@minhpnq1807/contextos 0.5.45 → 0.5.49

package/CHANGELOG.md

@@ -1,5 +1,31 @@
 # Changelog
 
+## 0.5.49
+
+- **Cold-cache MCP smoke:** `npm run test:mcp` now verifies the MCP tool contract without requiring a pre-downloaded ContextOS embedding model. When the model exists it still runs the semantic/performance smoke; otherwise it asserts cold-cache fallback behavior so publish CI does not fail before model warmup.
+
+## 0.5.48
+
+- **Publish-safe embedding warm:** `ctx embeddings warm` now skips remote model downloads by default in CI and returns best-effort warm results when local embedding model loading fails, preventing npm publish workflows from failing on Hugging Face 429/rate-limit errors.
+- **CI graph embedding skip:** Code-review-graph embedding refresh is skipped when remote embeddings are disabled, avoiding extra Hugging Face calls during publish jobs.
+
+## 0.5.47
+
+- **Semantic-only skill discovery:** Replaced taxonomy/keyword skill ranking with fused semantic retrieval. ContextOS now embeds `prompt + project profile` and compares it against cached skill vectors, so mixed Vietnamese/English prompts can use package, language, and recent-file signals without hard-coded categories.
+- **Project profile cache:** Added a bounded project profiler that builds an embeddable string from root/workspace `package.json` dependencies, scripts, detected languages, and recent git files. The cache invalidates when package metadata or git `HEAD` changes.
+- **Skill dedupe and activation format:** Skill catalogs are deduplicated by normalized skill name before indexing/searching, preferring project and Codex roots. Prompt output now renders skill activations as `$skill-name`.
+- **Refresh rebuilds skill index:** `ctx install` and `ctx refresh` now rebuild the skill embedding index by default because prompt-time skill discovery no longer falls back to taxonomy or keyword ranking for large catalogs.
+- **Skill hook timeout split:** Prompt hooks now use a separate skill embedding timeout, so large indexed skill catalogs can still produce suggestions after direct fallback without slowing rule scoring.
+- **Read-only embedding cache close:** Prompt-time embedding reads no longer rewrite the whole `embeddings.db` when no cache entries changed.
+
+## 0.5.46
+
+- **Configurable prompt suggestion limits:** `ctx --config` and interactive `ctx setup` now let users choose how many suggested files, skills, and workflows appear in prompt context. Defaults are five each, with caps of 20 files, 10 skills, and 5 workflows.
+- **Limit-aware prompt hooks and debug:** `UserPromptSubmit` hooks, direct fallback scoring, the private `ctx-mcp` bridge request, and `ctx debug` now all honor the saved suggestion limits instead of using hard-coded counts.
+- **Document authoring skill intent:** Prompts that create, edit, update, or maintain documents, workspace docs, README files, wiki pages, manuals, guides, specs, or ADRs now prioritize documentation skills such as `doc-coauthoring`, `documentation`, `docs-architect`, `readme`, and `wiki-page-writer`.
+- **Safer document skill gating:** Document-processing and workspace-automation skills such as Azure Document Intelligence, DocuSign, Asana, Slack, Google Docs, and Notion no longer win generic document-writing prompts unless the provider or processing task is explicitly named.
+- **Setup summary clarity:** The setup wizard summary now reports the saved prompt suggestion limits alongside the enabled prompt sections so users can review output volume immediately.
+
 ## 0.5.45
 
 - **Project-aware MCP skill suggestions:** Skill ranking now reads `package.json` keywords and dependencies such as `@modelcontextprotocol/sdk`. MCP projects can recommend `mcp-builder`, `mcp-management`, `mcp-tool-developer`, and `agent-memory-mcp` for context retrieval, scorer, hook, and prompt-injection debugging tasks even when the prompt does not explicitly say `mcp`.

package/README.md

@@ -444,8 +444,8 @@ This warning comes from a transitive dependency in the local embedding/WASM stac
 | `ctx sync --workflows --dry-run` | Previews workflow sync without writing files. | You want to inspect source workflows and target roots first. | Prints planned sync/index output and skips copying target files. |
 | `ctx skills` | Installs community skill libraries. | You want curated skills without running the full setup wizard. | Opens the community installer, uses a portable shell on Windows/Linux/macOS, repairs unsafe skill symlinks, and syncs installed skills to selected agents. |
 | `ctx embeddings warm -- "task"` | Prepares local semantic embedding caches. | First install, CI smoke checks, or after changing AGENTS.md/project files/skills/workflows. | Loads/downloads `Xenova/all-MiniLM-L6-v2` and writes rule, file-path, skill, and workflow vectors to `~/.ctx/contextos/embeddings.db`. |
-| `ctx --config` | Opens an interactive multi-select panel for prompt sections. | You want to reduce ContextOS prompt output noise. | Toggles critical rules, suggested files, suggested skills, and suggested workflows globally under `~/.ctx/contextos/output-config.json`. |
-| `ctx refresh` | Refreshes the active Codex marketplace plugin and rebuilds local indexes. | Local development updates or a stale file retrieval index. | Copies the current package to `$CODEX_HOME/marketplaces/contextos`, rebuilds file-path embeddings and import adjacency, and refreshes code-review-graph embeddings when available. |
+| `ctx --config` | Opens an interactive panel for prompt sections and suggestion limits. | You want to reduce ContextOS prompt output noise. | Toggles critical rules, suggested files, suggested skills, and suggested workflows globally under `~/.ctx/contextos/output-config.json`, then lets you set suggestion counts for files, skills, and workflows. |
+| `ctx refresh` | Refreshes the active Codex marketplace plugin and rebuilds local indexes. | Local development updates or stale file/skill retrieval indexes. | Copies the current package to `$CODEX_HOME/marketplaces/contextos`, rebuilds file-path embeddings, skill embeddings, import adjacency, and refreshes code-review-graph embeddings when available. |
 | `ctx ruler -- <args>` | Forwards args to the installed `ruler` CLI. | You need native Ruler commands such as `init`, `apply`, or `revert`. | Preserves Ruler stdout/stderr and exit status. |
 | `ctx skillshare -- <args>` | Forwards args to the installed `skillshare` CLI. | You need native skillshare commands such as `status`, `target list`, `doctor`, `push`, or `pull`. | Preserves skillshare stdout/stderr and exit status. |
 | `ctx --version` | Prints the installed ContextOS CLI version. | You want to confirm which npm version is being executed. | Prints the version from package metadata. |
@@ -511,17 +511,23 @@ This keeps the hook fast and local while still using graph semantics when availa
 
 Prompt scoring does not walk the repository for file candidates or import expansion. `ctx install` and `ctx embeddings warm` rebuild the persisted file-vector index and one-hop import adjacency index by walking source paths once; prompt hooks query those indexes directly. Rules, files, skills, and workflows are scored concurrently with `Promise.all()`.
 
-`ctx embeddings warm` automatically refreshes the active Codex marketplace payload before rebuilding indexes. Use `ctx refresh` when you want the same marketplace sync plus install-style file/import index and code-review-graph embedding refresh in one command.
+`ctx embeddings warm` automatically refreshes the active Codex marketplace payload before rebuilding indexes. Use `ctx refresh` when you want the same marketplace sync plus install-style file, skill, import, and code-review-graph embedding refresh in one command.
 
 If a prompt has no usable context candidates, the hook fails open without emitting an empty `hook context` block, records `emptyContextReason` in the workspace runtime file, and starts a detached `autowarm` rebuild with a cooldown. That background rebuild refreshes file vectors, skill/workflow vectors, import adjacency, and available code-review-graph node embeddings for the next prompt while keeping repository walking out of the current prompt hot path.
 
-Use `ctx --config` to choose which prompt sections ContextOS injects. Interactive `ctx setup` now includes the same multi-select step, while `ctx setup --yes` keeps the current saved config for automation. The panel supports multiple selection with `Space` and persists the global choice in `~/.ctx/contextos/output-config.json`. Disabling rules hides both critical and additional relevant rule sections; compliance metadata remains available for reports.
+Use `ctx --config` to choose which prompt sections ContextOS injects and how many suggestions each section may show. Interactive `ctx setup` includes the same section picker and limit prompts, while `ctx setup --yes` keeps the current saved config for automation. The panel supports multiple selection with `Space` and persists the global choice in `~/.ctx/contextos/output-config.json`. Defaults are five suggested files, five skills, and five workflows; caps are 20 files, 10 skills, and 5 workflows. Disabling rules hides both critical and additional relevant rule sections; compliance metadata remains available for reports.
 
-Injected prompt sections are intentionally compact: rules show only detected rule text, files show basenames without paths, skills show unique names as a comma-separated inline list without descriptions, and workflows show names with their agent chain. Stop hooks persist reports silently; run `ctx report` or `ctx evidence` when you want the detailed compliance output.
+Injected prompt sections are intentionally compact: rules show only detected rule text, files show a comma-separated inline list of basenames without paths, skills show unique `$skill-name` activations as a comma-separated inline list without descriptions, and workflows show names with their agent chain. Stop hooks persist reports silently; run `ctx report` or `ctx evidence` when you want the detailed compliance output.
 
 Codex may flatten newlines in its `UserPromptSubmit hook (completed)` preview. The injected `additionalContext` payload remains multiline; this is a Codex preview display limitation.
 
-Skill ranking uses bounded project hints from root/workspace `package.json` files and known mobile config files such as `app.json`, `app.config.*`, and `eas.json`. This lets Expo/EAS tasks activate specialized skills without walking the source tree on every prompt.
+Skill ranking is semantic-only. ContextOS builds a fused query from the user prompt plus a cached project profile, then compares that vector with cached skill vectors:
+
+```text
+embed(prompt + project profile) -> cosine -> embed(skill name + description)
+```
+
+The project profile is an embeddable string built from bounded root/workspace `package.json` metadata, dependencies, scripts, detected languages, and recent git files. It is cached under the ContextOS workspace data directory and invalidated when package metadata or git `HEAD` changes. ContextOS does not maintain a skill taxonomy or domain gate list for ranking; if the skill index is cold for a large catalog, prompt hooks fail open instead of falling back to arbitrary keyword matches. Skill catalogs are deduplicated by normalized skill name before indexing and rendering.
 
 After `ctx refresh`, ContextOS invalidates the private hook bridge socket so prompts fall back to direct scoring until Codex restarts the long-running `ctx-mcp` process. Hook clients also discard a same-inode socket if an older bridge revision is detected.
 
@@ -538,6 +544,8 @@ CONTEXTOS_HOOK_DEADLINE_MS=8500 hard fail-open deadline for prompt hooks
 CONTEXTOS_DIRECT_FALLBACK_TIMEOUT_MS=6000 direct scoring timeout when the bridge is unavailable
 CONTEXTOS_HOOK_EMBEDDING_TIMEOUT_MS=500 rule embedding timeout during hook direct fallback
 CONTEXTOS_EMBEDDING_TIMEOUT_MS=800 embedding scoring timeout inside ctx-mcp/debug
+CONTEXTOS_HOOK_SKILL_EMBEDDING_TIMEOUT_MS=2000 skill retrieval timeout during hook direct fallback
+CONTEXTOS_SKILL_EMBEDDING_TIMEOUT_MS=2000 skill retrieval timeout inside ctx-mcp/debug
 CONTEXTOS_FILE_EMBEDDINGS=0       disable file-path embedding retrieval
 CONTEXTOS_HOOK_FILE_EMBEDDING_TIMEOUT_MS=500 file retrieval timeout during hook direct fallback
 CONTEXTOS_FILE_EMBEDDING_TIMEOUT_MS=1000 file-path embedding retrieval timeout

package/bin/ctx.js

@@ -34,7 +34,7 @@ import { scanSkills, warmSkillEmbeddings } from "../plugins/ctx/lib/skill-discov
 import { parsePassthroughArgs, runPassthrough } from "../plugins/ctx/lib/passthrough.js";
 import { parseAgentList, parseSetupArgs, setupSummaryLines } from "../plugins/ctx/lib/setup-wizard.js";
 import { multiSelect } from "../plugins/ctx/lib/multi-select.js";
-import { configureOutputSections, enabledOutputSectionsLabel, loadOutputConfig } from "../plugins/ctx/lib/output-config.js";
+import { configureOutputSections, enabledOutputSectionsLabel, loadOutputConfig, outputConfigLimits, outputConfigLimitsLabel } from "../plugins/ctx/lib/output-config.js";
 import { syncWorkflows, warmWorkflowEmbeddings } from "../plugins/ctx/lib/workflow-discoverer.js";
 import { checkForUpdate } from "../plugins/ctx/lib/update-notifier.js";
 import { fetchSkillsForAgents, printSkillRecommendations, getAllLibraries, getInstallCommands } from "../plugins/ctx/lib/skill-library.js";
@@ -487,6 +487,7 @@ function graphStrategyForInstall() {
 async function warmInstallEmbeddings() {
   const dataDir = contextOSDataDir();
   const modelReady = isModelCacheReady(dataDir);
+  const allowRemote = shouldAllowRemoteWarm(modelReady);
   const result = await warmRuleEmbeddings({
     rules: [
       { content: "Always use project rules that are semantically relevant to the user prompt." },
@@ -496,29 +497,29 @@ async function warmInstallEmbeddings() {
     task: "kiểm duyệt upload moderation semantic code search",
     dataDir,
     sources: [],
-    allowRemote: !modelReady
+    allowRemote
   });
   const fileResult = await warmFileEmbeddings({
     cwd: process.cwd(),
     dataDir,
-    allowRemote: !modelReady
+    allowRemote
+  });
+  const skillResult = await warmSkillEmbeddings({
+    cwd: process.cwd(),
+    dataDir,
+    allowRemote
   });
   const warmDiscovery = process.env.CONTEXTOS_INSTALL_WARM_DISCOVERY === "1";
-  const skillResult = warmDiscovery
-    ? await warmSkillEmbeddings({
-      cwd: process.cwd(),
-      dataDir,
-      allowRemote: !modelReady
-    })
-    : { count: 0 };
   const workflowResult = warmDiscovery
     ? await warmWorkflowEmbeddings({
       cwd: process.cwd(),
       dataDir,
-      allowRemote: !modelReady
+      allowRemote
     })
     : { count: 0 };
-  const graphEmbedding = embedCodeReviewGraph({ cwd: process.cwd() });
+  const graphEmbedding = allowRemote
+    ? embedCodeReviewGraph({ cwd: process.cwd() })
+    : { status: "skipped", reason: "remote-embedding-disabled" };
   return { ...result, modelAlreadyCached: modelReady, fileCount: fileResult.count, skillCount: skillResult.count, workflowCount: workflowResult.count, graphEmbedding };
 }
 
@@ -586,18 +587,20 @@ function contextOSWorkspaceDataDir(cwd = process.cwd()) {
 
 async function debug(task) {
   const cwd = process.cwd();
+  const limits = outputConfigLimits(loadOutputConfig({ dataRoot: contextOSDataDir() }));
   const scored = await scoreContext({
     cwd,
     prompt: task,
     dataDir: contextOSDataDir(),
-    maxFiles: 7,
-    maxSkills: 7,
+    maxFiles: limits.files,
+    maxSkills: limits.skills,
+    maxWorkflows: limits.workflows,
     embeddingTimeoutMs: Number(process.env.CONTEXTOS_EMBEDDING_DEBUG_TIMEOUT_MS || 5000)
   });
   const rules = scored.scoredRules;
-  const relevantFiles = scored.suggestedFiles.slice(0, 7);
-  const suggestedSkills = (scored.suggestedSkills || []).slice(0, 7);
-  const suggestedWorkflows = (scored.suggestedWorkflows || []).slice(0, 2);
+  const relevantFiles = scored.suggestedFiles.slice(0, limits.files);
+  const suggestedSkills = (scored.suggestedSkills || []).slice(0, limits.skills);
+  const suggestedWorkflows = (scored.suggestedWorkflows || []).slice(0, limits.workflows);
   const scheduled = scheduleContext({ rules, relevantFiles, suggestedSkills, suggestedWorkflows });
 
   console.log("ContextOS debug");
@@ -661,31 +664,36 @@ async function warmEmbeddings(task, { syncMarketplace = true, quiet = false } =
 
 async function warmWorkspaceIndexes({ task = "project context" } = {}) {
   const cwd = process.cwd();
+  const dataDir = contextOSDataDir();
+  const modelReady = isModelCacheReady(dataDir);
+  const allowRemote = shouldAllowRemoteWarm(modelReady);
   const merged = readAgentsChain({ cwd });
   const rules = scoreRules(filterActionableRules(parseRules(merged.content)), task, []);
   const result = await warmRuleEmbeddings({
     rules,
     task,
-    dataDir: contextOSDataDir(),
+    dataDir,
     sources: merged.sources,
-    allowRemote: true
+    allowRemote
   });
   const fileResult = await warmFileEmbeddings({
     cwd,
-    dataDir: contextOSDataDir(),
-    allowRemote: true
+    dataDir,
+    allowRemote
   });
   const skillResult = await warmSkillEmbeddings({
     cwd,
-    dataDir: contextOSDataDir(),
-    allowRemote: true
+    dataDir,
+    allowRemote
   });
   const workflowResult = await warmWorkflowEmbeddings({
     cwd,
-    dataDir: contextOSDataDir(),
-    allowRemote: true
+    dataDir,
+    allowRemote
   });
-  const graphEmbedding = embedCodeReviewGraph({ cwd });
+  const graphEmbedding = allowRemote
+    ? embedCodeReviewGraph({ cwd })
+    : { status: "skipped", reason: "remote-embedding-disabled" };
   return {
     ruleCount: result.count,
     fileCount: fileResult.count,
@@ -696,12 +704,28 @@ async function warmWorkspaceIndexes({ task = "project context" } = {}) {
   };
 }
 
+function shouldAllowRemoteWarm(modelReady) {
+  if (modelReady) return false;
+  if (process.env.CONTEXTOS_EMBEDDING_ALLOW_REMOTE !== undefined) {
+    return process.env.CONTEXTOS_EMBEDDING_ALLOW_REMOTE === "1";
+  }
+  return !isCiEnvironment();
+}
+
+function isCiEnvironment() {
+  return process.env.CI === "true"
+    || process.env.GITHUB_ACTIONS === "true"
+    || process.env.CONTINUOUS_INTEGRATION === "true"
+    || process.env.BUILD_ID !== undefined
+    || process.env.RUN_ID !== undefined;
+}
+
 async function refresh() {
   const marketplaceSync = syncActiveCodexMarketplace();
   const invalidatedBridge = invalidateCtxMcpSocket(contextOSDataDir());
   const warmResult = await warmInstallEmbeddings();
   console.log(`Marketplace: ${marketplaceSync.synced ? "synced" : "already active"} (${marketplaceSync.targetRoot})`);
-  console.log(`Indexes: ${warmResult.fileCount || 0} file paths rebuilt`);
+  console.log(`Indexes: ${warmResult.fileCount || 0} file paths rebuilt, ${warmResult.skillCount || 0} skills indexed`);
   console.log(`Graph embeddings: ${formatCodeReviewGraphEmbedding(warmResult.graphEmbedding)}`);
   if (invalidatedBridge) console.log("Bridge: stale private socket invalidated");
   console.log("Restart Codex if ctx-mcp was already running.");
@@ -728,6 +752,21 @@ async function askSetupYesNo(rl, question, defaultValue = true) {
   return !/^n(o)?$/i.test(answer.trim());
 }
 
+async function askOutputLimit({ option, currentValue }) {
+  if (!process.stdin.isTTY) return currentValue;
+  const rl = readline.createInterface({ input, output });
+  try {
+    const answer = await rl.question(`◇ ${option.label} limit (0-${option.max}, current ${currentValue}): `);
+    const trimmed = answer.trim();
+    if (!trimmed) return currentValue;
+    const value = Number(trimmed);
+    if (!Number.isFinite(value)) return currentValue;
+    return Math.max(0, Math.min(option.max, Math.trunc(value)));
+  } finally {
+    rl.close();
+  }
+}
+
 async function setup({ args = [], cwd = process.cwd() } = {}) {
   const options = parseSetupArgs(args);
   const interactive = !options.yes && process.stdin.isTTY;
@@ -777,7 +816,8 @@ async function setup({ args = [], cwd = process.cwd() } = {}) {
     console.log("◇ Configure prompt output:");
     outputConfig = await configureOutputSections({
       dataRoot: contextOSDataDir(),
-      select: multiSelect
+      select: multiSelect,
+      askLimit: askOutputLimit
     });
   }
 
@@ -786,7 +826,8 @@ async function setup({ args = [], cwd = process.cwd() } = {}) {
   for (const line of setupSummaryLines({
     cwd,
     ...options,
-    promptSections: enabledOutputSectionsLabel(outputConfig)
+    promptSections: enabledOutputSectionsLabel(outputConfig),
+    promptLimits: outputConfigLimitsLabel(outputConfig)
   })) console.log(`│  ${line}`);
   console.log("");
 
@@ -871,7 +912,8 @@ try {
   } else if (command === "--config" || command === "config") {
     await configureOutputSections({
       dataRoot: contextOSDataDir(),
-      select: multiSelect
+      select: multiSelect,
+      askLimit: askOutputLimit
     });
   } else if (command === "install") {
     const copy = args.includes("--copy");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@minhpnq1807/contextos",
-  "version": "0.5.45",
+  "version": "0.5.49",
   "description": "Task-aware AGENTS.md context injection and compliance reporting for Codex, Claude Code, and Antigravity.",
   "type": "module",
   "bin": {

package/plugins/ctx/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ctx",
-  "version": "0.5.45",
+  "version": "0.5.49",
   "description": "Inject task-relevant AGENTS.md rules into Codex through plugin hooks.",
   "author": {
     "name": "ContextOS"

package/plugins/ctx/lib/embedding-scorer.js

@@ -65,13 +65,28 @@ export async function warmRuleEmbeddings({
     ...rules.map((rule) => rule.content || "")
   ].filter((text) => String(text).trim()))];
 
-  const cache = await openEmbeddingCache(dataDir);
-  const embedder = await getExtractor({ allowRemote, dataDir });
-  for (const text of texts) {
-    await getCachedEmbedding({ cache, embedder, text, sources });
+  let cache;
+  try {
+    cache = await openEmbeddingCache(dataDir);
+    const embedder = await getExtractor({ allowRemote, dataDir });
+    for (const text of texts) {
+      await getCachedEmbedding({ cache, embedder, text, sources, flush: false });
+    }
+    cache.close();
+    return { count: texts.length, cachePath: cache.path, status: "enabled" };
+  } catch (error) {
+    try {
+      cache?.close();
+    } catch {
+      // Ignore close failures while reporting a best-effort warm result.
+    }
+    return {
+      count: 0,
+      cachePath: path.join(dataDir, "embeddings.db"),
+      status: "warm-failed",
+      error: error?.message || String(error)
+    };
   }
-  cache.close();
-  return { count: texts.length, cachePath: cache.path };
 }
 
 export async function searchIndexedEmbeddings({
@@ -106,20 +121,35 @@ export async function warmIndexedEmbeddings({
     return { count: 0, cachePath: path.join(dataDir, "embeddings.db"), status: "missing-model" };
   }
 
-  const cache = await openEmbeddingCache(dataDir);
-  const embedder = await getExtractor({ allowRemote, dataDir });
-  if (String(task || "").trim()) await getCachedEmbedding({ cache, embedder, text: task, sources });
-
-  const indexed = [];
-  for (const item of items) {
-    const text = String(item.text || "");
-    if (!item.id || !text.trim()) continue;
-    const vector = await getCachedEmbedding({ cache, embedder, text, sources });
-    indexed.push({ id: item.id, text, vector });
+  let cache;
+  try {
+    cache = await openEmbeddingCache(dataDir);
+    const embedder = await getExtractor({ allowRemote, dataDir });
+    if (String(task || "").trim()) await getCachedEmbedding({ cache, embedder, text: task, sources });
+
+    const indexed = [];
+    for (const item of items) {
+      const text = String(item.text || "");
+      if (!item.id || !text.trim()) continue;
+      const vector = await getCachedEmbedding({ cache, embedder, text, sources, flush: false });
+      indexed.push({ id: item.id, text, vector });
+    }
+    cache.replaceIndex(kind, indexed);
+    cache.close();
+    return { count: indexed.length, cachePath: cache.path, status: "enabled" };
+  } catch (error) {
+    try {
+      cache?.close();
+    } catch {
+      // Ignore close failures while reporting a best-effort warm result.
+    }
+    return {
+      count: 0,
+      cachePath: path.join(dataDir, "embeddings.db"),
+      status: "warm-failed",
+      error: error?.message || String(error)
+    };
   }
-  cache.replaceIndex(kind, indexed);
-  cache.close();
-  return { count: indexed.length, cachePath: cache.path };
 }
 
 async function enhanceRuleScores(rules, task, { dataDir, sources, allowRemote }) {
@@ -187,7 +217,10 @@ async function getExtractor({ allowRemote, dataDir }) {
       return transformers.pipeline("feature-extraction", DEFAULT_MODEL, {
         quantized: true
       });
-    })());
+    })().catch((error) => {
+      extractorPromises.delete(key);
+      throw error;
+    }));
   }
   return extractorPromises.get(key);
 }
@@ -206,7 +239,7 @@ export function isModelCacheReady(dataDir = defaultDataRoot()) {
   ].every((relativePath) => fs.existsSync(path.join(modelDir, relativePath)));
 }
 
-async function getCachedEmbedding({ cache, embedder, text, sources }) {
+async function getCachedEmbedding({ cache, embedder, text, sources, flush = true }) {
   const key = cacheKey(text, sources);
   const existing = cache.get(key);
   if (existing) return existing;
@@ -216,7 +249,7 @@ async function getCachedEmbedding({ cache, embedder, text, sources }) {
     normalize: true
   });
   const embedding = Array.from(output.data || []);
-  cache.set(key, embedding);
+  cache.set(key, embedding, { flush });
   return embedding;
 }
 
@@ -225,6 +258,7 @@ export async function openEmbeddingCache(dataDir) {
   const cachePath = path.join(dataDir, "embeddings.db");
   const SQL = await getSql();
   const db = initializeEmbeddingDatabase(SQL, cachePath);
+  let dirty = false;
 
   return {
     path: cachePath,
@@ -238,12 +272,16 @@ export async function openEmbeddingCache(dataDir) {
         stmt.free();
       }
     },
-    set(key, vector) {
+    set(key, vector, { flush = true } = {}) {
       db.run(
         "INSERT OR REPLACE INTO embeddings (key, model, vector, updated_at) VALUES (?, ?, ?, ?)",
         [key, DEFAULT_MODEL, JSON.stringify(vector), new Date().toISOString()]
       );
-      writeDatabaseAtomically(cachePath, db);
+      dirty = true;
+      if (flush) {
+        writeDatabaseAtomically(cachePath, db);
+        dirty = false;
+      }
     },
     listIndexed(kind) {
       const stmt = db.prepare("SELECT id, text, vector FROM embedding_index WHERE kind = ? AND model = ?");
@@ -268,9 +306,10 @@ export async function openEmbeddingCache(dataDir) {
         );
       }
       writeDatabaseAtomically(cachePath, db);
+      dirty = false;
     },
     close() {
-      writeDatabaseAtomically(cachePath, db);
+      if (dirty) writeDatabaseAtomically(cachePath, db);
       db.close();
     }
   };

package/plugins/ctx/lib/graph-strategy.js

@@ -91,6 +91,7 @@ export function formatCodeReviewGraphEmbedding(result) {
   }
   if (result.reason === "missing-graph-index") return "skipped (no .code-review-graph/graph.db)";
   if (result.reason === "missing-code-review-graph-python") return "skipped (code-review-graph Python unavailable)";
+  if (result.reason === "remote-embedding-disabled") return "skipped (remote embedding disabled)";
   return `skipped (${result.error || result.reason || "unavailable"})`;
 }
 

package/plugins/ctx/lib/output-config.js

@@ -13,9 +13,16 @@ export const OUTPUT_SECTION_OPTIONS = [
   { value: "workflows", label: "Suggested workflow for this task", hint: "Include matching workflow recommendations." }
 ];
 
+export const OUTPUT_LIMIT_OPTIONS = [
+  { value: "files", label: "Suggested files", defaultValue: 5, max: 20 },
+  { value: "skills", label: "Suggested skills", defaultValue: 5, max: 10 },
+  { value: "workflows", label: "Suggested workflows", defaultValue: 5, max: 5 }
+];
+
 export function defaultOutputConfig() {
   return {
-    sections: Object.fromEntries(OUTPUT_SECTION_OPTIONS.map((option) => [option.value, true]))
+    sections: Object.fromEntries(OUTPUT_SECTION_OPTIONS.map((option) => [option.value, true])),
+    limits: Object.fromEntries(OUTPUT_LIMIT_OPTIONS.map((option) => [option.value, option.defaultValue]))
   };
 }
 
@@ -49,9 +56,19 @@ export function enabledOutputSectionsLabel(config = loadOutputConfig()) {
   return enabled.length ? enabled.join(", ") : "(none)";
 }
 
+export function outputConfigLimits(config = loadOutputConfig()) {
+  return normalizeOutputConfig(config).limits;
+}
+
+export function outputConfigLimitsLabel(config = loadOutputConfig()) {
+  const limits = outputConfigLimits(config);
+  return OUTPUT_LIMIT_OPTIONS.map((option) => `${option.value}: ${limits[option.value]}`).join(", ");
+}
+
 export async function configureOutputSections({
   dataRoot = defaultDataRoot(),
   select,
+  askLimit,
   logger = console.log
 } = {}) {
   if (typeof select !== "function") throw new Error("configureOutputSections requires a multi-select function");
@@ -64,11 +81,19 @@ export async function configureOutputSections({
     }))
   });
   const selectedSet = new Set(selected);
+  const limits = {};
+  for (const option of OUTPUT_LIMIT_OPTIONS) {
+    limits[option.value] = typeof askLimit === "function"
+      ? await askLimit({ option, currentValue: current.limits[option.value] })
+      : current.limits[option.value];
+  }
   const saved = saveOutputConfig({
-    sections: Object.fromEntries(OUTPUT_SECTION_OPTIONS.map((option) => [option.value, selectedSet.has(option.value)]))
+    sections: Object.fromEntries(OUTPUT_SECTION_OPTIONS.map((option) => [option.value, selectedSet.has(option.value)])),
+    limits
   }, { dataRoot });
   logger(`│  Saved ContextOS prompt section config: ${outputConfigPath(dataRoot)}`);
   logger(`│  Enabled sections: ${enabledOutputSectionsLabel(saved)}`);
+  logger(`│  Suggest limits: ${outputConfigLimitsLabel(saved)}`);
   return saved;
 }
 
@@ -80,6 +105,16 @@ function normalizeOutputConfig(config = {}) {
       typeof config.sections?.[option.value] === "boolean"
         ? config.sections[option.value]
         : defaults.sections[option.value]
+    ])),
+    limits: Object.fromEntries(OUTPUT_LIMIT_OPTIONS.map((option) => [
+      option.value,
+      normalizeLimit(config.limits?.[option.value], option)
     ]))
   };
 }
+
+function normalizeLimit(value, option) {
+  const number = Number(value);
+  if (!Number.isFinite(number)) return option.defaultValue;
+  return Math.max(0, Math.min(option.max, Math.trunc(number)));
+}