feat(workspace): mirror memory blocks to the bound workspace

packages/opencode/src/altimate/training/store.ts

@@ -1,6 +1,7 @@
 // altimate_change - Training store wrapping MemoryStore for learned knowledge
 import { MemoryStore, type MemoryBlock } from "../../memory"
 import {
+  TRAINING_META_COMMENT,
   TRAINING_TAG,
   TRAINING_MAX_PATTERNS_PER_KIND,
   TrainingKind,
@@ -164,5 +165,5 @@ export namespace TrainingStore {
 }
 
 function stripTrainingMeta(content: string): string {
-  return content.replace(/^<!--\s*training\n[\s\S]*?-->\n*/, "").trim()
+  return content.replace(TRAINING_META_COMMENT, "").trim()
 }

packages/opencode/src/altimate/training/types.ts

@@ -2,6 +2,12 @@
 import z from "zod"
 
 export const TRAINING_TAG = "training"
+
+/** Matches the metadata comment ``embedTrainingMeta`` writes at the head of a
+ * training block's content. Exported so readers that normalise it away — the
+ * workspace mirror hashes content without it, because the applied counter is
+ * rewritten every session — cannot drift from the writer. */
+export const TRAINING_META_COMMENT = /^<!--\s*training\n[\s\S]*?-->\n*/
 export const TRAINING_ID_PREFIX = "training"
 // altimate_change start — increase training limits for enterprise teams
 // 20 entries per kind is too restrictive for teams with 200+ dbt models spanning
@@ -70,6 +76,6 @@ export function embedTrainingMeta(content: string, meta: TrainingBlockMeta): str
     "-->",
   ].join("\n")
   // Strip existing training meta block if present
-  const stripped = content.replace(/^<!--\s*training\n[\s\S]*?-->\n*/, "")
+  const stripped = content.replace(TRAINING_META_COMMENT, "")
   return header + "\n" + stripped
 }

packages/opencode/src/altimate/workspace/api-client.ts

@@ -17,6 +17,10 @@ const REQUEST_TIMEOUT_MS = 15_000
 export interface DatamateRef {
   id: number
   name: string
+  /** Whether the workspace has memory switched on. Surfaced as a user-facing
+   * toggle in the workspace app, so callers that write memory must respect it.
+   * Undefined when the backend omitted the field. */
+  memoryEnabled?: boolean
 }
 
 export interface Binding {
@@ -237,6 +241,13 @@ async function req<T>(
   return json as T
 }
 
+/** Shared wire helper for sibling Altimate routers. Exported so callers that
+ * need the same credential resolution, abort budget and typed-error mapping do
+ * not duplicate any of it — see ./memory-api.ts, which drives
+ * ``/datamates/memory/*`` through this exact path. Always pass an explicit
+ * ``base``; the default is this module's own namespace. */
+export { req as altimateRequest }
+
 export namespace WorkspaceApi {
   /** Server-authoritative pre-check by git remote. Returns null on 404. */
   export async function getBindingForRemote(remote: string): Promise<GetBindingResponse | null> {
@@ -350,7 +361,7 @@ export namespace WorkspaceApi {
     // bare ``[...]``, and a generic ``{data: [...]}`` — so a backend
     // contract change (or compat layer) doesn't silently empty the picker.
     // (cubic-dev-ai round 3.)
-    type Row = { id: number | string; name: string }
+    type Row = { id: number | string; name: string; memory_enabled?: boolean }
     const body = await req<Row[] | { datamates?: Row[]; data?: Row[] }>("GET", "/", {
       base: "/datamates",
     })
@@ -378,7 +389,7 @@ export namespace WorkspaceApi {
     // per-element rather than per-envelope malformed value.
     return rows
       .filter((d): d is Row => d !== null && typeof d === "object")
-      .map((d) => ({ id: Number(d.id), name: d.name }))
+      .map((d) => ({ id: Number(d.id), name: d.name, memoryEnabled: d.memory_enabled }))
       .filter((d) => Number.isInteger(d.id) && d.id > 0 && typeof d.name === "string")
   }
 }

packages/opencode/src/altimate/workspace/memory-api.ts

@@ -0,0 +1,177 @@
+// altimate_change - new file
+//
+// Wire client for the memory routes in altimate-backend (/datamates/memory/*).
+// Those routes proxy to a Lambda which runs an LLM extractor before storing, so
+// the behaviour here is shaped by two facts established against the live
+// service rather than inferred:
+//
+//   * A create runs the extractor. It rewrites the submitted text and
+//     overwrites the ``memory_type`` and ``title`` metadata keys, and it can
+//     decline the content entirely and store nothing while still answering 200.
+//     ``update`` does NOT run the extractor — it writes text and metadata
+//     verbatim. Every create is therefore followed by an update that restores
+//     the block exactly as the user wrote it.
+//   * ``extracted`` is not a stored/not-stored signal. It is false both when
+//     the extractor declined (no ``result``) and when the extractor errored and
+//     the raw messages were stored as a fallback (``result`` present, content
+//     stored). Presence of ``result`` is the only reliable signal, which is why
+//     ``add`` reports an id rather than a boolean.
+//
+// Records are tagged with ``metadata.source`` so the backend can keep them out
+// of ordinary Datamate reads; this client opts back in explicitly on every read.
+import { altimateRequest } from "./api-client"
+
+/** Stamped on every record this CLI writes, and the value the backend filters
+ * on. Reads must opt in by name or they come back empty. */
+export const MIRROR_SOURCE = "altimate-code"
+
+const BASE = "/datamates/memory"
+
+/** Upper bound on records requested per read. The service does not honour
+ * paging parameters — a repeated request re-runs the identical query — so this
+ * is a single bounded fetch rather than the first page of several. */
+export const LIST_LIMIT = 200
+
+/** A record as returned by ``/list``. */
+export interface CloudMemoryRecord {
+  id: string
+  memory: string
+  created_at?: string
+  updated_at?: string
+  metadata?: Record<string, unknown> | null
+}
+
+/** Metadata written on every mirrored record. Names and types are fixed so
+ * that lookup and filtering agree across independent call sites. */
+export interface MirrorMetadata {
+  source: typeof MIRROR_SOURCE
+  block_id: string
+  block_scope: "global" | "project"
+  /** Always "private" in v0. Written so a later sharing feature can promote a
+   * record without a backfill; nothing reads it today. */
+  visibility: "private"
+  block_created: string
+  block_updated: string
+  /** Absent for global blocks — that is what makes them span workspaces. */
+  datamate_id?: string
+  datamate_name?: string
+  /** Project provenance and the cross-machine convergence key. */
+  repo_remote?: string
+  project_path?: string
+  block_tags?: string
+  /** ISO-8601. Mirrored so a TTL'd block expires everywhere rather than living
+   * forever in the workspace once it has left the machine that wrote it. */
+  block_expires?: string
+  archived?: "true"
+  archived_at?: string
+}
+
+export function isMirrorRecord(record: CloudMemoryRecord): boolean {
+  const meta = record.metadata
+  if (!meta || typeof meta !== "object") return false
+  return meta.source === MIRROR_SOURCE
+}
+
+export function isArchived(record: CloudMemoryRecord): boolean {
+  return record.metadata?.archived === "true"
+}
+
+/** Pull every created record id out of a create response.
+ *
+ * A create runs an extractor server-side, and an extractor is free to split one
+ * submission into several records — each carrying the metadata we sent, so each
+ * looks like our block. Returning only the first would leave the rest holding
+ * rewritten text, never repaired, never indexed and never archived, while the
+ * read path injected all of them under one block id.
+ *
+ * The payload shape is not a published contract, so this accepts the observed
+ * forms rather than guessing: a bare array, or an object wrapping one under
+ * ``results``/``memories``/``data``. An empty result is the expected outcome
+ * when the extractor declines the content. */
+export function extractRecordIds(result: unknown): string[] {
+  const rows = (() => {
+    if (Array.isArray(result)) return result
+    if (result && typeof result === "object") {
+      const obj = result as Record<string, unknown>
+      for (const key of ["results", "memories", "data"]) {
+        if (Array.isArray(obj[key])) return obj[key] as unknown[]
+      }
+    }
+    return []
+  })()
+
+  const ids: string[] = []
+  for (const row of rows) {
+    if (!row || typeof row !== "object") continue
+    const id = (row as Record<string, unknown>).id ?? (row as Record<string, unknown>).memory_id
+    if (typeof id === "string" && id) ids.push(id)
+  }
+  return ids
+}
+
+/** Convenience for callers that only need to know whether anything was stored. */
+export function extractRecordId(result: unknown): string | undefined {
+  return extractRecordIds(result)[0]
+}
+
+export namespace MemoryApi {
+  /** Create a record and report the ids it produced.
+   *
+   * Returns an empty array when the service stored nothing. That is not an error —
+   * the extractor declines content it judges unremarkable — so the caller
+   * should leave the block unindexed and let a later edit retry, rather than
+   * treating it as a failure. */
+  export async function add(content: string, metadata: MirrorMetadata): Promise<string[]> {
+    const res = await altimateRequest<{ message?: string; result?: unknown }>("POST", "/", {
+      base: BASE,
+      allowEmptyBody: true,
+      body: {
+        messages: [{ role: "user", content }],
+        memory_options: { metadata },
+      },
+    })
+    return extractRecordIds(res?.result)
+  }
+
+  /** Overwrite a record verbatim. Does not run the extractor, and replaces the
+   * metadata dict wholesale, so callers must pass the complete metadata. */
+  export async function update(
+    memoryId: string,
+    content: string,
+    metadata: MirrorMetadata,
+  ): Promise<void> {
+    await altimateRequest<{ message?: string }>("PATCH", `/${encodeURIComponent(memoryId)}`, {
+      base: BASE,
+      allowEmptyBody: true,
+      body: { memory: content, metadata },
+    })
+  }
+
+  /** Read this user's mirrored records.
+   *
+   * ``include_sources`` is required: the backend excludes this client's records
+   * from list/search by default so they do not surface in Datamate sessions.
+   * No workspace filter is sent — the service's own query for a caller's
+   * records is not scoped by workspace, so narrowing happens in the caller.
+   *
+   * Deliberately NOT capped at ``LIST_LIMIT``. The service ignores paging, so a
+   * cap here would discard real records before anything could rank them, and
+   * the user would lose memory with no signal. Session context is already
+   * bounded downstream: ``MemoryPrompt.inject`` scores every block and appends
+   * only while it fits the caller's budget. ``LIST_LIMIT`` is used to recognise
+   * a possibly-cut-short read (see ``fetchKnownRecords``), not to trim one. */
+  export async function list(): Promise<CloudMemoryRecord[]> {
+    const rows = await altimateRequest<CloudMemoryRecord[] | { memories?: CloudMemoryRecord[] }>(
+      "GET",
+      "/list",
+      {
+        base: BASE,
+        allowEmptyBody: true,
+        query: { include_sources: MIRROR_SOURCE, page_size: String(LIST_LIMIT) },
+      },
+    )
+    if (!rows) return []
+    if (Array.isArray(rows)) return rows
+    return Array.isArray(rows.memories) ? rows.memories : []
+  }
+}

packages/opencode/src/altimate/workspace/memory-backfill.ts

@@ -0,0 +1,51 @@
+// altimate_change - new file
+//
+// Seeds a freshly bound workspace with the memory this machine already holds.
+// Without it only blocks written AFTER the bind would ever reach the store, and
+// a user's existing memory would stay invisible in the workspace.
+//
+// Lives in its own module rather than inside ./state.ts because the sweep needs
+// MemoryStore, whose write path already reaches ./memory-sync — importing it
+// directly from state.ts would close an eval-order cycle
+// (state -> backfill -> memory -> store -> memory-sync -> state). state.ts
+// reaches this through a lazy dynamic import instead.
+import { MemoryStore } from "@/memory/store"
+import { Log } from "@/altimate/util/log"
+import { backfill, isEnabled } from "./memory-sync"
+import type { CachedBinding } from "./state"
+
+const log = Log.create({ service: "altimate-workspace-memory-backfill" })
+
+/** Push every non-expired local block. Throttled and resumable inside
+ * ``backfill`` — blocks already synced at their current payload are skipped, so
+ * repeated binds cost index reads rather than uploads.
+ *
+ * Covers both scopes: project blocks attach to the workspace just bound, and
+ * global blocks go up account-level. A bind is the only moment global memory is
+ * swept; blocks written later ride the ordinary per-write mirror. */
+export async function backfillOnBind(directory: string, binding: CachedBinding): Promise<boolean> {
+  if (!isEnabled()) return false
+  try {
+    // The directory and binding are passed in rather than rediscovered. The
+    // `link` subcommand binds from a plain yargs handler with no instance
+    // context, so resolving project scope from the ambient instance throws
+    // there — silently, because this catch turns it into a log line while the
+    // CLI still prints "Linked". Reading project memory was the entire point.
+    const blocks = await MemoryStore.listAll({ directory })
+    if (blocks.length === 0) return true
+    const result = await backfill(blocks, binding)
+    log.info("workspace memory seeded after bind", result)
+    // Only a sweep that stored everything it meant to counts as seeded. A
+    // failure here must leave the binding eligible for a retry, or local blocks
+    // stay absent from the workspace until a rebind or an unrelated edit.
+    // ``declined`` counts too — a block the service explicitly refused (quota,
+    // permissions) is still absent from the workspace and the binding should
+    // stay unseeded so a later rebind retries it. Without this a partially-
+    // rejected backfill left the binding treated as fully seeded. (altimate-
+    // harness-bot #1116 comment 3840503346.)
+    return !result.gated && result.failed === 0 && result.declined === 0
+  } catch (err) {
+    log.warn("workspace memory backfill after bind failed", { err: String(err) })
+    return false
+  }
+}