modelstat 0.0.23 → 0.0.24

package/LICENSE

@@ -0,0 +1,87 @@
+Copyright (c) 2026 ModelState Inc
+
+Source-Available License
+
+1. Grant of Rights
+
+Subject to the terms of this License, you are granted a non-exclusive, worldwide,
+non-transferable, non-sublicensable license to:
+
+- View, read, and inspect the source code
+- Build, modify, and run the software
+- Use the software internally, including in production, solely to access or interact
+  with ModelState Inc’s hosted services
+
+2. Permitted Use
+
+You may use the software as a client, agent, or self-hosted component that connects
+to and depends on ModelState Inc’s cloud or hosted services.
+
+3. Restrictions
+
+You may NOT, without explicit prior written permission from ModelState Inc:
+
+- Use the software to provide a hosted, managed, or SaaS service to third parties
+- Use the software in any product or service that competes with ModelState Inc
+- Use the software to build or operate an alternative to ModelState Inc’s services
+- Redistribute, sublicense, sell, license, or commercially exploit the software
+- Offer the software (modified or unmodified) as part of a commercial offering
+- Make the software available to third parties as a service
+- Use the software for the benefit of third parties (including multi-tenant or shared environments)
+- Circumvent or attempt to circumvent the limitations of this License
+- Remove or alter any licensing, copyright, or attribution notices
+
+4. Definition of Competing Service
+
+“Competing Service” means any product or service that provides substantially similar
+functionality to ModelState Inc’s offerings, including but not limited to:
+
+- AI or LLM usage tracking, monitoring, or observability systems
+- Model analytics platforms or dashboards
+- Inference tracking, logging, or telemetry pipelines
+- Evaluation, benchmarking, or quality analysis systems for AI/ML models
+- Cost tracking, performance tracking, or optimization systems for model inference
+- Any system that collects, processes, analyzes, or visualizes usage or behavior of AI or machine learning models, including large language models (LLMs), when offered as a product or service
+
+5. Internal Use
+
+Use of the software is permitted only for your internal business or personal use,
+and not for the benefit of third parties.
+
+6. Network Use Restriction
+
+You may not use the software to expose APIs, endpoints, dashboards, or services to external
+users except as part of accessing ModelState Inc’s services.
+
+7. Data Extraction Restriction
+
+You may not use the software to replicate, extract, reverse engineer, or reconstruct
+ModelState Inc’s service behavior, APIs, data models, or system design for the purpose
+of building, improving, or operating a competing system or service.
+
+8. Ownership
+
+All rights, title, and interest in the software remain exclusively with ModelState Inc.
+The software is licensed, not sold.
+
+9. Contributions
+
+Unless explicitly agreed otherwise in writing, any contributions submitted to the software
+grant ModelState Inc a perpetual, irrevocable, worldwide, royalty-free license to use,
+modify, and distribute those contributions.
+
+10. Termination
+
+This License terminates automatically if you violate any of its terms.
+Upon termination, you must immediately cease all use and delete all copies of the software.
+
+11. Disclaimer of Warranty
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED.
+
+12. Limitation of Liability
+
+IN NO EVENT SHALL MODELSTATE INC BE LIABLE FOR ANY CLAIM, DAMAGES,
+OR OTHER LIABILITY ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE
+OR ITS USE OR OTHER DEALINGS IN THE SOFTWARE.

package/dist/cli.mjs

@@ -4339,6 +4339,18 @@ var init_schemas = __esm({
       tool_calls: external_exports.record(external_exports.string(), external_exports.number().int().nonnegative()).default({}),
       // Files touched, relative to git root. Never absolute — scrubbed by agent.
       files_touched: external_exports.array(external_exports.string().max(512)).max(256).default([]),
+      // Redacted excerpt of the conversation turn (user prompt or
+      // assistant response). The PARSER is responsible for:
+      //   1. Pulling a representative snippet from the turn (≤320 chars).
+      //   2. Running it through @modelstat/core/redact PLUS, when
+      //      available, the on-device Privacy Filter adapter.
+      //   3. Stripping code blocks and file-path noise.
+      // Optional — events without it fall back to metadata-only abstracts
+      // (the historical behaviour). The companion-core pipeline runs
+      // redact() over it again as defence-in-depth before building the
+      // summarize prompt; it never gets stored long-term server-side, only
+      // used to construct the summarize input.
+      content_excerpt: external_exports.string().max(320).optional(),
       // Reference to originating file for reparsing
       source_file: external_exports.string().max(1024).nullable(),
       source_byte_offset: external_exports.number().int().nonnegative().nullable(),
@@ -4354,7 +4366,7 @@ var init_schemas = __esm({
       secrets_found: external_exports.number().int().nonnegative().default(0),
       emails_redacted: external_exports.number().int().nonnegative().default(0),
       paths_redacted_absolute: external_exports.number().int().nonnegative().default(0)
-    });
+    }).catchall(external_exports.number().int().nonnegative());
     TaxonomyHintRooted = external_exports.object({
       root_key: external_exports.string().max(60),
       name: external_exports.string().max(120),
@@ -4651,6 +4663,28 @@ import { createHash } from "crypto";
 import { createReadStream } from "fs";
 import { stat } from "fs/promises";
 import { createInterface } from "readline";
+function extractExcerpt(content) {
+  if (!content) return void 0;
+  let text = "";
+  if (typeof content === "string") {
+    text = content;
+  } else if (Array.isArray(content)) {
+    const parts = [];
+    for (const block of content) {
+      if (block && block.type === "text" && typeof block.text === "string") {
+        parts.push(block.text);
+      }
+    }
+    text = parts.join(" ");
+  }
+  if (!text) return void 0;
+  text = text.replace(/```[\s\S]*?```/g, " ").replace(/`[^`]*`/g, " ");
+  text = text.replace(/\s+/g, " ").trim();
+  if (!text) return void 0;
+  const cleaned = redact(text).text;
+  const truncated = cleaned.slice(0, 320);
+  return truncated.length > 0 ? truncated : void 0;
+}
 async function parseClaudeCodeJsonl(ctx) {
   const events = [];
   let rawLines = 0;
@@ -4701,6 +4735,7 @@ async function parseClaudeCodeJsonl(ctx) {
         continue;
       }
       const slug = guessRepoSlugFromPath(cwd);
+      const excerpt = extractExcerpt(a.message?.content);
       events.push({
         source_event_id: sourceEventId(ctx.deviceId, ctx.sourceFile, offsetAtLineStart),
         ts: a.timestamp,
@@ -4729,6 +4764,7 @@ async function parseClaudeCodeJsonl(ctx) {
         duration_ms: null,
         tool_calls: {},
         files_touched: [],
+        ...excerpt ? { content_excerpt: excerpt } : {},
         source_file: ctx.sourceFile,
         source_byte_offset: offsetAtLineStart,
         // Files in ~/.claude/projects/ come from the Claude Code app
@@ -4743,6 +4779,7 @@ async function parseClaudeCodeJsonl(ctx) {
         skipped += 1;
         continue;
       }
+      const excerpt = extractExcerpt(u.message?.content);
       events.push({
         source_event_id: sourceEventId(ctx.deviceId, ctx.sourceFile, offsetAtLineStart),
         ts: u.timestamp,
@@ -4759,6 +4796,7 @@ async function parseClaudeCodeJsonl(ctx) {
         duration_ms: null,
         tool_calls: {},
         files_touched: [],
+        ...excerpt ? { content_excerpt: excerpt } : {},
         source_file: ctx.sourceFile,
         source_byte_offset: offsetAtLineStart,
         billing: "subscription"
@@ -44197,9 +44235,9 @@ var OLLAMA_CHAT_MODEL, OLLAMA_EMBED_MODEL, SUMMARISER_SYSTEM_PROMPT, SUMMARISER_
 var init_prompts = __esm({
   "../../packages/companion-core/src/pipeline/prompts.ts"() {
     "use strict";
-    OLLAMA_CHAT_MODEL = "qwen3:0.6b";
+    OLLAMA_CHAT_MODEL = "qwen3.5:0.8b";
     OLLAMA_EMBED_MODEL = "bge-small-en-v1.5";
-    SUMMARISER_SYSTEM_PROMPT = "You summarise an AI coding session in ONE sentence, \u2264 240 characters. Focus on what the human was trying to accomplish. No quotes, no PII, no code literals, no file paths. Reply with only the sentence.";
+    SUMMARISER_SYSTEM_PROMPT = "You summarise an AI coding session in ONE sentence, \u2264 240 characters. If the user message includes sampled conversation excerpts, base your summary on what the developer was actually working on (the substance \u2014 what was being built, debugged, refactored, or designed). If only metadata is given, paraphrase the metadata. Never quote the excerpts verbatim. No PII, no code literals, no file paths, no API keys. Reply with only the sentence.";
     SUMMARISER_MAX_TOKENS = 120;
     SUMMARISER_TEMPERATURE = 0.2;
     QWEN_CHARS_PER_TOKEN = 3.3;
@@ -44310,7 +44348,14 @@ async function summariseSlice(sessionId, slice, adapters2) {
     first.files_touched?.length ? `files touched: ${first.files_touched.slice(0, 5).join(", ")}` : null,
     Object.keys(first.tool_calls ?? {}).length ? `tool calls: ${Object.keys(first.tool_calls).slice(0, 5).join(", ")}` : null
   ].filter(Boolean).join("; ");
-  const prompt = `Session context: ${promptFacts || "generic coding session"}.
+  const excerpts = sampleAndRedactExcerpts(slice);
+  const excerptBlock = excerpts.length ? excerpts.map((e, i) => `  [turn ${i + 1}] "${e.replace(/\s+/g, " ").trim()}"`).join("\n") : "";
+  const prompt = excerptBlock ? `Session context: ${promptFacts || "generic coding session"}.
+
+Sampled excerpts from the conversation (already redacted of PII and secrets):
+${excerptBlock}
+
+Write ONE sentence (\u2264240 chars) describing what the human was working on. Focus on the substance \u2014 what was being built, debugged, or designed. No quotes, no PII, no code literals, no file paths.` : `Session context: ${promptFacts || "generic coding session"}.
 Write one sentence describing what the human was doing.`;
   let rawAbstract;
   try {
@@ -44318,7 +44363,20 @@ Write one sentence describing what the human was doing.`;
   } catch {
     rawAbstract = promptFacts || `${first.tool} session with ${slice.length} turns`;
   }
-  const redacted = redact(rawAbstract);
+  const regexPass = redact(rawAbstract);
+  let abstractText = regexPass.text;
+  const counts = { ...regexPass.counts };
+  if (adapters2.redact) {
+    try {
+      const modelPass = await adapters2.redact(regexPass.text);
+      abstractText = modelPass.text;
+      for (const [k, v] of Object.entries(modelPass.counts)) {
+        if (k.startsWith("pf_")) counts[k] = v;
+      }
+    } catch {
+    }
+  }
+  const redacted = { text: abstractText, counts };
   const tags = [
     { root_key: "tools", name: first.tool, confidence: 1 },
     { root_key: "providers", name: first.provider, confidence: 1 }
@@ -44359,11 +44417,39 @@ Write one sentence describing what the human was doing.`;
     abstract: redacted.text.slice(0, ABSTRACT_MAX_CHARS),
     tokens,
     tags,
+    // counts is `Record<string, number>` after the optional model
+    // merge; the schema's RedactionReport requires the three regex
+    // counters (always populated from regexPass.counts) plus a
+    // number-valued catchall for pf_*.
     redaction: redacted.counts,
     source_event_ids: sourceEventIds,
     abstract_embedding: segmentEmbedding && segmentEmbedding.length === 384 ? segmentEmbedding : void 0
   };
 }
+function sampleAndRedactExcerpts(slice) {
+  const withContent = [];
+  for (let i = 0; i < slice.length; i++) {
+    const c = slice[i]?.content_excerpt;
+    if (c && c.trim().length > 0) withContent.push({ idx: i, text: c });
+  }
+  if (withContent.length === 0) return [];
+  const picks = [0];
+  if (withContent.length > 1) picks.push(withContent.length - 1);
+  for (const frac of [0.25, 0.5, 0.75]) {
+    const idx = Math.floor(withContent.length * frac);
+    if (!picks.includes(idx)) picks.push(idx);
+    if (picks.length >= 5) break;
+  }
+  picks.sort((a, b) => a - b);
+  const out = [];
+  for (const i of picks) {
+    const raw = withContent[i]?.text;
+    if (!raw) continue;
+    const redacted = redact(raw).text;
+    out.push(redacted.slice(0, 200));
+  }
+  return out;
+}
 function turnSurface(e) {
   const parts = [e.kind, e.tool];
   if (e.model) parts.push(e.model);