om-memory-system 3.2.0-next.8 → 3.3.0-next.17

package/README.md

@@ -122,7 +122,10 @@ import guides, and the [CLI reference](docs/cli.md).
 Pi shows a notice when a new version is out; update with
 `pi update npm:om-memory-system`. On OpenCode v2, run `opencode plugin check`
 and `opencode plugin update om-memory-system`. Restart the agent afterwards.
-Coming from `opencode-mem`? Your memories move over automatically, with a
+To stay on one version, install it with the number, for example
+`pi install npm:om-memory-system@3.1.1` or
+`opencode plugin add om-memory-system@3.1.1`; a pinned install is never
+updated. Coming from `opencode-mem`? Your memories move over automatically, with a
 backup first. See [Updating and upgrading](docs/upgrading.md) and
 [CHANGELOG.md](CHANGELOG.md).
 

package/dist/adapters/opencode/auto-capture-summary.js

@@ -2,7 +2,7 @@ import { randomUUID } from "node:crypto";
 import { CONFIG } from "../../config.js";
 import { resolveOpencodeHostModel } from "../../services/ai/live-model-choice.js";
 import { buildBoundedSummaryPrompt } from "../../core/capture-context.js";
-import { parseCaptureSummary } from "../../core/extraction.js";
+import { classifyCaptureReply, normalizeStopReason, parseCaptureSummary, } from "../../core/extraction.js";
 import { log } from "../../services/logger.js";
 import { loadOpencodeProvider } from "../../services/ai/opencode-provider-loader.js";
 function buildSystemPrompt(languageName) {
@@ -26,8 +26,19 @@ FORMAT:
 SKIP if: greetings, casual chat, no code/decisions made
 CAPTURE if: code changed, bug fixed, feature added, decision made`;
 }
+/** Reason code for an error thrown on the OpenCode host-model path. */
+function hostModelFailureReason(error) {
+    if (error instanceof Error && error.name === "ZodError")
+        return "schema-mismatch";
+    if (String(error).includes("returned no structured output"))
+        return "empty-text";
+    return "call-error";
+}
 export async function generateOpenCodeAutoCaptureSummary(request, notify) {
     let opencodeProviderError;
+    // Filled per path; when the host model falls back, the external API attempt
+    // overwrites it, so the record describes the call that decided the outcome.
+    const diagnostics = request.diagnostics ?? {};
     const hostModel = resolveOpencodeHostModel(CONFIG);
     if (hostModel) {
         try {
@@ -40,6 +51,9 @@ export async function generateOpenCodeAutoCaptureSummary(request, notify) {
                 ...hostModel,
                 prompt: request.prompt,
             });
+            diagnostics.path = "host-model";
+            diagnostics.provider = providerID;
+            diagnostics.model = modelID;
             if (!isProviderConnected(providerID)) {
                 throw new Error(`opencode provider '${providerID}' is not connected. Check your opencode provider configuration.`);
             }
@@ -59,6 +73,8 @@ export async function generateOpenCodeAutoCaptureSummary(request, notify) {
                 tags: z.array(z.string()),
             });
             const aiPrompt = buildBoundedSummaryPrompt(request.context, systemPrompt, z.toJSONSchema(schema));
+            diagnostics.systemPrompt = systemPrompt;
+            diagnostics.userPrompt = aiPrompt;
             const result = await generateStructuredOutput({
                 client: v2Client,
                 providerID,
@@ -66,6 +82,12 @@ export async function generateOpenCodeAutoCaptureSummary(request, notify) {
                 systemPrompt,
                 userPrompt: aiPrompt,
                 schema,
+                onReply: (reply) => {
+                    diagnostics.stopReason = normalizeStopReason(reply.finish);
+                    diagnostics.blockTypes = reply.partTypes;
+                    diagnostics.rawReply =
+                        reply.structuredOutput === undefined ? "" : JSON.stringify(reply.structuredOutput);
+                },
             });
             return {
                 summary: result.summary,
@@ -75,6 +97,7 @@ export async function generateOpenCodeAutoCaptureSummary(request, notify) {
         }
         catch (error) {
             opencodeProviderError = error;
+            diagnostics.failureReason = hostModelFailureReason(error);
             log("auto-capture: opencode provider failed, falling back to external API", {
                 error: String(error),
             });
@@ -134,19 +157,38 @@ export async function generateOpenCodeAutoCaptureSummary(request, notify) {
     };
     const aiPrompt = buildBoundedSummaryPrompt(request.context, systemPrompt, toolSchema);
     const captureSessionID = `auto-capture-${request.prompt?.id ?? request.sessionId}-${randomUUID()}`;
-    const result = await provider.executeToolCall(systemPrompt, aiPrompt, toolSchema, captureSessionID);
+    // Replace whatever the failed host-model attempt recorded.
+    Object.assign(diagnostics, {
+        path: "external-api",
+        provider: CONFIG.memoryProvider,
+        model: CONFIG.memoryModel,
+        systemPrompt,
+        userPrompt: aiPrompt,
+        stopReason: undefined,
+        blockTypes: undefined,
+        rawReply: undefined,
+        failureReason: undefined,
+    });
+    let result;
+    try {
+        result = await provider.executeToolCall(systemPrompt, aiPrompt, toolSchema, captureSessionID);
+    }
+    catch (error) {
+        diagnostics.failureReason = "call-error";
+        throw error;
+    }
+    diagnostics.stopReason = normalizeStopReason(result.stopReason);
     if (!result.success || !result.data) {
+        diagnostics.failureReason = diagnostics.stopReason === "length" ? "truncated" : "call-error";
         throw new Error(result.error || "Failed to generate summary");
     }
     const rawReply = JSON.stringify(result.data);
+    diagnostics.rawReply = rawReply;
     const summary = parseCaptureSummary(rawReply);
     if (!summary) {
-        log("OpenCode capture: model reply was not a valid capture summary", {
-            provider: CONFIG.memoryProvider,
-            modelId: CONFIG.memoryModel,
-            // The reply can carry conversation content, so log only its size.
-            replyLength: rawReply.length,
-        });
+        // The capture pipeline logs the attempt's metadata; never the reply.
+        diagnostics.failureReason =
+            classifyCaptureReply({ text: rawReply, stopReason: result.stopReason }) ?? "schema-mismatch";
         throw new Error("omms: OpenCode extraction returned an invalid summary payload");
     }
     return summary;

package/dist/adapters/opencode/import-command.d.ts

@@ -1,4 +1,6 @@
-import { type HistoryImportModels } from "../../importer/run-import.js";
+import { type OpencodeModelRef } from "../../services/ai/opencode-import-models.js";
+export { createOpencodeImportModels } from "../../services/ai/opencode-import-models.js";
+export type { OpencodeModelRef } from "../../services/ai/opencode-import-models.js";
 /**
  * OpenCode command surface for the historical-session importer:
  *
@@ -12,10 +14,6 @@ import { type HistoryImportModels } from "../../importer/run-import.js";
 export declare const OPENCODE_IMPORT_COMMAND = "memory-import-opencode-history";
 export declare const OPENCODE_IMPORT_DESCRIPTION = "Import historical OpenCode sessions into the shared memory store (try --dry-run first)";
 export declare const OPENCODE_IMPORT_USAGE: string;
-export interface OpencodeModelRef {
-    providerID: string;
-    modelID: string;
-}
 export interface OpencodeImportCommandInput {
     argsText: string;
     /** Session directory; relative paths and the default project resolve here. */
@@ -24,8 +22,6 @@ export interface OpencodeImportCommandInput {
     sessionModel: () => Promise<OpencodeModelRef | null>;
     notify?: (message: string) => void;
 }
-/** Capture and profile calls routed through OpenCode's structured-output sessions. */
-export declare function createOpencodeImportModels(ref: OpencodeModelRef, directory: string): Promise<Required<HistoryImportModels>>;
 /** Parse, run, and describe one import; always resolves to text for the session. */
 export declare function runOpencodeImportCommand(input: OpencodeImportCommandInput): Promise<string>;
 /** The model of the session's latest user message, from the V1 message list. */

package/dist/adapters/opencode/import-command.js

@@ -1,7 +1,8 @@
-import { CONFIG, isConfigured } from "../../config.js";
-import { buildBoundedSummaryPrompt } from "../../core/capture-context.js";
+import { isConfigured } from "../../config.js";
 import { historyImportUsage, importNeedsModel, parseHistoryImportArgs, tokenizeImportArgs, } from "../../importer/import-args.js";
 import { formatHistoryImportReport, summarizeHistoryImportReport, runHistoryImport, } from "../../importer/run-import.js";
+import { createOpencodeImportModels, } from "../../services/ai/opencode-import-models.js";
+export { createOpencodeImportModels } from "../../services/ai/opencode-import-models.js";
 import { loadOpencodeProvider } from "../../services/ai/opencode-provider-loader.js";
 import { memoryClient } from "../../services/client.js";
 import { log } from "../../services/logger.js";
@@ -23,58 +24,6 @@ function splitModel(value) {
     const separator = value.indexOf("/");
     return { providerID: value.slice(0, separator), modelID: value.slice(separator + 1) };
 }
-/** Capture and profile calls routed through OpenCode's structured-output sessions. */
-export async function createOpencodeImportModels(ref, directory) {
-    const { getV2Client, generateStructuredOutput } = await loadOpencodeProvider();
-    const client = getV2Client();
-    if (!client)
-        throw new Error("the OpenCode client is not ready; retry in a moment");
-    const { z } = await import("zod");
-    const { buildCaptureSystemPrompt, createUserProfileAnalysisSchema, parseCaptureSummary } = await import("../../core/extraction.js");
-    const { detectLanguage, getLanguageName } = await import("../../services/language-detector.js");
-    const captureSchema = z.object({
-        summary: z.string().optional(),
-        type: z.string(),
-        tags: z.array(z.string()).optional(),
-    });
-    const capture = {
-        async summarize(request) {
-            const target = CONFIG.autoCaptureLanguage && CONFIG.autoCaptureLanguage !== "auto"
-                ? CONFIG.autoCaptureLanguage
-                : detectLanguage(request.userPrompt);
-            const systemPrompt = buildCaptureSystemPrompt(getLanguageName(target));
-            const result = await generateStructuredOutput({
-                client,
-                ...ref,
-                systemPrompt,
-                userPrompt: buildBoundedSummaryPrompt(request.context, systemPrompt, z.toJSONSchema(captureSchema)),
-                schema: captureSchema,
-                directory,
-            });
-            const parsed = parseCaptureSummary(JSON.stringify(result));
-            if (!parsed)
-                throw new Error("History capture returned an invalid summary");
-            return parsed;
-        },
-    };
-    const profileSchema = createUserProfileAnalysisSchema(z);
-    const profile = {
-        provider: ref.providerID,
-        modelId: ref.modelID,
-        async complete(systemPrompt, userPrompt) {
-            const result = await generateStructuredOutput({
-                client,
-                ...ref,
-                systemPrompt,
-                userPrompt,
-                schema: profileSchema,
-                directory,
-            });
-            return JSON.stringify(result);
-        },
-    };
-    return { capture, profile };
-}
 /** Parse, run, and describe one import; always resolves to text for the session. */
 export async function runOpencodeImportCommand(input) {
     const name = OPENCODE_IMPORT_COMMAND;

package/dist/adapters/pi/extension.js

@@ -1,5 +1,6 @@
 import { Type } from "typebox";
 import { CONFIG, initConfigWithLegacyMigration, isConfigured } from "../../config.js";
+import { pruneTraces } from "../../services/capture-diagnostics.js";
 import { executeMemoryOperation } from "../../core/memory-operations.js";
 import { getLanguageName } from "../../services/language-detector.js";
 import { log } from "../../services/logger.js";
@@ -97,6 +98,8 @@ export default function ommsPiExtension(pi) {
         try {
             latestCtx = ctx;
             initConfigWithLegacyMigration(ctx.cwd);
+            // Runs even with tracing off, so turning it off does not leave old traces behind.
+            pruneTraces(CONFIG);
             captureState = createPiCaptureState();
             const globalScope = globalThis;
             if (!globalScope[GLOBAL_PLUGIN_WARMUP_KEY] && isConfigured()) {

package/dist/adapters/pi/profile.js

@@ -1,4 +1,4 @@
-import { CONFIG } from "../../config.js";
+import { CONFIG, refreshConfigIfChanged } from "../../config.js";
 import { analyzeProfile, } from "../../core/profile-analysis.js";
 import { log } from "../../services/logger.js";
 import { getTags } from "../../services/tags.js";
@@ -47,6 +47,7 @@ export async function performPiProfileLearning(input) {
     if (input.prompts.length === 0)
         return;
     try {
+        refreshConfigIfChanged(input.directory);
         const tags = getTags(input.directory);
         if (!tags.user.userEmail) {
             log("pi profile learning: skipped (no user email resolved)", {

package/dist/adapters/pi/provider.js

@@ -1,8 +1,7 @@
 import { CONFIG } from "../../config.js";
 import { buildBoundedSummaryPrompt } from "../../core/capture-context.js";
-import { buildCaptureSystemPrompt, captureSummaryToolSchema, parseCaptureSummary, } from "../../core/extraction.js";
+import { buildCaptureReplyInstruction, buildCaptureSystemPrompt, captureSummaryToolSchema, classifyCaptureReply, normalizeStopReason, parseCaptureSummary, } from "../../core/extraction.js";
 import { detectLanguage, getLanguageName } from "../../services/language-detector.js";
-import { log } from "../../services/logger.js";
 /**
  * Active-model-first resolution: the explicit `piProvider`/`piModel`
  * configuration when both are set, otherwise the active `ctx.model`.
@@ -80,23 +79,51 @@ export function createPiCaptureProvider(resolveModel) {
                 ? detectLanguage(request.userPrompt)
                 : CONFIG.autoCaptureLanguage;
             const systemPrompt = buildCaptureSystemPrompt(getLanguageName(targetLang));
-            const userPrompt = buildBoundedSummaryPrompt(request.context, systemPrompt, captureSummaryToolSchema);
-            const reply = await model.complete({
-                systemPrompt,
-                messages: [{ role: "user", content: userPrompt, timestamp: Date.now() }],
-            });
+            const replyInstruction = buildCaptureReplyInstruction();
+            // The instruction is counted as schema bytes so the context budget leaves room for it.
+            const userPrompt = `${buildBoundedSummaryPrompt(request.context, systemPrompt, {
+                schema: captureSummaryToolSchema,
+                replyInstruction,
+            })}\n\n${replyInstruction}`;
+            const diagnostics = request.diagnostics;
+            if (diagnostics) {
+                diagnostics.path = "host-model";
+                diagnostics.provider = model.provider;
+                diagnostics.model = model.modelId;
+                diagnostics.systemPrompt = systemPrompt;
+                diagnostics.userPrompt = userPrompt;
+            }
+            let reply;
+            try {
+                reply = await model.complete({
+                    systemPrompt,
+                    messages: [{ role: "user", content: userPrompt, timestamp: Date.now() }],
+                });
+            }
+            catch (error) {
+                if (diagnostics)
+                    diagnostics.failureReason = "call-error";
+                throw error;
+            }
+            const rawReply = replyText(reply);
+            if (diagnostics) {
+                diagnostics.stopReason = normalizeStopReason(reply.stopReason);
+                diagnostics.blockTypes = reply.content.map((block) => block.type);
+                diagnostics.rawReply = rawReply;
+            }
             if (reply.stopReason === "error") {
+                if (diagnostics)
+                    diagnostics.failureReason = "call-error";
                 throw new Error(`omms: Pi extraction call failed: ${reply.errorMessage || "unknown error"}`);
             }
-            const rawReply = replyText(reply);
             const summary = parseCaptureSummary(rawReply);
             if (!summary) {
-                log("Pi capture: model reply was not a valid capture summary", {
-                    provider: model.provider,
-                    modelId: model.modelId,
-                    // The reply can carry conversation content, so log only its size.
-                    replyLength: rawReply.length,
-                });
+                // The capture pipeline logs the attempt's metadata; never the reply.
+                if (diagnostics) {
+                    diagnostics.failureReason =
+                        classifyCaptureReply({ text: rawReply, stopReason: reply.stopReason }) ??
+                            "schema-mismatch";
+                }
                 throw new Error("omms: Pi extraction returned an invalid summary payload");
             }
             return summary;

package/dist/config.d.ts

@@ -38,6 +38,10 @@ interface OmmsConfig {
     piProvider?: string;
     piModel?: string;
     aiSessionRetentionDays?: number;
+    /** Write full capture prompts and replies to ~/.omms/traces. Global config only. */
+    captureTrace?: boolean;
+    captureTraceRetentionDays?: number;
+    captureAttemptRetentionDays?: number;
     webServerEnabled?: boolean;
     webServerPort?: number;
     webServerHost?: string;
@@ -85,6 +89,7 @@ interface OmmsConfig {
         injectOn?: "first" | "always";
     };
 }
+export declare const CONFIG_TEMPLATE = "{\n  // ============================================\n  // omms (Opinionated Modular Memory System) Configuration\n  // ============================================\n  \n  // Storage location for the vector database. Leave unset to use the\n  // default ~/.omms/data (a legacy ~/.opencode-mem/data store is migrated\n  // there automatically on first start; see docs/omms-migration.md).\n  // \"storagePath\": \"~/.omms/data\",\n\n  \"userEmailOverride\": \"\",\n  \"userNameOverride\": \"\",\n  \n  // ============================================\n  // Embedding Model (for similarity search)\n  // ============================================\n  // Local = Hugging Face / ONNX via @huggingface/transformers (not Apple MLX).\n  // Remote = set BOTH embeddingApiUrl and embeddingApiKey (OpenAI-compatible /embeddings).\n  \n  // Default: Nomic Embed v1 (768 dimensions, 8192 context, multilingual)\n  \"embeddingModel\": \"Xenova/nomic-embed-text-v1\",\n\n  // Opt-in Nomic task prefixes (search_document: / search_query:). After enabling,\n  // re-index existing memories so store and query vectors stay aligned.\n  // \"embeddingUseTaskPrefixes\": true,\n  \n  // Auto-detected dimensions (no need to set manually)\n  // \"embeddingDimensions\": 768,\n  \n  // Other recommended local models:\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-base-en\",  // 768 dims, English-only, 8192 context\n  // \"embeddingModel\": \"Xenova/jina-embeddings-v2-small-en\", // 512 dims, faster, 8192 context\n  // \"embeddingModel\": \"Xenova/all-MiniLM-L6-v2\",            // 384 dims, very fast, 512 context\n  // \"embeddingModel\": \"Xenova/all-mpnet-base-v2\",           // 768 dims, good quality, 512 context\n  \n  // Optional: OpenAI-compatible API for embeddings (both URL and key required)\n  // \"embeddingApiUrl\": \"https://api.openai.com/v1\",\n  // \"embeddingApiKey\": \"env://OPENAI_API_KEY\",  // or \"sk-...\" / \"file:///path/to/key\"\n  // \"embeddingModel\": \"text-embedding-3-small\",  // 1536 dims, auto-detected\n \n  // ============================================\n  // Web Server Settings\n  // ============================================\n  \n  // Enable web UI for managing memories (accessible at http://localhost:4747)\n  \"webServerEnabled\": true,\n  \n  // Port for web UI server\n  \"webServerPort\": 4747,\n  \n  // Host address for web UI (use 127.0.0.1 for local only, 0.0.0.0 for network access)\n  \"webServerHost\": \"127.0.0.1\",\n\n  // Optional HTTP Basic Auth for the web UI. Accepts literal, env://, or file:// secrets.\n  // \"webServerAuthPassword\": \"\",\n  // \"webServerAuthUsername\": \"\",\n\n  // Required when webServerHost is not loopback. Protects /api/* with Bearer / X-Opencode-Mem-Token.\n  // \"webServerApiToken\": \"env://OMMS_WEB_TOKEN\",\n  \n  // ============================================\n  // Database Settings\n  // ============================================\n  \n  // Maximum vectors per database shard (auto-creates new shard when limit reached)\n  \"maxVectorsPerShard\": 50000,\n  \n  // Automatically delete old memories based on retention period\n  \"autoCleanupEnabled\": true,\n  \n  // Days to keep memories before auto-cleanup (only if autoCleanupEnabled is true)\n  \"autoCleanupRetentionDays\": 30,\n  \n  // Automatically detect and remove duplicate memories\n  \"deduplicationEnabled\": true,\n  \n   // Similarity threshold (0-1) for detecting duplicates (higher = stricter)\n   \"deduplicationSimilarityThreshold\": 0.90,\n   \n  // ============================================\n  // Memory Scope Settings\n  // ============================================\n\n  // Default scope for memory list/search queries\n  // \"project\" keeps queries within the current project, \"all-projects\" searches across all project shards\n  \"memory\": {\n    \"defaultScope\": \"project\"\n  },\n\n  // ============================================\n  // Model for auto-capture and profile learning (same rule in OpenCode and Pi)\n  // ============================================\n\n   // Which model summarises your work, in this order:\n   //   1. The host model below: opencodeProvider/opencodeModel in OpenCode,\n   //      piProvider/piModel in Pi. Set the model to \"inherit\" to follow\n   //      whatever model the session is using.\n   //   2. If no host model is set: the external API (memoryModel/memoryApiUrl/\n   //      memoryApiKey further down).\n   //   3. If neither is set: the session's own model.\n   // If the host model fails and the external API is configured, the external\n   // API is used instead.\n   //\n   // Host models go through OpenCode's or Pi's own sign-in (OAuth like Claude\n   // Pro/Max, GitHub Copilot, ChatGPT, or your own key there), so no separate API\n   // key is needed here.\n   //\n   // Examples (OpenCode names from 'opencode providers list'; Pi names from its model list):\n   //   \"opencodeProvider\": \"anthropic\",    \"opencodeModel\": \"claude-haiku-4-5-20251001\"\n   //   \"opencodeProvider\": \"openai\",       \"opencodeModel\": \"inherit\"\n   //   \"piProvider\": \"openai-codex\",       \"piModel\": \"gpt-5.6-luna\"\n   //\n   // \"opencodeProvider\": \"anthropic\",\n   // \"opencodeModel\": \"claude-haiku-4-5-20251001\",\n   // \"piProvider\": \"openai-codex\",\n   // \"piModel\": \"gpt-5.6-luna\",\n\n   // ============================================\n   // Auto-Capture Settings\n   // ============================================\n  \n  // Auto-capture runs in the background without blocking your main session.\n  // Note: Ollama may not support tool calling. Use OpenAI, Anthropic, or Groq for best results.\n  \n  \"autoCaptureEnabled\": true,\n  \n  // Provider type: \"openai-chat\" | \"openai-responses\" | \"anthropic\" | \"minimax\" | \"orcarouter\"\n  // Note: \"openai-chat\" is a generic OpenAI API-compatible mode.\n  // Any service that follows the OpenAI Chat Completions API can use it via custom \"memoryApiUrl\".\n  \"memoryProvider\": \"openai-chat\",\n  \n  // External API (step 2 above, and the fallback when a host model fails).\n  // Uncomment all 3 lines and replace memoryApiKey before use:\n  // \"memoryModel\": \"gpt-4o-mini\",\n  // \"memoryApiUrl\": \"https://api.openai.com/v1\",\n  // \"memoryApiKey\": \"sk-...\",\n\n  // API Key Formats:\n  // Direct value:        \"sk-...\"\n  // From file:           \"file://~/.config/litellm-key.txt\"\n  // From env variable:   \"env://LITELLM_API_KEY\"\n  \n  // Examples for different providers:\n  // Any OpenAI-compatible endpoint can use the \"openai-chat\" provider pattern below.\n  // Common examples: DeepSeek, Qwen (via Alibaba Cloud ModelStudio),\n  // Zhipu GLM (BigModel platform), and Kimi (Moonshot AI platform).\n\n  // OpenAI Chat Completion (default, backward compatible):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"gpt-4o-mini\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n\n  // DeepSeek (OpenAI-compatible example):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"deepseek-chat\"\n  //   \"memoryApiUrl\": \"https://api.deepseek.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // OpenAI Responses API (recommended, with session support):\n  //   \"memoryProvider\": \"openai-responses\"\n  //   \"memoryModel\": \"gpt-4o\"\n  //   \"memoryApiUrl\": \"https://api.openai.com/v1\"\n  //   \"memoryApiKey\": \"sk-...\"\n  \n  // Anthropic (with session support):\n  //   \"memoryProvider\": \"anthropic\"\n  //   \"memoryModel\": \"claude-3-5-haiku-20241022\"\n  //   \"memoryApiUrl\": \"https://api.anthropic.com/v1\"\n  //   \"memoryApiKey\": \"sk-ant-...\"\n\n  // MiniMax (Anthropic Messages-compatible endpoint, with session support):\n  //   \"memoryProvider\": \"minimax\"\n  //   \"memoryModel\": \"MiniMax-M3\"\n  //   \"memoryApiUrl\": \"https://api.minimax.io\"        // global endpoint\n  //   \"memoryApiKey\": \"<MiniMax API key>\"\n  //   // China endpoint: \"memoryApiUrl\": \"https://api.minimaxi.com\"\n  //   // Optional adaptive thinking for MiniMax-M3:\n  //   \"memoryExtraParams\": { \"thinking\": { \"type\": \"adaptive\" } }\n\n  // OrcaRouter (OpenAI-compatible gateway, namespaced model IDs, with session support):\n  //   \"memoryProvider\": \"orcarouter\"\n  //   \"memoryApiKey\": \"<OrcaRouter API key>\"\n  //   // memoryApiUrl and memoryModel are optional \u2014 they default to\n  //   // https://api.orcarouter.ai/v1 and \"orcarouter/auto\" (a routing alias).\n  //   // OrcaRouter rejects bare model names, so if you set memoryModel, use a\n  //   // namespaced ID such as \"openai/gpt-5.5\" or \"deepseek/deepseek-v4-flash\".\n  //   \"memoryModel\": \"openai/gpt-5.5\"\n\n  // Groq (OpenAI-compatible, use openai-chat provider):\n  //   \"memoryProvider\": \"openai-chat\"\n  //   \"memoryModel\": \"llama-3.3-70b-versatile\"\n  //   \"memoryApiUrl\": \"https://api.groq.com/openai/v1\"\n  //   \"memoryApiKey\": \"gsk_...\"\n  \n  // Maximum iterations for multi-turn AI analysis (for openai-responses, anthropic, and minimax)\n  \"autoCaptureMaxIterations\": 5,\n   \n  // Timeout per iteration in milliseconds (30 seconds default)\n  \"autoCaptureIterationTimeout\": 30000,\n\n  // Maximum number of times to retry capturing a prompt if it fails (due to network, API errors, etc.)\n  \"autoCaptureMaxRetries\": 3,\n\n  // Maximum UTF-8 bytes for the auto-capture markdown context sent to the summary model.\n  // Prevents HTTP 400 context overflows on models with ~131K token windows (e.g. Groq Llama).\n  // Rough guide: tokens \u2248 bytes / 4 for mixed code/prose.\n  \"autoCaptureMaxContextBytes\": 131072,\n   \n  // Days to keep AI session history before cleanup\n  \"aiSessionRetentionDays\": 7,\n\n  // Capture diagnostics: every capture attempt always writes one metadata line\n  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no\n  // conversation text. Set captureTrace to true to also write each attempt's\n  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain\n  // conversation content: <private> text and common API key formats are\n  // redacted, files are readable only by you, and they are deleted after\n  // captureTraceRetentionDays. Only this global file can turn tracing on.\n  // \"captureTrace\": false,\n  // \"captureTraceRetentionDays\": 7,\n  // \"captureAttemptRetentionDays\": 30,\n\n  // Temperature for AI API requests (set to false to omit parameter for models that don't support it)\n  // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter\n  // Set to false and add \"memoryTemperature\": false in config when using such models\n  \"memoryTemperature\": 0.3,\n\n  // Extra parameters to include in API request body\n  // Useful for local inference servers (e.g. llama-server with --jinja) that support\n  // additional parameters like disabling thinking/reasoning mode\n  // Example for Qwen3 models: { \"enable_thinking\": false }\n  // \"memoryExtraParams\": {},\n\n  // Language for auto-capture summaries (default: \"auto\" for auto-detection)\n  // Options: \"auto\", \"en\", \"id\", \"zh\", \"ja\", \"es\", \"fr\", \"de\", \"ru\", \"pt\", \"ar\", \"ko\"\n  // \"autoCaptureLanguage\": \"auto\",\n\n  // ============================================\n  // Toast Notifications\n  // ============================================\n\n  // Show toast when memory is auto-captured\n  \"showAutoCaptureToasts\": true,\n\n  // Show toast when user profile is updated\n  \"showUserProfileToasts\": true,\n\n  // Show toast for error messages\n  \"showErrorToasts\": true,\n\n  // ============================================\n  // User Profile System\n  // ============================================\n\n  // Analyze user prompts every N prompts to build/update your user profile\n  // When N uncaptured prompts accumulate, AI will analyze them to identify:\n  // - User preferences (code style, communication style, tool preferences)\n  // - User patterns (recurring topics, problem domains, technical interests)\n  // - User workflows (development habits, sequences, learning style)\n  // - Skill level (overall and per-domain assessment)\n  \"userProfileAnalysisInterval\": 10,\n\n  // Days before inactive items (all types) are eligible for removal\n  \"userProfileStaleDays\": 2,\n\n  // Number of preferences shown in UI\n  \"userProfileDisplayPreferences\": 20,\n  \n  // Number of patterns shown in UI\n  \"userProfileDisplayPatterns\": 15,\n  \n  // Number of workflows shown in UI\n  \"userProfileDisplayWorkflows\": 10,\n  \n  // Number of preferences injected into LLM conversation context\n  // Keep this small \u2014 the strongest signals are enough; more dilute LLM attention\n  \"userProfileInjectPreferences\": 5,\n  \n  // Number of patterns injected into LLM conversation context\n  \"userProfileInjectPatterns\": 5,\n  \n  // Number of workflows injected into LLM conversation context\n  \"userProfileInjectWorkflows\": 3,\n  \n  // Days before preference confidence starts to decay (if not reinforced)\n  // Preferences that aren't seen again will gradually lose confidence and be removed\n  \"userProfileConfidenceDecayDays\": 30,\n  \n  // Number of profile versions to keep in changelog (for rollback/debugging)\n  // Older versions are automatically cleaned up\n  \"userProfileChangelogRetentionCount\": 5,\n\n  // Minimum evidence count for a preference/pattern to survive confidence decay\n  // Items confirmed fewer times are more likely to be pruned when confidence decays\n  \"userProfileMinEvidenceForRetention\": 3,\n\n  // Periodically merge duplicate or irrelevant profile items with the configured AI provider\n  \"userProfileAutoCleanupEnabled\": true,\n  // Number of analyzed user prompts between automatic AI cleanup runs\n  \"userProfileAutoCleanupInterval\": 100,\n\n  // Enable LLM validation of existing preferences against recent behavior.\n  // When enabled, each analysis round checks if top-5 preferences still match recent prompts.\n  // Experimental \u2014 disabled by default.\n  \"userProfileValidationEnabled\": false,\n\n  // ============================================\n  // Search Settings\n  // ============================================\n  \n  // Minimum similarity score (0-1) for memory search results\n  \"similarityThreshold\": 0.6,\n\n  // Maximum number of memories to return in search results\n  \"maxMemories\": 10,\n\n  // ============================================\n  // Advanced Settings\n  // ============================================\n  \n  // Inject user profile into AI context (preferences, patterns, workflows)\n  \"injectProfile\": true\n}\n";
 export declare function normalizeAutoCaptureMaxContextBytes(value: number): number;
 export declare function normalizeAutoCleanupRetentionDays(value: number): number;
 declare function buildConfig(fileConfig: OmmsConfig): {
@@ -119,6 +124,9 @@ declare function buildConfig(fileConfig: OmmsConfig): {
     piModel: string | undefined;
     autoCaptureProviderStatus: import("./config.js").AutoCaptureProviderStatus;
     aiSessionRetentionDays: number;
+    captureTrace: boolean;
+    captureTraceRetentionDays: number;
+    captureAttemptRetentionDays: number;
     webServerEnabled: boolean;
     webServerPort: number;
     webServerHost: string;
@@ -169,6 +177,9 @@ declare function buildConfig(fileConfig: OmmsConfig): {
         injectOn: "first" | "always";
     };
 };
+export declare function getGlobalConfigSourcePath(): string | undefined;
+export declare function getGlobalConfigWritePath(): string;
+export declare function validateGlobalConfig(value: unknown): void;
 export declare let CONFIG: {
     storagePath: string;
     userEmailOverride: string | undefined;
@@ -201,6 +212,9 @@ export declare let CONFIG: {
     piModel: string | undefined;
     autoCaptureProviderStatus: import("./config.js").AutoCaptureProviderStatus;
     aiSessionRetentionDays: number;
+    captureTrace: boolean;
+    captureTraceRetentionDays: number;
+    captureAttemptRetentionDays: number;
     webServerEnabled: boolean;
     webServerPort: number;
     webServerHost: string;
@@ -255,7 +269,10 @@ type RuntimeConfig = ReturnType<typeof buildConfig>;
 export { getAutoCaptureProviderStatus, type AutoCaptureProviderStatus, } from "./services/ai/live-model-choice.js";
 export { isPlaceholderApiKey };
 export declare function hasAutoCaptureProviderConfig(config?: RuntimeConfig): boolean;
-export declare function initConfig(directory: string): void;
+export declare function refreshConfigIfChanged(directory: string): void;
+export declare function initConfig(directory: string, options?: {
+    strict?: boolean;
+}): void;
 /**
  * initConfig plus the one-time legacy store migration (decision D13).
  *

package/dist/config.js

@@ -1,4 +1,4 @@
-import { existsSync, readFileSync, mkdirSync, writeFileSync } from "node:fs";
+import { existsSync, readFileSync, mkdirSync, writeFileSync, statSync } from "node:fs";
 import { join } from "node:path";
 import { homedir } from "node:os";
 import { stripJsoncComments } from "./services/jsonc.js";
@@ -48,6 +48,9 @@ const DEFAULTS = {
     autoCaptureMaxRetries: 3,
     autoCaptureMaxContextBytes: 131072,
     aiSessionRetentionDays: 7,
+    captureTrace: false,
+    captureTraceRetentionDays: 7,
+    captureAttemptRetentionDays: 30,
     webServerEnabled: true,
     webServerPort: 4747,
     webServerHost: "127.0.0.1",
@@ -108,8 +111,9 @@ function expandPath(path) {
  * Load the first config file that exists, in priority order. A file that
  * exists but cannot be read or parsed is not skipped: falling through would let
  * a lower-priority (legacy) file silently supply settings such as storagePath.
+ * `strict` (live reload) throws instead, so running hosts keep their settings.
  */
-function loadConfigFromPaths(paths) {
+function loadConfigFromPaths(paths, strict = false) {
     const path = paths.find((candidate) => existsSync(candidate));
     if (!path)
         return {};
@@ -119,6 +123,9 @@ function loadConfigFromPaths(paths) {
         return JSON.parse(json);
     }
     catch (error) {
+        // The parser's message can quote file content, so a reload names the file only.
+        if (strict)
+            throw new Error(`Config file cannot be parsed: ${path}`, { cause: error });
         log("Config file is invalid; using defaults instead of lower-priority files", {
             path,
             error: String(error),
@@ -140,7 +147,7 @@ function assertProjectRemoteProviderConfigIsSafe(projectConfig) {
             `Move them to the global config at ${CONFIG_FILES[0]}.`);
     }
 }
-const CONFIG_TEMPLATE = `{
+export const CONFIG_TEMPLATE = `{
   // ============================================
   // omms (Opinionated Modular Memory System) Configuration
   // ============================================
@@ -352,6 +359,17 @@ const CONFIG_TEMPLATE = `{
   // Days to keep AI session history before cleanup
   "aiSessionRetentionDays": 7,
 
+  // Capture diagnostics: every capture attempt always writes one metadata line
+  // (model, stop reason, sizes, outcome) to ~/.omms/omms.log, with no
+  // conversation text. Set captureTrace to true to also write each attempt's
+  // prompt and raw reply to ~/.omms/traces/ for debugging. Traces can contain
+  // conversation content: <private> text and common API key formats are
+  // redacted, files are readable only by you, and they are deleted after
+  // captureTraceRetentionDays. Only this global file can turn tracing on.
+  // "captureTrace": false,
+  // "captureTraceRetentionDays": 7,
+  // "captureAttemptRetentionDays": 30,
+
   // Temperature for AI API requests (set to false to omit parameter for models that don't support it)
   // Some reasoning models (like o1, o3, gpt-5) don't support temperature parameter
   // Set to false and add "memoryTemperature": false in config when using such models
@@ -580,6 +598,9 @@ function buildConfig(fileConfig) {
             memoryApiKey,
         }),
         aiSessionRetentionDays: fileConfig.aiSessionRetentionDays ?? DEFAULTS.aiSessionRetentionDays,
+        captureTrace: fileConfig.captureTrace === true,
+        captureTraceRetentionDays: Math.max(1, Math.floor(fileConfig.captureTraceRetentionDays ?? DEFAULTS.captureTraceRetentionDays)),
+        captureAttemptRetentionDays: normalizeAutoCleanupRetentionDays(fileConfig.captureAttemptRetentionDays ?? DEFAULTS.captureAttemptRetentionDays),
         webServerEnabled: fileConfig.webServerEnabled ?? DEFAULTS.webServerEnabled,
         webServerPort: fileConfig.webServerPort ?? DEFAULTS.webServerPort,
         webServerHost: fileConfig.webServerHost ?? DEFAULTS.webServerHost,
@@ -638,6 +659,21 @@ function buildConfig(fileConfig) {
         },
     };
 }
+export function getGlobalConfigSourcePath() {
+    return CONFIG_FILES.find((candidate) => existsSync(candidate));
+}
+export function getGlobalConfigWritePath() {
+    const source = getGlobalConfigSourcePath();
+    return source && !LEGACY_CONFIG_FILES.includes(source)
+        ? source
+        : join(OMMS_CONFIG_DIR, "omms.jsonc");
+}
+export function validateGlobalConfig(value) {
+    if (!value || typeof value !== "object" || Array.isArray(value)) {
+        throw new Error("Global config must be an object");
+    }
+    buildConfig(value);
+}
 const _globalFileConfig = loadConfigFromPaths(CONFIG_FILES);
 let lastFileConfig = _globalFileConfig;
 export let CONFIG = buildConfig(_globalFileConfig);
@@ -646,23 +682,81 @@ export { isPlaceholderApiKey };
 export function hasAutoCaptureProviderConfig(config = CONFIG) {
     return getAutoCaptureProviderStatus(config).ready;
 }
-export function initConfig(directory) {
-    // omms project overrides win; the legacy opencode-mem file is still read.
-    const projectPaths = [
+function projectConfigPaths(directory) {
+    return [
         join(directory, ".opencode", "omms.jsonc"),
         join(directory, ".opencode", "omms.json"),
         join(directory, ".opencode", "opencode-mem.jsonc"),
         join(directory, ".opencode", "opencode-mem.json"),
     ];
-    const globalConfig = loadConfigFromPaths(CONFIG_FILES);
-    const projectConfig = loadConfigFromPaths(projectPaths);
+}
+function configSignature(directory) {
+    return [...CONFIG_FILES, ...projectConfigPaths(directory)]
+        .filter((path) => existsSync(path))
+        .map((path) => {
+        try {
+            const stat = statSync(path);
+            return `${path}:${stat.mtimeMs}:${stat.size}`;
+        }
+        catch (error) {
+            if (error.code === "ENOENT")
+                return `${path}:missing`;
+            throw error;
+        }
+    })
+        .join("|");
+}
+let lastConfigDirectory;
+let lastConfigSignature;
+/** Directory and signature of the last reload that failed. */
+let lastFailedConfig;
+export function refreshConfigIfChanged(directory) {
+    // Both hosts initialise their config at startup. Direct callers that have
+    // not initialised it keep their in-memory settings untouched.
+    if (lastConfigDirectory === undefined)
+        return;
+    let attempt;
+    try {
+        const signature = configSignature(directory);
+        if (directory === lastConfigDirectory && signature === lastConfigSignature)
+            return;
+        attempt = `${directory}\0${signature}`;
+        if (attempt === lastFailedConfig)
+            return;
+        initConfig(directory, { strict: true });
+        lastFailedConfig = undefined;
+    }
+    catch (error) {
+        // A bad hand edit must not stop capture: keep the last good settings, and
+        // skip this exact file state until it changes, instead of re-reading it on
+        // every unit. The last good directory and signature stay as they were.
+        log("Config reload failed; keeping the previous settings", {
+            error: error instanceof Error ? error.message : String(error),
+        });
+        lastFailedConfig = attempt;
+    }
+}
+export function initConfig(directory, options = {}) {
+    // omms project overrides win; the legacy opencode-mem file is still read.
+    const globalConfig = loadConfigFromPaths(CONFIG_FILES, options.strict);
+    const projectConfig = loadConfigFromPaths(projectConfigPaths(directory), options.strict);
     assertProjectRemoteProviderConfigIsSafe(projectConfig);
     const projectOverrides = { ...projectConfig };
     delete projectOverrides.autoCleanupEnabled;
     delete projectOverrides.autoCleanupRetentionDays;
+    // A checked-in project file must not start recording conversations; it may
+    // only turn tracing off for its project.
+    if (projectOverrides.captureTrace === true) {
+        log("Project config cannot turn on captureTrace; the value was ignored", { directory });
+        delete projectOverrides.captureTrace;
+    }
+    delete projectOverrides.captureTraceRetentionDays;
     const merged = { ...globalConfig, ...projectOverrides };
+    const nextConfig = buildConfig(merged);
     lastFileConfig = merged;
-    CONFIG = buildConfig(merged);
+    CONFIG = nextConfig;
+    lastConfigDirectory = directory;
+    lastConfigSignature = configSignature(directory);
 }
 /**
  * Rebuild CONFIG from the last loaded file config. Used after the legacy

package/dist/core/capture-context.js

@@ -4,7 +4,7 @@ const DEFAULT_AUTO_CAPTURE_MAX_CONTEXT_BYTES = 131072;
 const CONTEXT_TRUNCATION_MARKER = "\n[... truncated to autoCaptureMaxContextBytes ...]\n";
 const SUMMARY_REQUEST_OVERHEAD_BYTES = 1024;
 const SUMMARY_OUTPUT_RESERVE_BYTES = 16384;
-const SUMMARY_ANALYSIS_SUFFIX = 'Analyze this conversation. If it contains technical work (code, bugs, features, decisions), create a concise summary and relevant tags. If it\'s non-technical (greetings, casual chat, incomplete requests), return type="skip" with empty summary.';
+const SUMMARY_ANALYSIS_SUFFIX = 'Analyze this conversation. If it contains technical work (code, bugs, features, decisions), create a concise summary and relevant tags. If it\'s non-technical (greetings, casual chat, incomplete requests), set "type" to "skip" with an empty "summary".';
 function fitTextResponses(textResponses, maxBytes) {
     if (textResponses.length === 0 || maxBytes <= 0)
         return "";

package/dist/core/capture.d.ts

@@ -12,4 +12,8 @@ export type CaptureResult = {
     status: "skipped";
     type?: string;
 };
+/**
+ * Run one capture attempt and write exactly one diagnostics record for it,
+ * whether it is saved, skipped, or fails.
+ */
 export declare function captureConversation(workUnit: CaptureWorkUnit, provider: CaptureSummaryProvider): Promise<CaptureResult>;