mikser-io 9.43.1 → 9.45.0

package/docs/api-reference.md

@@ -538,6 +538,27 @@ write back a whole file built from a copy that is now stale.
 catalog's, which lags between builds — pass `diskChecksum`, or the
 `currentChecksum` a refusal hands back.
 
+### `deleteEntitySource(options)`
+
+Removes a source file with the same guards, plus the one only removal needs.
+
+```js
+const preview = await deleteEntitySource({ id: '/documents/old.md', dryRun: true })
+// preview.referencedBy → [{ id: '/documents/page.md', field: 'hero' }]
+```
+
+Takes the same addressing, `ifChecksum` and `dryRun` as the write. `referencedBy`
+reports everything still pointing at the entity — asked of the reference index
+through `lookupKeys`, so a ref by served path counts exactly as invalidation
+counts it, not just a ref by id. A delete is the one write with no content to
+inspect afterwards, which makes the preview the only chance to see its cost.
+
+### Change sets
+
+Both take `changeSet` and `summary`, grouping several writes into one unit a
+consumer can commit together and later take back together. See
+[Change sets](#change-sets-1) below.
+
 ### Advisories
 
 `contentAdvisories(entity, content)` names files a caller must not edit blind,
@@ -556,6 +577,46 @@ extension, which may render to the same destination.
 or `{ error }` — taken from the entity rather than by splitting the id, since the
 prefix is configurable and the extension may have been stripped.
 
+## Change sets
+
+Which writes belong together, and who asked for them.
+
+The engine does not otherwise care who wrote a file — a write is a write. That
+holds until something wants to undo one request without touching everything
+around it, and then it is the wrong resolution: an agent's three edits and a
+document created through the API a second later are indistinguishable, so
+removing one removes the other.
+
+```js
+await writeEntitySource({ id, content, changeSet: 'req-42', summary: 'Rewrite the hero copy' })
+await deleteEntitySource({ id: other, changeSet: 'req-42' })
+
+pendingChangeSets()
+// [{ id: 'req-42', summary: 'Rewrite the hero copy', paths: [...], deletions: [...] }]
+clearChangeSets(['req-42'])
+```
+
+| Export | Does |
+| --- | --- |
+| `recordChangeSetWrite({ changeSet, summary, principal, uri, operation, undoOf })` | attach one path to a set |
+| `pendingChangeSets()` | sets with unconsumed writes, oldest first |
+| `clearChangeSets(ids)` | drop what a consumer has committed |
+
+Paths come back repo-relative and POSIX-separated, ready for a git pathspec.
+
+**Not a transaction.** Nothing is held back, nothing rolls back on failure, and
+a half-finished set is a real set containing what actually landed. It is a
+label on work that already happened, which is what makes it safe on a write
+path that must never block. A path is claimed only after the write succeeds —
+claiming on intent would make a write that never happened undoable, and undoing
+it would delete whatever is actually at that path.
+
+**Unclaimed writes stay unclaimed.** A consumer must still handle them; they
+happened, and losing them would be worse than not attributing them.
+`mikser-io-git` commits claimed sets to their own scoped commits and sweeps the
+rest into unattributed ones, which is what lets it offer undo for the first and
+not the second.
+
 ## Search
 
 `queryEntities` sifts **meta**. `searchEntities` answers the other question —

package/index.js

@@ -13,6 +13,7 @@ export * from './src/journal.js'
 export * from './src/catalog.js'
 export * from './src/search.js'
 export * from './src/write.js'
+export * from './src/changeset.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": "9.43.1",
+  "version": "9.45.0",
   "description": "<p align=\"center\">   <img src=\"mikser-lockup-stacked.svg\" alt=\"mikser\" width=\"198\" /> </p>",
   "main": "index.js",
   "exports": {

package/src/changeset.js

@@ -0,0 +1,108 @@
+// Which writes belong together, and who asked for them.
+//
+// The engine deliberately does not care who wrote a file — files are the
+// source of truth and a write is a write. That holds right up until something
+// wants to UNDO one request without touching everything else that happened
+// around it, and then "a write is a write" is exactly the wrong resolution:
+// an agent's three edits and a document created through the API in the same
+// second are indistinguishable, so removing one removes the other.
+//
+// A change set is the missing grain. The caller names it, the writes
+// accumulate under it, and a consumer — mikser-io-git today — can commit
+// exactly those paths and later remove exactly that contribution.
+//
+// Deliberately NOT a transaction. Nothing is held back, nothing rolls back on
+// failure, and a half-finished set is a real set containing what actually
+// landed. It is a label on work that already happened, which is what makes it
+// safe to add to a write path that must never block on it.
+//
+// Unclaimed writes stay unclaimed. A consumer is expected to handle them —
+// they still happened, and losing them would be worse than not being able to
+// attribute them.
+
+import path from 'node:path'
+import runtime from './runtime.js'
+
+function store() {
+    runtime.changeSets ??= new Map()
+    return runtime.changeSets
+}
+
+// Repo-relative, POSIX-separated: these end up in a git pathspec, and a
+// consumer should not have to redo that conversion or guess the root.
+function relativeToWorkingFolder(uri) {
+    const root = runtime.options?.workingFolder
+    if (!root || !uri) return null
+    const rel = path.relative(root, uri)
+    if (!rel || rel.startsWith('..') || path.isAbsolute(rel)) return null
+    return rel.split(path.sep).join('/')
+}
+
+// Attach one write to a change set.
+//
+// `summary` is the caller's own description of what it is doing, kept because
+// nothing downstream will ever know it as well — a reader choosing what to
+// undo needs "changed the hero text on the devices page", not a file count.
+// First one wins: later writes in the same set are the same request.
+export function recordChangeSetWrite({ changeSet, summary, principal, uri, operation = 'write', undoOf } = {}) {
+    if (!changeSet || !uri) return null
+    const rel = relativeToWorkingFolder(uri)
+    // Outside the working folder there is nothing a repo-scoped consumer can
+    // do with the path, and silently keeping an absolute one would produce a
+    // pathspec that matches nothing.
+    if (!rel) return null
+
+    const sets = store()
+    let set = sets.get(changeSet)
+    if (!set) {
+        set = {
+            id: changeSet,
+            summary: summary ?? null,
+            principal: principal ?? null,
+            // Set when this change set exists to take another one back, so
+            // the undo is itself an ordinary, undoable change rather than a
+            // special history-rewriting operation.
+            undoOf: undoOf ?? null,
+            startedAt: Date.now(),
+            paths: new Map(),
+        }
+        sets.set(changeSet, set)
+    }
+    if (!set.summary && summary) set.summary = summary
+    if (!set.principal && principal) set.principal = principal
+    if (!set.undoOf && undoOf) set.undoOf = undoOf
+    set.paths.set(rel, operation)
+    return set.id
+}
+
+// Every set with writes not yet consumed, oldest first — the order a consumer
+// should commit them in, so history reads the way the work happened.
+export function pendingChangeSets() {
+    return [...store().values()]
+        .filter(set => set.paths.size)
+        .sort((a, b) => a.startedAt - b.startedAt)
+        .map(set => ({
+            id: set.id,
+            summary: set.summary,
+            principal: set.principal,
+            undoOf: set.undoOf,
+            startedAt: set.startedAt,
+            paths: [...set.paths.keys()],
+            deletions: [...set.paths.entries()].filter(([, op]) => op === 'delete').map(([p]) => p),
+        }))
+}
+
+// Drop sets a consumer has dealt with.
+//
+// Called after the paths are committed, not after they are written: a crash in
+// between loses the attribution but not the work, which then reaches the
+// consumer as an unclaimed write. That is the right way round — attribution is
+// a convenience, the bytes are not.
+export function clearChangeSets(ids = []) {
+    const sets = store()
+    for (const id of ids) sets.delete(id)
+}
+
+export function forgetAllChangeSets() {
+    store().clear()
+}

package/src/write.js

@@ -22,8 +22,9 @@ 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 { useCollection, checksum, readEntityContent, lookupKeys } from './utils.js'
 import { nextCycleId, whenCycleCompletes } from './report.js'
+import { recordChangeSetWrite } from './changeset.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.
@@ -167,6 +168,9 @@ export async function writeEntitySource({
     ifChecksum,
     dryRun = false,
     awaitCycle = false,
+    changeSet,
+    summary,
+    principal,
 } = {}) {
     if (id) {
         const located = await locateEntityFile(id)
@@ -260,11 +264,17 @@ export async function writeEntitySource({
     const cycleId = nextCycleId()
     await handle.write(relativePath, content)
 
+    // AFTER the write, so a set only ever claims paths that actually moved.
+    // Claiming on intent would make a failed write undoable, and undoing a
+    // write that never happened is a way to delete someone else's file.
+    if (changeSet) recordChangeSetWrite({ changeSet, summary, principal, uri })
+
     const result = {
         ok: true, collection, relativePath,
         checksum: await fileChecksum(uri),
         bytes: Buffer.byteLength(content),
         cycleId,
+        ...(changeSet ? { changeSet } : {}),
         siblingDestinations: await siblingDestinations(handle.folder, relativePath),
     }
     // Echoed on the way out, not only on read. A caller that never read the
@@ -277,3 +287,131 @@ export async function writeEntitySource({
     if (awaitCycle) result.report = await whenCycleCompletes(cycleId)
     return result
 }
+
+// Remove a source file, with the same checks the write has plus the one that
+// only matters for removal: what still points at it.
+//
+// Delete is the more destructive operation and had the fewest guards — no
+// precondition, no preview, no advisory, and nothing at all about the
+// references that break when an entity other documents point at disappears.
+//
+// Same contract as writeEntitySource: never throws for an expected outcome.
+export async function deleteEntitySource({
+    id,
+    collection,
+    relativePath,
+    ifChecksum,
+    dryRun = false,
+    changeSet,
+    summary,
+    principal,
+} = {}) {
+    if (id) {
+        const located = await locateEntityFile(id)
+        if (located.error) return { ok: false, refused: 'unresolvable-id', error: located.error }
+        if (collection && collection !== located.collection) {
+            return {
+                ok: false,
+                refused: 'collection-mismatch',
+                error: `id ${id} is in collection ${located.collection}, not ${collection}. Pass one or the other.`,
+            }
+        }
+        collection ??= located.collection
+        relativePath ??= located.relativePath
+    }
+    if (!collection || !relativePath) {
+        return {
+            ok: false,
+            refused: 'incomplete-target',
+            error: 'Pass either `id` (for an existing entity) or both `collection` and `relativePath`.',
+        }
+    }
+
+    let handle
+    let uri
+    try {
+        handle = useCollection(runtime, collection)
+        uri = handle.resolveWithin(relativePath)
+    } catch (err) {
+        return { ok: false, refused: 'invalid-target', collection, relativePath, error: err.message }
+    }
+
+    const existing = id ? await readEntity({ id }) : await findEntityAtUri(uri)
+    const currentChecksum = await fileChecksum(uri)
+    if (currentChecksum === null) {
+        return {
+            ok: false, refused: 'not-found', collection, relativePath,
+            error: 'Nothing on disk at that path.',
+        }
+    }
+
+    // Everything that would be left pointing at nothing. Asked of the
+    // reference index rather than of the text, so it sees a `$`-ref by served
+    // path exactly the way invalidation does.
+    const referencedBy = existing ? referrersOf(existing) : []
+    const wouldAffect = existing?.id ? (runtime.manifest?.affectedBy?.(existing) ?? []) : []
+
+    if (dryRun) {
+        return {
+            ok: true, dryRun: true, collection, relativePath,
+            id: existing?.id ?? null,
+            exists: true,
+            currentChecksum,
+            referencedBy,
+            wouldAffect,
+            wouldAffectCount: wouldAffect.length,
+            ...(referencedBy.length ? {
+                warning: `${referencedBy.length} entit${referencedBy.length === 1 ? 'y' : 'ies'} reference this. `
+                    + 'Deleting it leaves those references pointing at nothing.',
+            } : {}),
+        }
+    }
+
+    if (ifChecksum !== undefined && ifChecksum !== currentChecksum) {
+        return {
+            ok: false,
+            refused: 'checksum-mismatch',
+            collection, relativePath,
+            expectedChecksum: ifChecksum,
+            currentChecksum,
+            hint: 'The file changed since you read it. Re-read it before deciding to delete it, and retry with '
+                + '`currentChecksum` from THIS response.',
+        }
+    }
+
+    await handle.remove(relativePath)
+    if (changeSet) recordChangeSetWrite({ changeSet, summary, principal, uri, operation: 'delete' })
+
+    return {
+        ok: true, collection, relativePath,
+        id: existing?.id ?? null,
+        deletedChecksum: currentChecksum,
+        cycleId: nextCycleId(),
+        ...(changeSet ? { changeSet } : {}),
+        referencedBy,
+        ...(referencedBy.length ? {
+            warning: `${referencedBy.length} entit${referencedBy.length === 1 ? 'y' : 'ies'} still reference this.`,
+        } : {}),
+    }
+}
+
+// Who points at this entity, by any of the keys it can be referenced by.
+//
+// `lookupKeys` rather than the id alone: content refers to served paths far
+// more often than to catalog ids, and asking only about the id would report a
+// widely-linked image as unreferenced.
+function referrersOf(entity) {
+    const refs = runtime.refs
+    if (!refs?.inboundFor) return []
+    const found = new Map()
+    for (const key of [entity.id, ...(lookupKeys(entity) ?? [])]) {
+        if (!key) continue
+        let inbound = []
+        try { inbound = refs.inboundFor(key) ?? [] } catch { continue }
+        for (const ref of inbound) {
+            if (!ref?.id || ref.id === entity.id) continue
+            found.set(ref.id, { id: ref.id, ...(ref.field ? { field: ref.field } : {}) })
+        }
+    }
+    return [...found.values()]
+}