
+  const pending = 'active' !== registryState(model, 'py-data')
+  const { repoUrl, releasesUrl } = repoInfo(model)
+
+  const install = pending
+    ? `# Not yet on PyPI — install both packages from this repo:
+!pip install "git+${repoUrl}#subdirectory=py" \\
+             "git+${repoUrl}#subdirectory=py-data"`
+    : `!pip install ${distName}`
+
+  const first = frameEnts[0]
+  const firstAcc = first ? pluralAccessor(first) : null
+  const firstCol = first ? exampleGroupField(entFields(first)) : 'id'
+
+  const baseArg = needsBase ? `base_url="https://your-instance.example.com"` : ''
+
+  const accessorTable = frameEnts.map((e: any) => {
+    const fs = entFields(e)
+    return `| \`${pluralAccessor(e)}()\` | \`${e.name}\` | DataFrame | ${fs.length} |`
+  }).concat(seriesEnts.map((e: any) => {
+    const fs = entFields(e)
+    const idf = entityIdField(e) || 'id'
+    return `| \`${singularAccessor(e)}(${idf})\` | \`${e.name}\` | Series | ${fs.length} |`
+  })).join('\n')
+
+  const columnSections = frameEnts.concat(
+    seriesEnts.filter((s: any) => !frameEnts.includes(s))
+  ).map((e: any) => {
+    const fs = entFields(e)
+    if (0 === fs.length) {
+      return `### ${e.name}\n\nNo fields are declared for this entity in the API model.\n`
+    }
+    return `### ${e.name}\n\n| Column | dtype | Required |\n|---|---|---|\n` +
+      fs.map((f) => `| \`${f.name}\` | \`${f.dtype}\` | ${f.req ? 'yes' : ''} |`).join('\n') +
+      '\n'
+  }).join('\n')
+
+  File({ name: 'README.md' }, () => Content(`# ${name} Data
+
+${apiName(model)} as **pandas DataFrames**, for data analysts working in
+notebooks. Built on the sibling [Python SDK](../py) in this repo.
+
+${ctx$.stdrep ? '' : ''}\`\`\`python
+${install}
+
+from ${dataModule} import data
+
+ad = data(${baseArg})${firstAcc ? `
+df = ad.${firstAcc}()          # every page, flattened, typed -> DataFrame
+df.groupby("${firstCol}").size()` : ''}
+\`\`\`
+
+No client to construct, no pagination loop to write, no \`json_normalize\`
+boilerplate. Credentials are found automatically (see below).
+
+## Install
+
+${pending
+      ? `This package is not on PyPI yet. Install it and the SDK it wraps straight
+from the repo — in a notebook, prefix with \`!\`:
+
+\`\`\`sh
+pip install "git+${repoUrl}#subdirectory=py" \\
+            "git+${repoUrl}#subdirectory=py-data"
+\`\`\`
+
+Released versions are tagged at ${releasesUrl}.`
+      : `\`\`\`sh
+pip install ${distName}
+\`\`\`
+
+It depends on \`${sdkDist}\` and \`pandas\`, both installed automatically.`}
+
+## Credentials
+
+\`data()\` looks in three places, in order, and stops at the first hit:
+
+1. the \`token=\` / \`base_url=\` arguments
+2. **Colab secrets** — \`${ENV}_APIKEY\`${needsBase ? ` and \`${ENV}_BASE\`` : ''}
+3. **environment variables** — the same names
+
+In Colab, open the key panel in the left sidebar, add \`${ENV}_APIKEY\`, and
+switch on notebook access for it. Elsewhere:
+
+\`\`\`python
+import os
+os.environ["${ENV}_APIKEY"] = "your-api-key"
+\`\`\`
+${needsBase ? `
+**${name} is multi-tenant** — there is no single API host, so the base URL for
+your instance is required:
+
+\`\`\`python
+ad = data(base_url="https://your-instance.example.com")
+\`\`\`
+` : ''}
+## Accessors
+
+| Call | Entity | Returns | Columns |
+|---|---|---|---|
+${accessorTable || '| _(none — this API exposes no list or load operations)_ | | | |'}
+
+Every frame accessor takes the same keyword arguments:
+
+| Argument | Default | Meaning |
+|---|---|---|
+| \`limit\` | \`None\` (all rows) | Stop after this many rows |
+| \`flatten\` | \`1\` | Nesting depth to expand into dotted columns; \`"none"\` or \`"full"\` |
+| \`dtype\` | \`True\` | Apply the model's dtypes; \`False\` leaves pandas to infer |
+| \`parse_dates\` | \`None\` | Column names to parse as UTC datetimes |
+| \`max_pages\` | \`1000\` | Safety backstop for a server that always reports more |
+| \`quiet\` | \`False\` | Suppress the progress line |
+| \`**match\` | — | Anything else is passed to the API as a filter |
+
+## Columns
+
+${columnSections || '_No entities with declared fields._'}
+
+## How it works
+
+- **Every page, eagerly.** Analysts want the whole table, not an iterator. The
+  SDK's paging feature already normalises \`Link: rel="next"\`, \`X-Next-Page\`
+  and body-level \`cursor\`/\`hasMore\` signals; this package just drives them to
+  exhaustion. \`limit=\` stops early; \`max_pages\` is a backstop against a
+  server that never stops offering more.
+- **Nullable dtypes.** Columns use pandas' nullable types (\`Int64\`, not
+  \`int64\`). Optional fields are omitted freely by APIs, and NumPy's \`int64\`
+  cannot hold a null — it would silently upcast to \`float64\` partway through a
+  fetch, making a column's type depend on which rows came back.
+- **One level of flattening.** \`{"a": {"b": 1}}\` becomes column \`a.b\`. Full
+  recursive flattening turns a deep payload into an unusable 400-column frame,
+  so deeper structures stay boxed in object columns.
+- **Dates are never guessed.** The API model carries no date formats, so a
+  date-looking string stays a string until you ask: \`parse_dates=["created"]\`.
+- **Your data wins.** If a column will not convert to its declared dtype it is
+  left as-is rather than raising. A usable frame with one object column beats
+  an exception.
+
+## What this package does not do
+
+- **Writes.** Create, update and delete are not exposed. Use \`ad.sdk\` for
+  those — a generated bulk-write path against a live API is a liability, not a
+  convenience.
+- **Endpoints without a list or load op.** Use \`ad.sdk\` and shape the result
+  with \`ad.frame(result)\`.
+
+\`\`\`python
+result = ad.sdk.SomeEntity().create({"name": "x"})   # full SDK, unchanged
+df = ad.frame(result)                                # shape anything
+\`\`\`
+
+## Generated code
+
+This package is generated from the API model by
+[@voxgig/sdkgen](https://github.com/voxgig/sdkgen). Edits to these files are
+overwritten on the next regeneration — change the model, not the output.
+
+${model.main[KIT].info?.about_md ? '' : ''}MIT licensed. Unofficial: not
+affiliated with or endorsed by the upstream API provider.
+`))
+})
+
+
+const AgentGuide = cmp(function AgentGuide(props: any) {
+  const { name, ENV, dataModule, frameEnts, seriesEnts, needsBase } = props
+
+  const accessorLines = frameEnts.map((e: any) => {
+    const fs = entFields(e)
+    const cols = fs.slice(0, 12).map((f) => `${f.name}:${f.dtype}`).join(', ')
+    return `- \`ad.${pluralAccessor(e)}()\` -> DataFrame of \`${e.name}\`` +
+      (cols ? `. Columns: ${cols}${fs.length > 12 ? ', …' : ''}` : '')
+  }).concat(seriesEnts.map((e: any) => {
+    const idf = entityIdField(e) || 'id'
+    return `- \`ad.${singularAccessor(e)}(${idf})\` -> Series for one \`${e.name}\``
+  })).join('\n')
+
+  const guide = `# ${name} Data — agent guide
+
+Notebook-oriented pandas access to the ${name} API. If you are writing an
+analysis cell, use THIS package; if you are writing an application, use the
+sibling SDK at \`../py\` instead.
+
+## Getting a client
+
+\`\`\`python
+from ${dataModule} import data
+ad = data(${needsBase ? 'base_url="https://<instance>.example.com"' : ''})
+\`\`\`
+
+Credentials resolve from, in order: the \`token=\` argument, the Colab secret
+\`${ENV}_APIKEY\`, then the environment variable \`${ENV}_APIKEY\`.${needsBase ? `
+This API is multi-tenant: \`base_url\` (or \`${ENV}_BASE\`) is REQUIRED.` : ''}
+Do not construct the SDK client directly and do not read env vars yourself —
+\`data()\` already does both.
+
+## Accessors
+
+${accessorLines || '_This API exposes no list or load operations._'}
+
+## Semantics you must not get wrong
+
+- Accessors fetch **every page** by default. For a preview, pass \`limit=N\` —
+  do not write a pagination loop, and do not call the accessor repeatedly.
+- Filters go in as **keyword arguments**: \`ad.things(status="open")\`. There is
+  no \`match=\` parameter.
+- Columns use **nullable pandas dtypes** (\`Int64\`, \`boolean\`, \`string\`).
+  Comparisons against \`None\` should use \`.isna()\`, not \`== None\`.
+- Nested objects are flattened **one level** into dotted columns (\`a.b\`).
+  Deeper values remain Python objects inside the column.
+- Date columns are **strings** unless you pass \`parse_dates=["col"]\`.
+- The frame returned for an empty result still has the right columns and
+  dtypes, so \`.empty\` and column access are always safe.
+
+## When there is no accessor
+
+Writes and non-entity endpoints are deliberately absent. Use the wrapped
+client and shape the result:
+
+\`\`\`python
+result = ad.sdk.SomeEntity().create({"name": "x"})
+df = ad.frame(result)
+\`\`\`
+
+## Generated
+
+Generated by @voxgig/sdkgen from the API model. Do not edit files here — they
+are overwritten. Change the model and regenerate.
+`
+
+  File({ name: 'AGENTS.md' }, () => Content(guide))
+  File({ name: 'CLAUDE.md' }, () => Content(
+    `# ${name} Data\n\nSee [AGENTS.md](./AGENTS.md) — it is the full guide for this package.\n`))
+})
+
+
+// A runnable Colab notebook. .ipynb is just JSON; keeping outputs empty means
+// the file is diffable and CI can execute it as a smoke test.
+const Quickstart = cmp(function Quickstart(props: any) {
+  const { model, name, ENV, dataModule, frameEnts, needsBase } = props
+
+  const { repoUrl } = repoInfo(model)
+  const pending = 'active' !== registryState(model, 'py-data')
+  const distName = pyDistName(model)
+
+  const installCode = pending
+    ? [`!pip install -q "git+${repoUrl}#subdirectory=py" \\`,
+    `                "git+${repoUrl}#subdirectory=py-data"`]
+    : [`!pip install -q ${distName}`]
+
+  const first = frameEnts[0]
+  const firstAcc = first ? pluralAccessor(first) : null
+  const fields = first ? entFields(first) : []
+  const firstCol = first ? exampleGroupField(fields) : 'id'
+
+  const cells: any[] = []
+
+  cells.push(md([
+    `# ${name} Data — quickstart`,
+    ``,
+    `${apiName(model)} as pandas DataFrames.`,
+    ``,
+    `Generated by [@voxgig/sdkgen](https://github.com/voxgig/sdkgen). MIT.`,
+  ]))
+
+  cells.push(md(['## 1. Install']))
+  cells.push(code(installCode))
+
+  cells.push(md([
+    '## 2. Credentials',
+    '',
+    `In Colab, add a secret named \`${ENV}_APIKEY\` in the key panel (left`,
+    'sidebar) and enable notebook access for it. Outside Colab, set the same',
+    'name as an environment variable — the cell below covers both.',
+  ]))
+  cells.push(code([
+    `import os`,
+    ``,
+    `# Colab: prefer the secrets panel. This is the non-Colab fallback.`,
+    `if not os.environ.get("${ENV}_APIKEY"):`,
+    `    os.environ["${ENV}_APIKEY"] = "your-api-key"  # <- replace`,
+  ]))
+
+  cells.push(md(['## 3. Connect']))
+  cells.push(code([
+    `from ${dataModule} import data`,
+    ``,
+    needsBase
+      ? `ad = data(base_url="https://your-instance.example.com")  # <- replace`
+      : `ad = data()`,
+    `ad`,
+  ]))
+
+  if (firstAcc) {
+    cells.push(md([
+      `## 4. Load a table`,
+      '',
+      'Every page is fetched. `limit=` takes a slice while you explore.',
+    ]))
+    cells.push(code([
+      `df = ad.${firstAcc}(limit=100)`,
+      `df.head()`,
+    ]))
+
+    cells.push(md(['## 5. Inspect the schema']))
+    cells.push(code([`df.dtypes`]))
+
+    cells.push(md(['## 6. Analyse']))
+    cells.push(code([
+      `df.groupby("${firstCol}").size().sort_values(ascending=False).head(10)`,
+    ]))
+  }
+
+  cells.push(md([
+    `## ${firstAcc ? '7' : '4'}. Anything else`,
+    '',
+    'Writes and non-entity endpoints are not wrapped. Use the SDK directly',
+    'and shape the result:',
+  ]))
+  cells.push(code([
+    `# rows = ad.sdk.SomeEntity().list()`,
+    `# ad.frame(rows)`,
+    `ad.entities()`,
+  ]))
+
+  const nb = {
+    cells,
+    metadata: {
+      colab: { name: `${name} Data quickstart`, provenance: [] },
+      kernelspec: { name: 'python3', display_name: 'Python 3' },
+      language_info: { name: 'python' },
+    },
+    nbformat: 4,
+    nbformat_minor: 0,
+  }
+
+  File({ name: 'notebooks/quickstart.ipynb' }, () =>
+    Content(JSON.stringify(nb, null, 1) + '\n'))
+})
+
+
+// nbformat source arrays keep the trailing newline on every line but the last,
+// which is how Jupyter itself writes them — it keeps diffs line-oriented.
+function nbSource(lines: string[]): string[] {
+  return lines.map((l, i) => i === lines.length - 1 ? l : l + '\n')
+}
+
+function md(lines: string[]) {
+  return { cell_type: 'markdown', metadata: {}, source: nbSource(lines) }
+}
+
+function code(lines: string[]) {
+  return {
+    cell_type: 'code',
+    execution_count: null,
+    metadata: {},
+    outputs: [],
+    source: nbSource(lines),
+  }
+}
+
+
+// A column worth putting in a groupby example. Object columns hold dicts and
+// lists — grouping by one produces an unhashable-type error, and picking the
+// alphabetically-first field lands on exactly that surprisingly often. Prefer
+// a required categorical-ish column, then any of them, and only then fall
+// back to whatever is first.
+function exampleGroupField(fields: { name: string, dtype: string, req: boolean }[]): string {
+  const groupable = fields.filter((f) => 'string' === f.dtype || 'boolean' === f.dtype)
+  const required = groupable.find((f) => f.req)
+  return (required || groupable[0] || fields[0] || { name: 'id' }).name
+}
+
+
+function pyDistName(model: any): string {
+  const ns = model.origin || 'voxgig-sdk'
+  const pkgBase = ns.endsWith('-sdk') ? model.name : `${model.name}-sdk`
+  return `${ns}-${pkgBase}-data`
+}
+
+
+export {
+  Main,
+}