mikser-io 9.41.0 → 9.43.0

package/docs/api-reference.md

@@ -483,6 +483,138 @@ Query types throughout: function, lodash match object, or `undefined` for all.
 
 ---
 
+## Writing source files
+
+`updateEntity` is a catalog operation. `writeEntitySource` writes the FILE, with
+the checks that make a whole-file rewrite safe to perform without having watched
+the file the whole time.
+
+### `writeEntitySource(options)`
+
+```js
+import { writeEntitySource } from 'mikser-io'
+
+const preview = await writeEntitySource({
+    id: '/documents/about.md',
+    content: next,
+    dryRun: true,          // writes nothing; reports what it would re-render
+})
+
+const result = await writeEntitySource({
+    id: '/documents/about.md',
+    content: next,
+    ifChecksum: preview.currentChecksum,
+})
+```
+
+| Option | Meaning |
+| --- | --- |
+| `id` | Catalog id of an existing entity. Alternative to the pair below. |
+| `collection` + `relativePath` | Where to write. Required unless `id` is given. |
+| `content` | The COMPLETE file. Anything omitted is deleted — there is no patch mode. |
+| `ifChecksum` | Only write if the file's current DISK checksum equals this. |
+| `dryRun` | Write nothing; return the blast radius and any advisory. |
+| `awaitCycle` | Resolve once the cycle that picks the write up finishes, with its build report attached as `report`. |
+
+**Never throws for an expected outcome.** A bad id, a path that escapes the
+collection, a checksum that no longer matches — each returns
+`{ ok: false, refused }` with the facts needed to retry, because those are
+answers rather than faults. `refused` is one of `unresolvable-id`,
+`collection-mismatch`, `incomplete-target`, `invalid-target`,
+`checksum-mismatch`.
+
+**Containment.** `relativePath` cannot leave the collection folder. This
+matters because the path often comes from a request body or a CMS form, and
+`path.join(folder, '../../x')` resolves outside the folder and writes there. It
+is resolved and then contained rather than rejected on a literal `..`, so
+`blog/../about.md` still works. The refusal happens before anything stats the
+file — reporting a checksum for an out-of-tree path is a disclosure on its own.
+
+**The precondition is not a lock.** A writer landing between the check and the
+write still wins. It closes the window that matters in practice: read, think,
+write back a whole file built from a copy that is now stale.
+
+`ifChecksum` is compared against the DISK. `readEntity`'s `checksum` is the
+catalog's, which lags between builds — pass `diskChecksum`, or the
+`currentChecksum` a refusal hands back.
+
+### Advisories
+
+`contentAdvisories(entity, content)` names files a caller must not edit blind,
+from `meta.specLocked` / `meta.generated` or from a header in the first 40
+lines. Two kinds, kept apart because the instruction differs: `spec-locked`
+means the bytes answer to a document outside the repo, `generated` means
+editing the file is pointless because the next build overwrites it.
+`advisoryWarning(advisories)` renders one line of prose for a response meant to
+be read rather than parsed. Both are reported by `writeEntitySource` — on the
+dry run and again on the way out, since a caller that never read the file is
+exactly the one that needs telling.
+
+`siblingDestinations(folder, relativePath)` reports files differing only by
+extension, which may render to the same destination.
+
+## Search
+
+`queryEntities` sifts **meta**. `searchEntities` answers the other question —
+*where does this text appear?* — across structured values, source files and
+built output.
+
+### `searchEntities(options)`
+
+```js
+import { searchEntities } from 'mikser-io'
+
+const { hits, count, truncated } = await searchEntities({
+    query: 'NOVAPRESS',
+    in: ['meta', 'content'],   // default
+})
+```
+
+| Option | Meaning |
+| --- | --- |
+| `query` | Text to find. A plain substring unless `regex` is true. Required. |
+| `in` | `'meta'`, `'content'`, `'output'`. Default `['meta', 'content']`. |
+| `collection` | Restrict to one collection. |
+| `filter` | Sift filter narrowing **what may be searched at all** — see below. |
+| `regex` | Treat `query` as a JavaScript regular expression. |
+| `ignoreCase` | Case-insensitive matching. Default false. |
+| `limit` | Maximum hits (default 50). `truncated` says when it stopped early. |
+
+The three scopes answer different questions and none implies another. `meta`
+walks structured values as dotted paths and touches no files. `content` reads
+the **source**. `output` walks the **built** folder and reports `occurrences`
+per destination — the blast-radius question, and the one where a string can be
+present because a layout writes it, with no source entity containing it
+anywhere.
+
+Whether a file is text is decided by reading its bytes, not by its extension,
+so a `.njk` source or a `.webmanifest` output is searched without first being
+added to a list of known formats.
+
+`truncated` is load-bearing: it separates "these are the hits" from "these are
+the first N", which a caller acting on a blast radius cannot afford to guess.
+
+> **Unscoped by default.** A bare call reads every entity and every source
+> file — drafts, unpublished documents, layouts, sidecar JavaScript. The `api`
+> plugin narrows `list` through each endpoint's own sift scope; nothing
+> narrows this. Anything that puts search behind a request **must** pass that
+> same scope as `filter`, or a public endpoint listing only published
+> documents will happily search the unpublished ones. There is deliberately no
+> `search` operation in the `api` plugin and no route in this module.
+
+### Primitives
+
+Exported so a caller building a different question out of the same parts does
+not reimplement them: `countMatches`, `snippetAround`, `lineOfFirstMatch`,
+`flattenMeta`, `findOccurrences`, `walkFiles`.
+
+`findOccurrences(text, needle)` returns every occurrence with `line`, `col`,
+the line's `text`, and `leading` — whether the match begins its line. That one
+flag separates the file **declaring** something from the files merely using
+it, in any text format, with no per-language grammar involved. It returns the
+line rather than a verdict about it, so where the heuristic is wrong the
+evidence is in the result.
+
 ## Database
 
 ```js

package/index.js

@@ -11,6 +11,8 @@ export * from './src/lifecycle.js'
 export * from './src/database/index.js'
 export * from './src/journal.js'
 export * from './src/catalog.js'
+export * from './src/search.js'
+export * from './src/write.js'
 export * from './src/refs.js'
 export * from './src/manifest.js'
 export * from './src/provenance.js'

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mikser-io",
-  "version": "9.41.0",
+  "version": "9.43.0",
   "description": "<p align=\"center\">   <img src=\"mikser-lockup-stacked.svg\" alt=\"mikser\" width=\"198\" /> </p>",
   "main": "index.js",
   "exports": {

package/src/builtin-tools.js

@@ -96,6 +96,45 @@ export function registerBuiltinTools() {
         },
     )
 
+    registerTool(
+        'search',
+        {
+            description:
+                'Find a string across the catalog in ONE call — "where does this appear?". Searches entity meta '
+                + 'values and, when asked, the source files themselves and the built output, returning '
+                + '{ id, collection, path, field, snippet } per hit.\n\n'
+                + 'This is how you locate content you can only describe by what it says: a menu label, a phone '
+                + 'number, a sentence you were asked to change. Paging the catalog to find it means reading '
+                + 'everything, most of which is fonts and image derivatives, and finding a SECOND copy of the same '
+                + 'label somewhere else is then a matter of luck.\n\n'
+                + 'in: ["meta"] searches structured values, no file I/O. in: ["content"] reads source files, '
+                + 'binaries skipped. Default is both. in: ["output"] searches the BUILT files instead and reports '
+                + '`occurrences` per destination — the blast-radius question, which is what you want before editing '
+                + 'anything shared. The scopes answer different questions and none implies another: a string can be '
+                + 'in the output because a layout writes it, with no source entity containing it anywhere.',
+            inputSchema: {
+                query: { type: 'string', required: true,
+                    description: 'Text to find. A plain substring unless `regex` is true. Case-sensitive by default.' },
+                collection: { type: 'string',
+                    description: 'Restrict to one collection (e.g. "documents"). Omit to search all.' },
+                in: { type: 'array',
+                    description: 'Where to look: "meta", "content", "output". Default is meta + content.' },
+                regex: { type: 'boolean',
+                    description: 'Treat `query` as a JavaScript regular expression rather than a literal substring.' },
+                ignoreCase: { type: 'boolean', description: 'Case-insensitive matching. Default false.' },
+                limit: { type: 'number', description: 'Maximum hits to return (default 50).' },
+            },
+        },
+        async (args) => {
+            try {
+                const { searchEntities } = await import('./search.js')
+                return ok(await searchEntities(args ?? {}))
+            } catch (err) {
+                return fail(err.message)
+            }
+        },
+    )
+
     registerTool(
         'sources',
         {

package/src/search.js

@@ -0,0 +1,284 @@
+// Find a string across the catalog and the built output.
+//
+// `queryEntities` sifts META. This answers the other question — "where does
+// this text appear?" — over structured values, source files and built output,
+// which is what locating a phone number in a document body or a class name in
+// the shipped CSS requires.
+//
+// Three scopes, answering three different questions, none implying another:
+//
+//   meta     — structured values, walked as dotted paths. No file I/O.
+//   content  — the SOURCE files. What an editor is going to change.
+//   output   — the BUILT files. The blast radius: a string can be in the
+//              output because a layout writes it, with no source entity
+//              containing it anywhere.
+//
+// UNSCOPED BY DEFAULT, and that is a security property, not a detail. A bare
+// call reads every entity and every source file — drafts, unpublished
+// documents, layouts, sidecar JavaScript. The `api` plugin's public endpoints
+// narrow `list` through the endpoint's own sift scope; nothing narrows this.
+// So anything that puts search behind a request MUST pass that same scope as
+// `filter`, or a public endpoint that lists only published documents will
+// happily search the unpublished ones. There is deliberately no `search`
+// operation in the api plugin, and no route here.
+//
+// The primitives are exported alongside `searchEntities` so a caller building
+// a different question out of the same parts does not reimplement counting,
+// snippets or line numbers to do it.
+
+import path from 'node:path'
+import { readdir, readFile } from 'node:fs/promises'
+
+import runtime from './runtime.js'
+import { findEntities } from './catalog.js'
+import { readEntityContent, looksTextual } from './utils.js'
+
+// How much of an output file to read before deciding it is text. Same rule
+// the source read uses, so the two scopes cannot disagree about a file that
+// appears in both.
+const SNIFF_BYTES = 8 * 1024
+
+// Every leaf value under meta as [dottedPath, value], arrays included.
+// The same shape refs_inbound reports, so a hit here and a referrer there
+// name the same field.
+export function* flattenMeta(node, prefix = '') {
+    if (node === null || node === undefined) return
+    if (Array.isArray(node)) {
+        for (let i = 0; i < node.length; i++) yield* flattenMeta(node[i], `${prefix}[${i}]`)
+        return
+    }
+    if (typeof node === 'object') {
+        for (const [key, value] of Object.entries(node)) {
+            yield* flattenMeta(value, prefix ? `${prefix}.${key}` : key)
+        }
+        return
+    }
+    yield [prefix, node]
+}
+
+// How many times the needle appears. A count, not a boolean, because "this
+// class is on nine pages" and "this class is on nine pages and seven times on
+// one of them" are different facts, and the second one names the component.
+export function countMatches(text, query, regex, ignoreCase) {
+    if (regex) {
+        const re = new RegExp(query, ignoreCase ? 'gi' : 'g')
+        let n = 0
+        // Guard the zero-width case: /a*/g on a non-matching position returns
+        // an empty match forever and never advances.
+        for (let m = re.exec(text); m; m = re.exec(text)) {
+            n++
+            if (m.index === re.lastIndex) re.lastIndex++
+        }
+        return n
+    }
+    const haystack = ignoreCase ? text.toLowerCase() : text
+    const needle = ignoreCase ? query.toLowerCase() : query
+    if (!needle) return 0
+    let n = 0
+    for (let at = haystack.indexOf(needle); at !== -1; at = haystack.indexOf(needle, at + needle.length)) n++
+    return n
+}
+
+// Offset of the first match, or -1. The one place the regex/substring
+// difference is resolved, so the line number and the snippet cannot end up
+// pointing at different matches.
+function firstMatchAt(text, query, regex, ignoreCase) {
+    if (regex) {
+        const m = new RegExp(query, ignoreCase ? 'i' : '').exec(text)
+        return m ? m.index : -1
+    }
+    const haystack = ignoreCase ? text.toLowerCase() : text
+    return haystack.indexOf(ignoreCase ? query.toLowerCase() : query)
+}
+
+// 1-based line of the first match, so a hit is somewhere to go rather than
+// something to grep for again.
+export function lineOfFirstMatch(text, query, regex, ignoreCase) {
+    const at = firstMatchAt(text, query, regex, ignoreCase)
+    if (at < 0) return null
+    let line = 1
+    for (let i = 0; i < at; i++) if (text.charCodeAt(i) === 10) line++
+    return line
+}
+
+// Enough text around the match to recognise it without returning the file.
+export function snippetAround(text, query, regex, ignoreCase) {
+    const at = firstMatchAt(text, query, regex, ignoreCase)
+    if (at < 0) return text.slice(0, 120)
+    const start = Math.max(0, at - 60)
+    const end = Math.min(text.length, at + query.length + 60)
+    return (start > 0 ? '…' : '') + text.slice(start, end).replace(/\s+/g, ' ') + (end < text.length ? '…' : '')
+}
+
+// Every occurrence of a needle in a text, with the one signal that separates
+// "this file DECLARES it" from "this file uses it" without knowing the
+// language: a declaration begins its line in nearly every text format, while a
+// use sits mid-line.
+//
+// The LINE is returned rather than a computed verdict about it. A caller can
+// tell a declaration from something that merely starts with the same token by
+// reading it, which needs no grammar here at all — and where the heuristic is
+// wrong, the evidence for that is right there in the result.
+export function findOccurrences(text, needle, { limit = 200 } = {}) {
+    const sites = []
+    if (!needle || typeof text !== 'string') return sites
+    let line = 1
+    let lineStart = 0
+    let scanned = 0
+    for (let at = text.indexOf(needle); at !== -1; at = text.indexOf(needle, at + needle.length)) {
+        while (scanned < at) {
+            if (text.charCodeAt(scanned) === 10) { line++; lineStart = scanned + 1 }
+            scanned++
+        }
+        const lineEnd = text.indexOf('\n', at)
+        sites.push({
+            line,
+            col: at - lineStart,
+            leading: text.slice(lineStart, at).trim() === '',
+            text: text.slice(lineStart, lineEnd === -1 ? text.length : lineEnd).trim().slice(0, 160),
+        })
+        if (sites.length >= limit) break
+    }
+    return sites
+}
+
+// Every file under a folder, depth-first. A generator so a search that hits
+// its limit stops walking rather than materializing the tree.
+export async function* walkFiles(folder) {
+    let entries
+    try {
+        entries = await readdir(folder, { withFileTypes: true })
+    } catch {
+        return
+    }
+    for (const entry of entries) {
+        const full = path.join(folder, entry.name)
+        if (entry.isDirectory()) yield* walkFiles(full)
+        else if (entry.isFile()) yield full
+    }
+}
+
+// Read a built file as text, or null when it is not text.
+//
+// Decided by the bytes, so a `.webmanifest`, a `.map` or whatever a renderer
+// plugin emits is searchable without first being added to a list of known
+// extensions — and a font still comes back null rather than as convincing
+// garbage.
+async function readOutputText(file) {
+    let buf
+    try {
+        buf = await readFile(file)
+    } catch {
+        return null
+    }
+    if (!looksTextual(buf.subarray(0, Math.min(buf.length, SNIFF_BYTES)))) return null
+    return buf.toString('utf8')
+}
+
+// Build the test used for meta values and for a whole-file contains check.
+// Throws on an invalid regex so the caller can report it as bad input rather
+// than as an empty result.
+function buildMatcher(query, regex, ignoreCase) {
+    if (regex) {
+        const re = new RegExp(query, ignoreCase ? 'i' : '')
+        return { test: (value) => re.test(String(value)) }
+    }
+    const needle = ignoreCase ? query.toLowerCase() : query
+    return {
+        test: (value) => (ignoreCase ? String(value).toLowerCase() : String(value)).includes(needle),
+    }
+}
+
+// Find `query` across the catalog and, when asked, the built output.
+//
+// Returns { query, scopes, count, truncated, searched, hits }. Each hit
+// carries where it was found and enough to go there:
+//
+//   meta     { id, collection, path, where, field, snippet }
+//   content  { id, collection, path, where, field: null, occurrences, line, snippet }
+//   output   { destination, where, occurrences, snippet }
+//
+// `truncated` says the limit stopped the walk, which is the difference
+// between "these are the hits" and "these are the first N" — a distinction a
+// caller acting on a blast radius cannot afford to guess at.
+export async function searchEntities({
+    query,
+    collection,
+    filter,
+    in: where,
+    regex = false,
+    ignoreCase = false,
+    limit = 50,
+} = {}) {
+    if (!query) throw new Error('searchEntities: query is required')
+    const scopes = where?.length ? where : ['meta', 'content']
+    const matcher = buildMatcher(query, regex, ignoreCase)
+
+    const hits = []
+    let truncated = false
+
+    // The output scope walks the deployed folder rather than the catalog, so
+    // it runs on its own. Counting is the point here rather than first-match:
+    // seven occurrences on one page and one on nine others is the shape of a
+    // shared component, and a list of nine equal-looking filenames hides it.
+    if (scopes.includes('output')) {
+        const outputFolder = runtime.options?.outputFolder
+        if (!outputFolder) throw new Error('No output folder configured, so there is no built output to search.')
+        for await (const file of walkFiles(outputFolder)) {
+            if (hits.length >= limit) { truncated = true; break }
+            const text = await readOutputText(file)
+            if (text === null) continue
+            const occurrences = countMatches(text, query, regex, ignoreCase)
+            if (!occurrences) continue
+            hits.push({
+                destination: '/' + path.relative(outputFolder, file).split(path.sep).join('/'),
+                where: 'output',
+                occurrences,
+                snippet: snippetAround(text, query, regex, ignoreCase),
+            })
+        }
+        hits.sort((a, b) => b.occurrences - a.occurrences || a.destination.localeCompare(b.destination))
+    }
+
+    // `filter` is how a caller narrows what may be searched at all. It merges
+    // with `collection` and pushes into the catalog query, so a scoped caller
+    // never materializes a row it is not allowed to see rather than filtering
+    // one out after reading it off disk.
+    const searchesCatalog = scopes.some(scope => scope !== 'output')
+    const catalogQuery = { ...(filter ?? {}), ...(collection ? { collection } : {}) }
+    const entities = searchesCatalog
+        ? await findEntities(Object.keys(catalogQuery).length ? catalogQuery : undefined)
+        : []
+
+    for (const entity of entities) {
+        if (hits.length >= limit) { truncated = true; break }
+        if (scopes.includes('meta') && entity.meta) {
+            for (const [field, value] of flattenMeta(entity.meta)) {
+                if (!matcher.test(value)) continue
+                hits.push({
+                    id: entity.id, collection: entity.collection ?? null, path: entity.uri ?? null,
+                    where: 'meta', field, snippet: snippetAround(String(value), query, regex, ignoreCase),
+                })
+                break
+            }
+        }
+        if (hits.length >= limit) { truncated = true; break }
+        if (scopes.includes('content')) {
+            // readEntityContent owns the text/binary decision and makes it by
+            // reading the bytes, so a .njk or .toml is searched like anything
+            // else. Deciding it a second time here on an extension would make
+            // a skipped format indistinguishable from a real absence.
+            const { content } = await readEntityContent(entity)
+            if (typeof content !== 'string' || !matcher.test(content)) continue
+            hits.push({
+                id: entity.id, collection: entity.collection ?? null, path: entity.uri ?? null,
+                where: 'content', field: null,
+                occurrences: countMatches(content, query, regex, ignoreCase),
+                line: lineOfFirstMatch(content, query, regex, ignoreCase),
+                snippet: snippetAround(content, query, regex, ignoreCase),
+            })
+        }
+    }
+
+    return { query, scopes, count: hits.length, truncated, searched: entities.length, hits }
+}

package/src/utils.js

@@ -1077,20 +1077,38 @@ export function useCollection(runtime, name) {
         return folder
     }
 
+    // A path that cannot leave the collection folder.
+    //
+    // `path.join(folder, '../../x')` resolves outside the folder and writes
+    // there, which turns a collection handle into an arbitrary-write
+    // primitive the moment a relative path comes from a request body or a CMS
+    // form. Resolved and then contained rather than rejected on a literal
+    // `..`, so `a/../b.md` — which lands inside — still works.
+    function resolveWithin(relativePath) {
+        const folder = resolveFolder()
+        const uri = path.resolve(folder, relativePath ?? '')
+        const root = path.resolve(folder)
+        if (uri !== root && !uri.startsWith(root + path.sep)) {
+            throw new Error(
+                `Path escapes the ${name} collection: ${JSON.stringify(relativePath)} resolves outside ${root}`)
+        }
+        return uri
+    }
+
     return {
         name,
         get folder() { return resolveFolder() },
+        resolveWithin,
 
         async write(relativePath, content = '') {
-            const uri = path.join(resolveFolder(), relativePath)
+            const uri = resolveWithin(relativePath)
             await mkdir(path.dirname(uri), { recursive: true })
             await writeFile(uri, content, 'utf8')
             return uri
         },
 
         async remove(relativePath) {
-            const uri = path.join(resolveFolder(), relativePath)
-            await unlink(uri)
+            await unlink(resolveWithin(relativePath))
         },
     }
 }

package/src/write.js

@@ -0,0 +1,279 @@
+// Write a source file back, with the checks that make a whole-file rewrite
+// safe to perform without having watched the file the whole time.
+//
+// The only write mode is whole-file: there is no patch. That makes the write
+// itself the easy part and everything around it the point —
+//
+//   - a checksum precondition, so a rewrite built from a stale copy is
+//     refused instead of silently discarding whoever edited in between
+//   - a dry run reporting which destinations the write would re-render
+//   - advisories naming a file that is GENERATED or answers to an external
+//     spec, which a caller must not edit blind
+//   - siblings that could render to the same destination
+//   - containment, so a relative path from a form or a request body cannot
+//     write outside the collection folder
+//
+// An editing agent gets these through mikser-io-mcp. An application driving
+// mikser as a CMS writes its own files and gets none of them, which is the
+// gap this closes: the safety belongs to the write, not to one transport.
+
+import path from 'node:path'
+import { readdir } from 'node:fs/promises'
+
+import runtime from './runtime.js'
+import { readEntity, findEntities } from './catalog.js'
+import { useCollection, checksum, readEntityContent } from './utils.js'
+import { nextCycleId, whenCycleCompletes } from './report.js'
+
+// How far into a file to look for a marker. A header nobody reads is not a
+// header; one buried 200 lines down is not either.
+const HEADER_SCAN_LINES = 40
+
+// Files a caller must not edit blind, read from the bytes themselves.
+//
+// Two kinds, kept apart because the instruction differs. `spec-locked` means
+// the bytes answer to a document outside the repo — change it and the site
+// stops matching something a human signed off. `generated` means editing the
+// file is pointless, because the next build overwrites it.
+const HEADER_PATTERNS = [
+    { kind: 'spec-locked', re: /^\W*spec source:\s*(.+?)\s*$/i },
+    { kind: 'generated',   re: /^\W*(?:generated by|do not edit)\b:?\s*(.*?)\s*$/i },
+]
+
+export function contentAdvisories(entity, content) {
+    const found = []
+    const push = (kind, detail, via, line) => {
+        if (found.some(a => a.kind === kind)) return
+        found.push({ kind, detail: detail || null, via, ...(line ? { line } : {}) })
+    }
+    // Explicit meta wins: someone wrote it down as data, on purpose.
+    if (entity?.meta?.specLocked) {
+        push('spec-locked', typeof entity.meta.specLocked === 'string' ? entity.meta.specLocked : null, 'meta.specLocked')
+    }
+    if (entity?.meta?.generated) {
+        push('generated', typeof entity.meta.generated === 'string' ? entity.meta.generated : null, 'meta.generated')
+    }
+    if (typeof content === 'string') {
+        const lines = content.split('\n', HEADER_SCAN_LINES)
+        for (let i = 0; i < lines.length; i++) {
+            for (const { kind, re } of HEADER_PATTERNS) {
+                const m = re.exec(lines[i])
+                if (m) push(kind, m[1], 'header', i + 1)
+            }
+        }
+    }
+    return found
+}
+
+// One line of prose for a response that has to be read, not parsed.
+export function advisoryWarning(advisories) {
+    if (!advisories?.length) return null
+    return advisories.map(a => a.kind === 'spec-locked'
+        ? `SPEC-LOCKED: ${a.detail ?? 'this file answers to an external specification'}`
+            + ' — changing it may break a signed-off design. Confirm against the spec before writing.'
+        : `GENERATED: ${a.detail ?? 'this file is produced by the build'}`
+            + ' — edit its source instead; the next build overwrites this.').join(' ')
+}
+
+// Files beside this one that differ only by extension.
+//
+// An empty `index.md` sitting next to a real `index.yml` renders to the same
+// destination; whichever renders last wins and the other output is discarded.
+// The destination is not known until the cycle runs, but the COLLIDING SHAPE
+// is visible at write time, and the write is the cheapest moment to say so.
+export async function siblingDestinations(folder, relativePath) {
+    const dir = path.dirname(path.join(folder, relativePath))
+    const base = path.basename(relativePath, path.extname(relativePath))
+    try {
+        const entries = await readdir(dir, { withFileTypes: true })
+        return entries
+            .filter(e => e.isFile()
+                && path.basename(e.name, path.extname(e.name)) === base
+                && e.name !== path.basename(relativePath))
+            .map(e => ({
+                path: path.join(path.dirname(relativePath), e.name),
+                note: 'same name, different extension — may render to the same destination',
+            }))
+    } catch {
+        return []
+    }
+}
+
+// The checksum of a file, or null when there is nothing there. Not an error
+// path: "does not exist yet" is the normal case for a create.
+async function fileChecksum(uri) {
+    try {
+        return await checksum(uri)
+    } catch {
+        return null
+    }
+}
+
+// The catalog entity written from this file, when there is one.
+async function findEntityAtUri(uri) {
+    if (!uri) return null
+    const matches = await findEntities({ uri })
+    return matches?.[0] ?? null
+}
+
+// Resolve a catalog id to the (collection, relativePath) pair a write needs.
+//
+// Taken from the entity rather than by splitting the id: the id prefix is
+// `idPrefix ?? '/' + collection` and the extension may have been stripped, so
+// splitting on the first segment is a guess that is usually right and
+// silently wrong for any source configured either way.
+export async function locateEntityFile(id) {
+    const entity = await readEntity({ id })
+    if (!entity) return { error: `No entity with id ${id}.` }
+    if (!entity.collection) {
+        return { error: `Entity ${id} has no collection, so its file location cannot be derived.` }
+    }
+    let folder
+    try {
+        folder = useCollection(runtime, entity.collection).folder
+    } catch (err) {
+        return { error: `Entity ${id} is in collection ${entity.collection}, which has no folder: ${err.message}` }
+    }
+    if (!entity.uri) {
+        return { error: `Entity ${id} has no uri — it is synthetic (emitted by a plugin, not read from a file) `
+            + 'and has no file to rewrite.' }
+    }
+    const relativePath = path.relative(folder, entity.uri)
+    if (!relativePath || relativePath.startsWith('..') || path.isAbsolute(relativePath)) {
+        return { error: `Entity ${id} lives at ${entity.uri}, outside its collection folder ${folder}.` }
+    }
+    return { collection: entity.collection, relativePath }
+}
+
+// Create or overwrite a source file inside a collection.
+//
+// Takes either `id` (an existing entity) or `collection` + `relativePath`.
+// `content` is the COMPLETE file: anything omitted is deleted.
+//
+// Never throws for an expected outcome. A refusal — a bad id, a path that
+// escapes the collection, a checksum that no longer matches — comes back as
+// `{ ok: false, refused }` with the facts needed to retry, because those are
+// answers rather than faults.
+//
+//   dryRun      write nothing; report what the write would touch
+//   ifChecksum  only write if the file's current DISK checksum equals this
+//   awaitCycle  resolve once the cycle that picks the write up has finished,
+//               with its build report attached
+export async function writeEntitySource({
+    id,
+    collection,
+    relativePath,
+    content = '',
+    ifChecksum,
+    dryRun = false,
+    awaitCycle = false,
+} = {}) {
+    if (id) {
+        const located = await locateEntityFile(id)
+        if (located.error) return { ok: false, refused: 'unresolvable-id', error: located.error }
+        // An explicit pair still wins if a caller passes both, but disagreeing
+        // with the id is a mistake worth refusing rather than silently
+        // resolving one way.
+        if (collection && collection !== located.collection) {
+            return {
+                ok: false,
+                refused: 'collection-mismatch',
+                error: `id ${id} is in collection ${located.collection}, not ${collection}. Pass one or the other.`,
+            }
+        }
+        collection ??= located.collection
+        relativePath ??= located.relativePath
+    }
+    if (!collection || !relativePath) {
+        return {
+            ok: false,
+            refused: 'incomplete-target',
+            error: 'Pass either `id` (for an existing entity) or both `collection` and `relativePath`.',
+        }
+    }
+
+    let handle
+    let uri
+    try {
+        handle = useCollection(runtime, collection)
+        // Containment before anything reads or writes. A relative path that
+        // escapes the collection must not even be STATTED — reporting a
+        // checksum for /etc/passwd is a disclosure on its own.
+        uri = handle.resolveWithin(relativePath)
+    } catch (err) {
+        return { ok: false, refused: 'invalid-target', collection, relativePath, error: err.message }
+    }
+
+    // Everything a caller should know BEFORE the bytes move. Computed for the
+    // dry run and the real write alike, so the preview and the thing it
+    // previews cannot disagree.
+    const existing = id ? await readEntity({ id }) : await findEntityAtUri(uri)
+    const onDisk = await readEntityContent({ uri }, { reload: true })
+    const advisories = contentAdvisories(existing, typeof onDisk.content === 'string' ? onDisk.content : null)
+
+    if (dryRun) {
+        const wouldAffect = existing?.id ? (runtime.manifest?.affectedBy?.(existing) ?? []) : []
+        const touched = new Set(wouldAffect.map(a => a.destination))
+        return {
+            ok: true, dryRun: true, collection, relativePath,
+            id: existing?.id ?? null,
+            exists: existing != null,
+            currentChecksum: await fileChecksum(uri),
+            advisories,
+            warning: advisoryWarning(advisories),
+            wouldAffect,
+            wouldAffectCount: wouldAffect.length,
+            siblingDestinations: await siblingDestinations(handle.folder, relativePath),
+            // Collisions ALREADY standing at the outputs this write would
+            // touch. A write cannot be blamed for them, but re-rendering into
+            // one is how the wrong half of a contested destination wins.
+            collisionsAtAffected: (runtime.manifest?.collisions?.() ?? [])
+                .filter(c => touched.has(c.destination)),
+            note: existing?.id
+                ? 'Destinations are computed with the engine\'s own skip rule, so they match what a real cycle '
+                    + 'would do — EXCEPT for changes to the entity\'s own frontmatter, which is parsed during '
+                    + 'import and can move its destination.'
+                : 'This file is not in the catalog yet, so nothing depends on it and there is no blast radius '
+                    + 'to report.',
+        }
+    }
+
+    // Checked immediately before the write. Not a lock — a writer that lands
+    // between the check and the write still wins — but it closes the window
+    // that matters in practice: read, think, write back a whole file built
+    // from a copy that is now stale.
+    const before = await fileChecksum(uri)
+    if (ifChecksum !== undefined && ifChecksum !== before) {
+        return {
+            ok: false,
+            refused: 'checksum-mismatch',
+            collection, relativePath,
+            expectedChecksum: ifChecksum,
+            currentChecksum: before,
+            hint: before === null
+                ? 'The file does not exist. Omit ifChecksum to create it.'
+                : 'The file on disk changed since you read it. Re-read it for the CONTENT, re-apply your change, '
+                    + 'and retry with `currentChecksum` from THIS response.',
+        }
+    }
+
+    const cycleId = nextCycleId()
+    await handle.write(relativePath, content)
+
+    const result = {
+        ok: true, collection, relativePath,
+        checksum: await fileChecksum(uri),
+        bytes: Buffer.byteLength(content),
+        cycleId,
+        siblingDestinations: await siblingDestinations(handle.folder, relativePath),
+    }
+    // Echoed on the way out, not only on read. A caller that never read the
+    // file — or read past the header — is exactly the one that needs telling,
+    // and telling it after the write still names what to check before deploy.
+    if (advisories.length) {
+        result.advisories = advisories
+        result.warning = advisoryWarning(advisories)
+    }
+    if (awaitCycle) result.report = await whenCycleCompletes(cycleId)
+    return result
+}