@compilr-dev/sdk 0.34.0 → 0.36.0

package/dist/errors/classify.js

@@ -33,13 +33,50 @@ function statusFromText(text) {
     const m = text.match(/\b(400|401|403|404|408|409|413|422|429|500|502|503|504|529)\b/);
     return m ? Number(m[1]) : undefined;
 }
+/**
+ * Textual markers for "the key is wrong", measured against the real wire text on 2026-10-03.
+ *
+ * ⚠️ EVERY MARKER BUT ONE USED TO BE AN ANTHROPIC STRING, so a bad key classified as `auth` for
+ * exactly one of three first-party providers. What the others actually send:
+ *
+ *   Anthropic  401  {"error":{"type":"authentication_error","message":"API key is invalid."}}
+ *   OpenAI     401  {"error":{"message":"Incorrect API key provided: sk-…","code":"invalid_api_key"}}
+ *   Gemini     400  {"error":{"message":"API key not valid. Please pass a valid API key.",
+ *                             "details":[{"reason":"API_KEY_INVALID"}]}}
+ *
+ * Two independent reasons the old set missed them:
+ *
+ *  1. **Gemini answers 400, not 401**, so no status test can catch it — the text is the only
+ *     signal, and "API key not valid" is not the substring "invalid api key".
+ *  2. **Our own providers rewrite the message and interpolate their name into the middle of the
+ *     phrase**: `gemini-native` produces "Invalid Gemini API key. …" and `openai.ts` produces
+ *     "Invalid OpenAI API key. …". Neither contains "invalid api key". In-process the thrown
+ *     `ProviderError.statusCode` saved OpenAI; across an IPC boundary — Desktop, the path this
+ *     module exists for — only the string crosses, and both classified as `unknown`.
+ *
+ * Hence `invalid(?: \w+)? api key`: it matches the bare phrase and the branded one, so a
+ * provider adding its name no longer silently breaks the classification.
+ */
+const AUTH_MARKERS = [
+    /authentication_error/,
+    /permission_error/,
+    /unauthorized/,
+    /x-api-key/,
+    // Gemini: the message, and the machine-readable reason in `details`.
+    /api key not valid/,
+    /api[_ -]?key[_ -]?invalid/,
+    // OpenAI: the message, and its `error.code`.
+    /incorrect api key/,
+    /invalid_api_key/,
+    // Ours and anyone else's, with or without a provider name in the middle.
+    /invalid(?: \w+)? api key/,
+    /\b(?:no|missing) api key/,
+];
 function categorize(statusCode, text) {
     const lower = text.toLowerCase();
-    // Auth: key issues. Match on status OR the provider's textual markers, since a
-    // stringified error may not preserve the numeric code.
-    if (statusCode === 401 ||
-        statusCode === 403 ||
-        /authentication_error|invalid x-api-key|invalid api key|no api key|x-api-key|permission_error|unauthorized/.test(lower)) {
+    // Auth: key issues. Match on status OR a textual marker, since a stringified error may not
+    // preserve the numeric code — and Gemini's bad-key status (400) is not an auth code at all.
+    if (statusCode === 401 || statusCode === 403 || AUTH_MARKERS.some((re) => re.test(lower))) {
         return 'auth';
     }
     if (statusCode === 429 || /rate.?limit|too many requests/.test(lower))

package/dist/index.d.ts

@@ -64,7 +64,7 @@ export { detectProject, suggestProjectType, detectCommon } from './detection/ind
 export type { DetectProjectOptions, DetectionResult, ContentSummary } from './detection/index.js';
 export { createGuideTool, SHARED_GUIDE_ENTRIES, searchGuideEntries, topicToGuideEntry, } from './guide/index.js';
 export type { GuideEntry, ContentTopic, ContentSection, GuideToolConfig } from './guide/index.js';
-export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, createSceneWriter, SCENE_JOURNAL_SIZE, createImageTools, ProjectAnchorStore, FileArtifactService, } from './platform/index.js';
+export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, resolveSelfAlias, createSceneWriter, SCENE_JOURNAL_SIZE, createImageTools, ProjectAnchorStore, FileArtifactService, } from './platform/index.js';
 export type { ProjectAnchorStoreConfig, FileArtifactServiceConfig, ImageToolsConfig, ImageResizer, } from './platform/index.js';
 export { STEP_ORDER, GUIDED_STEP_CRITERIA, getNextStep, isValidTransition, getStepCriteria, formatStepDisplay, getStepNumber, } from './platform/index.js';
 export type { CustomSkill, CompilrSkillExtension, ForkedFromMarker, InstalledFromMarker, SkillEligibilityContext, SkillCollision, SkillDiffLine, SkillValidationIssue, ScopeConfig, SkillResolution, } from './skills/index.js';

package/dist/index.js

@@ -147,7 +147,7 @@ export { createGuideTool, SHARED_GUIDE_ENTRIES, searchGuideEntries, topicToGuide
 // =============================================================================
 // Platform Tools (runtime — createPlatformTools factory + individual factories)
 // =============================================================================
-export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, createSceneWriter, SCENE_JOURNAL_SIZE, createImageTools, ProjectAnchorStore, FileArtifactService, } from './platform/index.js';
+export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, resolveSelfAlias, createSceneWriter, SCENE_JOURNAL_SIZE, createImageTools, ProjectAnchorStore, FileArtifactService, } from './platform/index.js';
 // =============================================================================
 // Platform Workflow (pure step-criteria logic)
 // =============================================================================

package/dist/models/model-registry.js

@@ -739,8 +739,24 @@ export function shouldClearHistoryOnModelChange(_fromModelId, _toModelId) {
 // Context Window Resolution
 // =============================================================================
 /**
- * Provider fallback context limits.
- * Used when a model is not in the registry or has no contextWindow set.
+ * What to assume when the endpoint is unknown.
+ *
+ * ⚠️ UNDER-REPORTING IS THE SAFE DIRECTION. Too small wastes context and compacts early; too
+ * large means compaction fires after the real limit, so the request 400s and reads as a provider
+ * fault. `custom` points at whatever the user is running — LM Studio, vLLM, a proxy — so it gets
+ * the same local-model assumption as Ollama rather than a number borrowed from a frontier model.
+ */
+const UNKNOWN_ENDPOINT_CONTEXT = 32000;
+/**
+ * Provider fallback context limits, used when a model is not in the registry or declares no
+ * `contextWindow` (every Groq / Together / Fireworks / Perplexity / OpenRouter / Ollama entry —
+ * those hosts define the window by deployment, so the registry does not guess per model).
+ *
+ * ⚠️ EXHAUSTIVE ON PURPOSE. This was a `default: return 200000`, which handed out ANTHROPIC's
+ * window to any provider not listed — so `getModelContextWindow('some-local-model', 'custom')`
+ * answered 200000 and a local 8k model would compact long after it had already overflowed. The
+ * `never` assignment below makes a new `ProviderType` a compile error here instead of a silent
+ * inheritance of someone else's limit.
  */
 function getProviderContextLimitFallback(provider) {
     switch (provider) {
@@ -763,8 +779,16 @@ function getProviderContextLimitFallback(provider) {
             return 131072;
         case 'openrouter':
             return 128000;
-        default:
-            return 200000;
+        case 'custom':
+        case undefined:
+            return UNKNOWN_ENDPOINT_CONTEXT;
+        default: {
+            // Unreachable while every ProviderType is handled above — and a COMPILE ERROR the moment
+            // one is added without a limit, which is the whole point of writing it this way.
+            const unhandled = provider;
+            void unhandled;
+            return UNKNOWN_ENDPOINT_CONTEXT;
+        }
     }
 }
 /**

package/dist/models.d.ts

@@ -9,6 +9,20 @@ import type { ProviderType } from './config.js';
  * Default model for each provider.
  * Derived from MODEL_REGISTRY — uses the 'balanced' tier default for each provider.
  */
+/**
+ * Default model for each provider — the 'balanced' tier entry from MODEL_REGISTRY.
+ *
+ * ⚠️ THE KEYS COME FROM `PROVIDER_METADATA`, NOT A HAND-WRITTEN LIST. The list used to be typed
+ * out here and was missing `ollama-anthropic`, so `DEFAULT_MODELS['ollama-anthropic']` was
+ * `undefined` behind a `Record<ProviderType, string>` — the `{} as Record<…>` cast is what let
+ * that compile. `PROVIDER_METADATA` is a true complete Record, so adding a provider to the union
+ * now forces an entry here too.
+ *
+ * ⚠️ `custom` IS DELIBERATELY EMPTY. It used to fall through to `'gpt-6-sol'`, so pointing
+ * `custom` at LM Studio or vLLM without naming a model requested one that endpoint has never
+ * heard of. A custom endpoint's model is the user's to name; there is no default to invent, and
+ * an empty string fails immediately and obviously instead of looking like an intent.
+ */
 export declare const DEFAULT_MODELS: Record<ProviderType, string>;
 /**
  * Default context window when model is unknown

package/dist/models.js

@@ -4,36 +4,39 @@
  * This file maintains backward compatibility with the original SDK API.
  * Internally, it delegates to the full model registry in ./models/.
  */
-import { getDefaultModelForTier, getModelContextWindow, MODEL_REGISTRY } from './models/index.js';
+import { getDefaultModelForTier, getModelContextWindow, MODEL_REGISTRY, PROVIDER_METADATA, } from './models/index.js';
 /**
  * Default model for each provider.
  * Derived from MODEL_REGISTRY — uses the 'balanced' tier default for each provider.
  */
+/**
+ * Default model for each provider — the 'balanced' tier entry from MODEL_REGISTRY.
+ *
+ * ⚠️ THE KEYS COME FROM `PROVIDER_METADATA`, NOT A HAND-WRITTEN LIST. The list used to be typed
+ * out here and was missing `ollama-anthropic`, so `DEFAULT_MODELS['ollama-anthropic']` was
+ * `undefined` behind a `Record<ProviderType, string>` — the `{} as Record<…>` cast is what let
+ * that compile. `PROVIDER_METADATA` is a true complete Record, so adding a provider to the union
+ * now forces an entry here too.
+ *
+ * ⚠️ `custom` IS DELIBERATELY EMPTY. It used to fall through to `'gpt-6-sol'`, so pointing
+ * `custom` at LM Studio or vLLM without naming a model requested one that endpoint has never
+ * heard of. A custom endpoint's model is the user's to name; there is no default to invent, and
+ * an empty string fails immediately and obviously instead of looking like an intent.
+ */
 export const DEFAULT_MODELS = (() => {
-    const providers = [
-        'claude',
-        'openai',
-        'gemini',
-        'ollama',
-        'together',
-        'groq',
-        'fireworks',
-        'perplexity',
-        'openrouter',
-        'custom',
-    ];
     const result = {};
-    for (const provider of providers) {
-        const balanced = getDefaultModelForTier(provider, 'balanced');
-        if (balanced) {
-            result[provider] = balanced.id;
-        }
-        else {
-            // Fallback: first supported model for this provider, or a sensible default
-            const firstModel = MODEL_REGISTRY.find((m) => m.provider === provider && m.status === 'supported');
-            // Fallback only when the registry has no model at all for the provider; gpt-4o is gone.
-            result[provider] = firstModel?.id ?? 'gpt-6-sol';
+    for (const provider of Object.keys(PROVIDER_METADATA)) {
+        if (provider === 'custom') {
+            result[provider] = '';
+            continue;
         }
+        // `ollama-anthropic` is Ollama behind an Anthropic-shaped shim: same models.
+        const lookup = provider === 'ollama-anthropic' ? 'ollama' : provider;
+        const balanced = getDefaultModelForTier(lookup, 'balanced');
+        result[provider] =
+            balanced?.id ??
+                MODEL_REGISTRY.find((m) => m.provider === lookup && m.status === 'supported')?.id ??
+                '';
     }
     return result;
 })();

package/dist/platform/context.d.ts

@@ -47,9 +47,29 @@ export interface PlatformHooks {
         id: number;
         name: string;
     }) => void;
-    /** Resolve owner aliases (e.g., 'self' → active agent ID) */
+    /**
+     * Resolve owner aliases (e.g., 'self' → active agent ID).
+     *
+     * ⚠️ MUST read the agent executing RIGHT NOW, not the one that was active when the tools were
+     * built — platform tools are built once and memoised by every host, so a captured value freezes
+     * for the session. Same trap as `cwd`, below. Implement it with {@link resolveSelfAlias}.
+     */
     resolveOwner?: (alias: 'self') => string | undefined;
 }
+/**
+ * Turn a host's notion of "the agent running now" into an owner alias.
+ *
+ * `'default'` is the coordinator slot every single-agent session uses, not an identity, so it
+ * resolves to `undefined` — which callers then render as their own fallback (`'agent'` for a work
+ * item's author) rather than writing the literal string `default` into stored data.
+ *
+ * ⚠️ This lives here because BOTH hosts need the same answer and only one of them had it. The CLI
+ * implemented the rule inline; Desktop passed no `resolveOwner` at all, so every scene edit
+ * recorded no agent and `consumeUserEdits` keyed every agent on the empty string — one shared
+ * "already told" mark, so only the FIRST agent to touch a canvas was ever told about the user's
+ * edits to it. The rule being in one host and not the other is a bug in both.
+ */
+export declare function resolveSelfAlias(agentId: string | null | undefined): string | undefined;
 export interface PlatformToolsConfig {
     /** Data access layer */
     context: PlatformContext;

package/dist/platform/context.js

@@ -4,6 +4,24 @@
  * Provides a single entry point for data access, abstracting over the
  * underlying storage backend (SQLite in CLI, PostgreSQL in web/API).
  */
+/**
+ * Turn a host's notion of "the agent running now" into an owner alias.
+ *
+ * `'default'` is the coordinator slot every single-agent session uses, not an identity, so it
+ * resolves to `undefined` — which callers then render as their own fallback (`'agent'` for a work
+ * item's author) rather than writing the literal string `default` into stored data.
+ *
+ * ⚠️ This lives here because BOTH hosts need the same answer and only one of them had it. The CLI
+ * implemented the rule inline; Desktop passed no `resolveOwner` at all, so every scene edit
+ * recorded no agent and `consumeUserEdits` keyed every agent on the empty string — one shared
+ * "already told" mark, so only the FIRST agent to touch a canvas was ever told about the user's
+ * edits to it. The rule being in one host and not the other is a bug in both.
+ */
+export function resolveSelfAlias(agentId) {
+    if (!agentId || agentId === 'default')
+        return undefined;
+    return agentId;
+}
 /**
  * Resolve a `cwd` that may be a value or a getter.
  *

package/dist/platform/index.d.ts

@@ -5,6 +5,8 @@ export type { ProjectType, ProjectStatus, RepoPattern, WorkflowMode, LifecycleSt
 export type { IProjectRepository, IWorkItemRepository, IDocumentRepository, IPlanRepository, ICommentRepository, ICanvasRepository, } from './repositories.js';
 export type { IAnchorService, IArtifactService, IEpisodeService, ICanvasRenderer, RenderedCanvasImage, AnchorData, ArtifactType, ArtifactData, ArtifactSummaryData, WorkEpisode, ProjectWorkSummary, } from './services.js';
 export type { PlatformContext, PlatformToolsConfig, PlatformHooks } from './context.js';
+/** Owner-alias rule shared by every host — see its doc comment for why it lives here. */
+export { resolveSelfAlias } from './context.js';
 export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, sceneOutline, createImageTools, } from './tools/index.js';
 export { createSceneWriter, SCENE_JOURNAL_SIZE } from './scene-writer.js';
 export type { ISceneWriter, SceneReadResult, SceneApplyResult, SceneCreateResult, SceneLastEdit, } from './scene-writer.js';

package/dist/platform/index.js

@@ -1,6 +1,8 @@
 /**
  * Platform — Repository interfaces, data models, tools, and workflow.
  */
+/** Owner-alias rule shared by every host — see its doc comment for why it lives here. */
+export { resolveSelfAlias } from './context.js';
 // Platform tools (runtime)
 export { createPlatformTools, createProjectTools, createWorkItemTools, createDocumentTools, createPlanTools, createBacklogTools, createAnchorTools, createArtifactTools, createEpisodeTools, createCanvasTools, createSceneTools, SCENE_TOOL_NAMES, sceneOutline, createImageTools, } from './tools/index.js';
 // The scene writer — one per host process, shared by the scene tools and the host's own

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@compilr-dev/sdk",
-  "version": "0.34.0",
+  "version": "0.36.0",
   "description": "Universal agent runtime for building AI-powered applications",
   "type": "module",
   "main": "dist/index.js",
@@ -81,7 +81,7 @@
     "node": ">=20.0.0"
   },
   "dependencies": {
-    "@compilr-dev/agents": "^0.7.1",
+    "@compilr-dev/agents": "^0.8.0",
     "@compilr-dev/logger": "^0.1.0",
     "ajv": "^6.14.0",
     "yaml": "^2.8.4"