@wrongstack/cli 0.9.20 → 0.10.0

package/data/README.md

@@ -0,0 +1,77 @@
+# Curated model-catalog overlay (`providers.json`)
+
+`providers.json` is a **curated override layer** that WrongStack deep-merges **on top of**
+the live `https://models.dev/api.json` catalog. models.dev stays the base/primary source; this
+file lets us **add** providers/models it doesn't carry and **fix** fields it gets wrong (a missing
+model, a stale context limit, etc.) without waiting for an upstream fix or a release.
+
+At runtime the registry resolves:
+
+```
+merged = mergeModelsPayload(modelsDev, providers.json)   // overlay wins
+```
+
+It is loaded from (first non-empty wins):
+1. this file fetched from our GitHub raw URL (so it can refresh between releases), then
+2. this file bundled in the installed package (offline floor).
+
+If models.dev is completely unreachable and there's no cache, a **non-empty** overlay still drives
+the catalog on its own. An empty `{}` overlay is a safe no-op.
+
+## Shape
+
+Same schema as `models.dev/api.json` — a map keyed by provider id. You only include the fields you
+want to add or override; everything else falls through to the base. (JSON has no comments, hence
+this README.)
+
+```jsonc
+{
+  // Override just one field on an existing model — here, fix a context window.
+  "deepseek": {
+    "models": {
+      "deepseek-v4-pro": { "limit": { "context": 128000 } }
+    }
+  },
+
+  // Add a provider models.dev doesn't list at all.
+  "myco": {
+    "id": "myco",
+    "name": "My Co",
+    "npm": "@ai-sdk/openai-compatible",   // determines the wire family
+    "api": "https://api.myco.example/v1",
+    "env": ["MYCO_API_KEY"],
+    "models": {
+      "myco-large": {
+        "id": "myco-large",
+        "name": "MyCo Large",
+        "tool_call": true,
+        "modalities": { "input": ["text"], "output": ["text"] },
+        "limit": { "context": 200000, "output": 16000 },
+        "cost": { "input": 0.5, "output": 1.5 }
+      }
+    }
+  }
+}
+```
+
+## Merge rules
+
+- Provider in both → overlay scalar fields (`name`, `npm`, `api`, `env`, `doc`) override the base;
+  `models` are merged by id.
+- Model in both → `{ ...baseModel, ...overlayModel }`, with the nested `limit` / `cost` /
+  `modalities` objects merged one level deeper — so `{"limit":{"context":…}}` overrides only the
+  context and keeps the base's `limit.output`.
+- Anything only in the overlay is added.
+
+## Editing / refreshing
+
+Use the helper to seed and sanity-check entries against upstream:
+
+```bash
+pnpm run sync:models -- --extract deepseek:deepseek-v4-pro   # print a paste-ready overlay snippet
+pnpm run sync:models -- --diff                               # what we override vs upstream + drift
+```
+
+Then edit `providers.json` and commit. Keep it **small and curated** — it is an override layer,
+not a mirror of models.dev. Once models.dev catches up, drop the now-redundant override (`--diff`
+flags those).

package/data/providers.json

@@ -0,0 +1 @@
+{}

package/dist/index.js

@@ -13,6 +13,7 @@ import { MCPRegistry } from '@wrongstack/mcp';
 import { buildProviderFactoriesFromRegistry, makeProviderFromConfig, capabilitiesFor } from '@wrongstack/providers';
 import { createDefaultContainer, routeImagesForModel, readClipboardImage } from '@wrongstack/runtime';
 import { builtinToolsPack, rememberTool, forgetTool } from '@wrongstack/tools';
+import { fileURLToPath } from 'url';
 import * as readline from 'readline';
 import * as fs10 from 'fs';
 import { writeFileSync } from 'fs';
@@ -2353,6 +2354,7 @@ var BOOLEAN_FLAGS = /* @__PURE__ */ new Set([
   "recover",
   "no-alt-screen",
   "alt-screen",
+  "mouse",
   "output-json",
   "prompt",
   "metrics",
@@ -9727,6 +9729,14 @@ function fmtTaskResultLine(r, color41) {
 
 // src/boot.ts
 init_update_check();
+var GITHUB_PROVIDERS_OVERLAY_URL = "https://raw.githubusercontent.com/WrongStack/WrongStack/main/packages/cli/data/providers.json";
+function resolveBundledOverlayFile() {
+  try {
+    return fileURLToPath(new URL("../data/providers.json", import.meta.url));
+  } catch {
+    return void 0;
+  }
+}
 function resolveBundledSkillsDir() {
   try {
     const req2 = createRequire(import.meta.url);
@@ -9742,6 +9752,10 @@ async function boot(argv) {
     flags["resume"] = positional[1];
     positional.splice(0, 2);
   }
+  if (positional.length === 0) {
+    if (flags["help"] === true) positional.push("help");
+    else if (flags["version"] === true) positional.push("version");
+  }
   let bootResult;
   try {
     bootResult = await bootConfig(flags);
@@ -9758,7 +9772,12 @@ async function boot(argv) {
   const reader = new ReadlineInputReader({ historyFile: wpaths.historyFile });
   const modelsRegistry = new DefaultModelsRegistry({
     cacheFile: wpaths.modelsCache,
-    ttlSeconds: 24 * 3600
+    ttlSeconds: 24 * 3600,
+    // Curated overlay merged on top of models.dev: fetched from GitHub raw for
+    // freshness, with the bundled file as the offline floor.
+    overlayUrl: GITHUB_PROVIDERS_OVERLAY_URL,
+    overlayFile: resolveBundledOverlayFile(),
+    overlayCacheFile: wpaths.modelsOverlayCache
   });
   let updateInfo;
   if (!flags["no-check"] && !process.env["WRONGSTACK_NO_CHECK"]) {
@@ -10911,9 +10930,15 @@ async function execute(deps) {
           // Default OFF so the terminal's native scrollback works for chat
           // history out of the box (mouse wheel / Shift+PgUp). Users who hit
           // resize/overlay-leak artifacts can opt back into alt-screen with
-          // `--alt-screen` or `/altscreen on`. `--no-alt-screen` still wins
-          // when both are passed.
+          // `--alt-screen`. `--no-alt-screen` still wins when both are passed.
           altScreen: flags["alt-screen"] === true && flags["no-alt-screen"] !== true,
+          // Full mouse mode: clickable pickers + in-app wheel scroll. Opt-in
+          // via --mouse. The TUI combines altScreen || mouse to determine
+          // whether to use the managed viewport (ScrollableHistory + in-app
+          // scroll); --mouse alone enables that without requiring --alt-screen.
+          // Mouse mode takes ownership of the terminal's mouse, suspending
+          // native scroll/copy until `/mouse off`.
+          mouse: flags.mouse === true,
           director,
           fleetRoster,
           onAfterExit: () => {
@@ -11208,7 +11233,11 @@ var MultiAgentHost = class {
   makeSubagentFactory(config) {
     return async (subCfg) => {
       const events = new EventBus();
-      const provider = await this.buildSubagentProvider(config, subCfg.provider);
+      const provider = await this.buildSubagentProvider(
+        config,
+        subCfg.provider,
+        subCfg.model ?? config.model
+      );
       const subCwd = subCfg.cwd ?? this.deps.cwd;
       const baseSystem = await this.deps.systemPromptBuilder.build({
         cwd: subCwd,
@@ -11346,17 +11375,28 @@ var MultiAgentHost = class {
    * not configured (so a typo doesn't crash the whole run — we just
    * use the leader and the calling code can decide to error later).
    */
-  async buildSubagentProvider(config, overrideId) {
+  async buildSubagentProvider(config, overrideId, model) {
     const providerId = overrideId && config.providers?.[overrideId] ? overrideId : config.provider;
     const newCfg = config.providers?.[providerId] ?? {
       type: providerId,
       apiKey: config.apiKey,
       baseUrl: config.baseUrl
     };
-    return makeProviderFromConfig(providerId, {
+    const provider = makeProviderFromConfig(providerId, {
       ...newCfg,
       type: providerId
     });
+    if (this.deps.modelsRegistry) {
+      const resolvedModel = model ?? config.model;
+      const caps = await capabilitiesFor(
+        this.deps.modelsRegistry,
+        providerId,
+        resolvedModel
+      ).catch(() => void 0);
+      const mc = caps?.maxContext ?? config.context?.effectiveMaxContext ?? provider.capabilities.maxContext;
+      if (mc && mc > 0) provider.capabilities.maxContext = mc;
+    }
+    return provider;
   }
   async spawnACP(subagentId, task, config) {
     const taskId = randomUUID();
@@ -12138,6 +12178,13 @@ function renderProgress3(ratio, width) {
   const capped = Math.min(width, filled);
   return FILLED2.repeat(capped) + EMPTY2.repeat(width - capped);
 }
+var createSessionEventBridge = (_writer, level) => ({
+  append: async (_e) => {
+  },
+  level: level ?? "standard",
+  allows: () => true
+});
+var resolveAuditLevel = (cfg) => cfg?.session?.auditLevel ?? "standard";
 function setupPipelines(params) {
   const { events, logger } = params;
   const pipelines = createDefaultPipelines();
@@ -12164,11 +12211,13 @@ function setupPipelines(params) {
   return pipelines;
 }
 async function setupCompaction(params) {
-  const { compactor, events, modelsRegistry, context, config, provider, pipelines } = params;
+  const { compactor, events, modelsRegistry, context, config, provider, pipelines, fullConfig, sessionWriter, sessionBridge: providedBridge } = params;
   const resolvedCaps = await capabilitiesFor(modelsRegistry, provider.id, context.model).catch(() => void 0);
   const effectiveMaxContext = config.context.effectiveMaxContext ?? resolvedCaps?.maxContext ?? provider.capabilities.maxContext;
   let autoCompactor;
   if (config.context.autoCompact !== false) {
+    const auditLevel = resolveAuditLevel(fullConfig ?? config);
+    const sessionBridge = providedBridge ?? createSessionEventBridge(sessionWriter, auditLevel);
     autoCompactor = new AutoCompactionMiddleware(
       compactor,
       effectiveMaxContext,
@@ -12180,7 +12229,12 @@ async function setupCompaction(params) {
         soft: config.context.softThreshold,
         hard: config.context.hardThreshold
       },
-      { aggressiveOn: "soft", failureMode: "throw_on_hard", events }
+      {
+        aggressiveOn: "soft",
+        failureMode: "throw_on_hard",
+        events,
+        sessionBridge
+      }
     );
     pipelines.contextWindow.use({ name: "AutoCompaction", handler: autoCompactor.handler() });
   }
@@ -12553,6 +12607,21 @@ async function setupSession(params) {
 }
 
 // src/index.ts
+var createSessionEventBridge2 = (_writer, level, _opts) => ({
+  append: async (_e) => {
+  },
+  level: level ?? "standard",
+  allows: () => true
+});
+var resolveAuditLevel2 = (cfg) => cfg?.session?.auditLevel ?? "standard";
+var resolveSessionLoggingConfig = (cfg) => ({
+  auditLevel: resolveAuditLevel2(cfg),
+  sampling: {
+    toolProgress: {
+      sampleRate: cfg?.session?.sampling?.toolProgress?.sampleRate ?? 8
+    }
+  }
+});
 function resolveBundledSkillsDir2() {
   try {
     const req2 = createRequire(import.meta.url);
@@ -12782,18 +12851,99 @@ async function main(argv) {
   const planPath = sessResult.planPath;
   const detachTodosCheckpoint = sessResult.detachTodosCheckpoint;
   const priorFleetState = sessResult.priorFleetState;
+  const sessionConfig = resolveSessionLoggingConfig(config);
+  const sessionBridge = createSessionEventBridge2(
+    session,
+    sessionConfig.auditLevel);
   const stats = new SessionStats(events, tokenCounter);
   const errorRing = [];
   events.on("error", (e) => {
     const err = e.err;
     const code = err && typeof err === "object" && "code" in err && typeof err.code === "string" ? err.code : "UNKNOWN";
     const message = e.err instanceof Error ? e.err.message : String(e.err);
-    errorRing.push({ ts: (/* @__PURE__ */ new Date()).toISOString(), phase: e.phase, code, message });
+    const ts = (/* @__PURE__ */ new Date()).toISOString();
+    errorRing.push({ ts, phase: e.phase, code, message });
     if (errorRing.length > 5) errorRing.shift();
+    sessionBridge.append({
+      type: "error",
+      ts,
+      message,
+      phase: e.phase
+    }).catch(() => {
+    });
+  });
+  events.on("tool.started", (e) => {
+    sessionBridge.append({
+      type: "tool_call_start",
+      ts: (/* @__PURE__ */ new Date()).toISOString(),
+      name: e.name,
+      id: e.id,
+      input: e.input
+    }).catch(() => {
+    });
+  });
+  events.on("tool.executed", (e) => {
+    sessionBridge.append({
+      type: "tool_call_end",
+      ts: (/* @__PURE__ */ new Date()).toISOString(),
+      name: e.name,
+      id: e.id ?? "",
+      durationMs: e.durationMs,
+      outputSize: e.outputBytes ?? 0,
+      ok: e.ok,
+      outputBytes: e.outputBytes,
+      outputTokens: e.outputTokens,
+      outputLines: e.outputLines
+    }).catch(() => {
+    });
+  });
+  events.on("tool.progress", (e) => {
+    sessionBridge.append({
+      type: "tool_progress",
+      ts: (/* @__PURE__ */ new Date()).toISOString(),
+      name: e.name,
+      id: e.id,
+      event: { type: e.event.type, text: e.event.text, data: e.event.data }
+    }).catch(() => {
+    });
+  });
+  events.on("provider.retry", (e) => {
+    sessionBridge.append({
+      type: "provider_retry",
+      ts: (/* @__PURE__ */ new Date()).toISOString(),
+      providerId: e.providerId,
+      attempt: e.attempt,
+      delayMs: e.delayMs,
+      status: e.status,
+      description: e.description
+    }).catch(() => {
+    });
+  });
+  events.on("provider.error", (e) => {
+    sessionBridge.append({
+      type: "provider_error",
+      ts: (/* @__PURE__ */ new Date()).toISOString(),
+      providerId: e.providerId,
+      status: e.status,
+      description: e.description,
+      retryable: e.retryable
+    }).catch(() => {
+    });
   });
   const pipelines = setupPipelines({ events, logger });
   const compactor = container.resolve(TOKENS.Compactor);
-  const { effectiveMaxContext, autoCompactor } = await setupCompaction({ compactor, events, modelsRegistry, context, config, provider, pipelines });
+  const { effectiveMaxContext, autoCompactor } = await setupCompaction({
+    compactor,
+    events,
+    modelsRegistry,
+    context,
+    config,
+    provider,
+    pipelines,
+    fullConfig: config,
+    sessionBridge
+    // share the same bridge for consistent audit logging (compaction + errors + future)
+  });
   const refreshMaxContext = async (providerId, modelId) => {
     if (!autoCompactor) return;
     const cap = await capabilitiesFor(modelsRegistry, providerId, modelId).catch(() => void 0);
@@ -12890,6 +13040,7 @@ async function main(argv) {
       toolRegistry,
       providerRegistry,
       configStore,
+      modelsRegistry,
       events,
       systemPromptBuilder: promptBuilder,
       session,