modelstat 0.0.52 → 0.1.0

package/dist/cli.mjs

@@ -63,17 +63,77 @@ import { execFile } from "child_process";
 import { promisify } from "util";
 import { dirname, resolve } from "path";
 import { existsSync } from "fs";
+function findRepoRoot(startCwd) {
+  let cwd = resolve(startCwd);
+  for (let i = 0; i < 10; i++) {
+    if (existsSync(`${cwd}/.git`)) return cwd;
+    const parent = dirname(cwd);
+    if (parent === cwd) return null;
+    cwd = parent;
+  }
+  return null;
+}
+function parseRemote(url) {
+  const ssh = /^(?:git@)?([^:]+):([^/]+)\/([^.]+?)(?:\.git)?$/.exec(url);
+  if (ssh) return { host: ssh[1] ?? null, slug: `${ssh[2]}/${ssh[3]}` };
+  try {
+    const u = new URL(url);
+    const m = /^\/([^/]+)\/([^/]+?)(?:\.git)?\/?$/.exec(u.pathname);
+    if (!m) return { host: u.hostname, slug: null };
+    return { host: u.hostname, slug: `${m[1]}/${m[2]}` };
+  } catch {
+    return { host: null, slug: null };
+  }
+}
+async function resolveGitContext(cwd) {
+  if (!cwd) return null;
+  if (cache.has(cwd)) return cache.get(cwd) ?? null;
+  const root = findRepoRoot(cwd);
+  if (!root) {
+    const empty = {
+      remote_url: null,
+      remote_host: null,
+      remote_slug: null,
+      branch: null,
+      commit_sha: null
+    };
+    cache.set(cwd, empty);
+    return empty;
+  }
+  const ran = async (args) => {
+    try {
+      const { stdout } = await pexec("git", args, { cwd: root, timeout: 2e3 });
+      return stdout.trim() || null;
+    } catch {
+      return null;
+    }
+  };
+  const remoteUrl = await ran(["config", "--get", "remote.origin.url"]);
+  const branch = await ran(["rev-parse", "--abbrev-ref", "HEAD"]);
+  const sha = await ran(["rev-parse", "HEAD"]);
+  const parsed = remoteUrl ? parseRemote(remoteUrl) : { host: null, slug: null };
+  const ctx = {
+    remote_url: remoteUrl,
+    remote_host: parsed.host,
+    remote_slug: parsed.slug,
+    branch,
+    commit_sha: sha
+  };
+  cache.set(cwd, ctx);
+  return ctx;
+}
 function guessRepoSlugFromPath(cwd) {
   if (!cwd) return null;
   const m = /\/(?:www|src|code|repos|projects)\/([^/]+)\/([^/]+)/i.exec(cwd);
   if (m) return `${m[1]}/${m[2]}`;
   return null;
 }
-var pexec;
+var pexec, cache;
 var init_git = __esm({
   "../../packages/parsers/src/git.ts"() {
     "use strict";
     pexec = promisify(execFile);
+    cache = /* @__PURE__ */ new Map();
   }
 });
 
@@ -4603,6 +4663,327 @@ var init_redact = __esm({
   }
 });
 
+// ../../packages/core/src/session-metadata.ts
+function emptyDetectedRefs() {
+  return { repos: [], pull_requests: [], commits: [], issues: [] };
+}
+function repoFrom(host, slug, source) {
+  return { host, slug, branches: [], source };
+}
+function detectReferences(text, source = "content") {
+  const out = emptyDetectedRefs();
+  if (!text) return out;
+  for (const m of text.matchAll(GITHUB_PR)) {
+    const slug = `${m[1]}/${m[2]}`;
+    out.pull_requests.push({
+      host: "github.com",
+      slug,
+      number: Number(m[3]),
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    out.repos.push(repoFrom("github.com", slug, source));
+  }
+  for (const m of text.matchAll(GITLAB_MR)) {
+    out.pull_requests.push({
+      host: "gitlab.com",
+      slug: m[1] ?? null,
+      number: Number(m[2]),
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    if (m[1]) out.repos.push(repoFrom("gitlab.com", m[1], source));
+  }
+  for (const m of text.matchAll(BITBUCKET_PR)) {
+    const slug = `${m[1]}/${m[2]}`;
+    out.pull_requests.push({
+      host: "bitbucket.org",
+      slug,
+      number: Number(m[3]),
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    out.repos.push(repoFrom("bitbucket.org", slug, source));
+  }
+  for (const m of text.matchAll(GITHUB_ISSUE)) {
+    const slug = `${m[1]}/${m[2]}`;
+    out.issues.push({
+      provider: "github",
+      key: m[3] ?? "",
+      slug,
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    out.repos.push(repoFrom("github.com", slug, source));
+  }
+  for (const m of text.matchAll(GITLAB_ISSUE)) {
+    out.issues.push({
+      provider: "gitlab",
+      key: m[2] ?? "",
+      slug: m[1] ?? null,
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    if (m[1]) out.repos.push(repoFrom("gitlab.com", m[1], source));
+  }
+  for (const m of text.matchAll(LINEAR_ISSUE)) {
+    out.issues.push({
+      provider: "linear",
+      key: m[1] ?? "",
+      slug: null,
+      url: m[0],
+      source,
+      confidence: 0.9
+    });
+  }
+  for (const m of text.matchAll(JIRA_ISSUE)) {
+    out.issues.push({
+      provider: "jira",
+      key: m[1] ?? "",
+      slug: null,
+      url: m[0],
+      source,
+      confidence: 0.9
+    });
+  }
+  for (const m of text.matchAll(GITHUB_COMMIT)) {
+    const slug = `${m[1]}/${m[2]}`;
+    out.commits.push({ sha: m[3] ?? "", slug, url: m[0], source, confidence: 0.95 });
+    out.repos.push(repoFrom("github.com", slug, source));
+  }
+  for (const m of text.matchAll(GITLAB_COMMIT)) {
+    out.commits.push({
+      sha: m[2] ?? "",
+      slug: m[1] ?? null,
+      url: m[0],
+      source,
+      confidence: 0.95
+    });
+    if (m[1]) out.repos.push(repoFrom("gitlab.com", m[1], source));
+  }
+  for (const m of text.matchAll(SLUG_HASH)) {
+    const slug = m[1] ?? "";
+    out.issues.push({
+      provider: "github",
+      key: m[2] ?? "",
+      slug,
+      url: null,
+      source,
+      confidence: 0.55
+    });
+  }
+  if (source === "model") {
+    for (const m of text.matchAll(BARE_TICKET)) {
+      out.issues.push({
+        provider: "other",
+        key: m[1] ?? "",
+        slug: null,
+        url: null,
+        source,
+        confidence: 0.4
+      });
+    }
+  }
+  return out;
+}
+function detectBranchTickets(branch) {
+  if (!branch) return [];
+  const out = [];
+  for (const m of branch.matchAll(BARE_TICKET)) {
+    out.push({
+      provider: "other",
+      key: m[1] ?? "",
+      slug: null,
+      url: null,
+      source: "git",
+      confidence: 0.7
+    });
+  }
+  return out;
+}
+function stronger(a, b) {
+  return SOURCE_RANK[a] >= SOURCE_RANK[b] ? a : b;
+}
+function dedupe(items, keyOf, merge) {
+  const byKey = /* @__PURE__ */ new Map();
+  for (const item of items) {
+    const key = keyOf(item);
+    const existing = byKey.get(key);
+    byKey.set(key, existing ? merge(existing, item) : item);
+  }
+  return [...byKey.values()];
+}
+function dedupeSessionMetadata(parts) {
+  const all = emptyDetectedRefs();
+  for (const p of parts) {
+    all.repos.push(...p.repos);
+    all.pull_requests.push(...p.pull_requests);
+    all.commits.push(...p.commits);
+    all.issues.push(...p.issues);
+  }
+  const repos = dedupe(
+    all.repos,
+    (r) => r.slug.toLowerCase(),
+    (a, b) => ({
+      host: a.host ?? b.host,
+      slug: a.slug,
+      branches: [.../* @__PURE__ */ new Set([...a.branches, ...b.branches])].slice(0, 50),
+      source: stronger(a.source, b.source)
+    })
+  );
+  const score = (r) => SOURCE_RANK[r.source] * 2 + (r.confidence ?? 0);
+  const pull_requests = dedupe(
+    all.pull_requests,
+    (p) => `${(p.slug ?? "").toLowerCase()}#${p.number}`,
+    (a, b) => {
+      const [win, lose] = score(a) >= score(b) ? [a, b] : [b, a];
+      return {
+        host: win.host ?? lose.host,
+        slug: win.slug ?? lose.slug,
+        number: win.number,
+        url: win.url ?? lose.url,
+        source: win.source,
+        confidence: Math.max(win.confidence, lose.confidence)
+      };
+    }
+  );
+  const commits = dedupe(
+    all.commits,
+    // Key on slug + sha: two different commits in two repos can share a short
+    // 7-hex prefix, and must not collapse into one.
+    (c) => `${(c.slug ?? "").toLowerCase()}@${c.sha.toLowerCase()}`,
+    (a, b) => {
+      const [win, lose] = score(a) >= score(b) ? [a, b] : [b, a];
+      return {
+        sha: win.sha,
+        slug: win.slug ?? lose.slug,
+        url: win.url ?? lose.url,
+        source: win.source,
+        confidence: Math.max(win.confidence, lose.confidence)
+      };
+    }
+  );
+  const issues = dedupe(
+    all.issues,
+    (i) => `${i.provider}:${(i.slug ?? "").toLowerCase()}#${i.key.toLowerCase()}`,
+    (a, b) => {
+      const [win, lose] = score(a) >= score(b) ? [a, b] : [b, a];
+      return {
+        provider: win.provider,
+        key: win.key,
+        slug: win.slug ?? lose.slug,
+        url: win.url ?? lose.url,
+        source: win.source,
+        confidence: Math.max(win.confidence, lose.confidence)
+      };
+    }
+  );
+  const prKeys = new Set(pull_requests.map((p) => `${(p.slug ?? "").toLowerCase()}#${p.number}`));
+  const reconciledIssues = issues.filter(
+    (i) => !(i.provider === "github" && prKeys.has(`${(i.slug ?? "").toLowerCase()}#${i.key}`))
+  );
+  const keepValid = (schema, items) => items.filter((item) => schema.safeParse(item).success);
+  return {
+    repos: keepValid(RepoRef, repos).slice(0, 50),
+    pull_requests: keepValid(PullRequestRef, pull_requests).slice(0, 100),
+    commits: keepValid(CommitRef, commits).slice(0, 200),
+    issues: keepValid(IssueRef, reconciledIssues).slice(0, 100)
+  };
+}
+function isEmptySessionMetadata(m) {
+  return m.repos.length === 0 && m.pull_requests.length === 0 && m.commits.length === 0 && m.issues.length === 0;
+}
+function detectEventReferences(text) {
+  if (!text) return null;
+  const m = dedupeSessionMetadata([detectReferences(text, "content")]);
+  if (isEmptySessionMetadata(m)) return null;
+  return {
+    repos: m.repos.slice(0, 24),
+    pull_requests: m.pull_requests.slice(0, 24),
+    commits: m.commits.slice(0, 24),
+    issues: m.issues.slice(0, 24)
+  };
+}
+var REF_SOURCES, RefSource, SOURCE_RANK, RepoRef, PullRequestRef, CommitRef, ISSUE_PROVIDERS, IssueRef, SessionMetadata, EventReferences, GITHUB_PR, GITLAB_MR, BITBUCKET_PR, GITHUB_ISSUE, GITLAB_ISSUE, GITHUB_COMMIT, GITLAB_COMMIT, LINEAR_ISSUE, JIRA_ISSUE, SLUG_HASH, BARE_TICKET;
+var init_session_metadata = __esm({
+  "../../packages/core/src/session-metadata.ts"() {
+    "use strict";
+    init_zod();
+    REF_SOURCES = ["git", "tool", "content", "model"];
+    RefSource = external_exports.enum(REF_SOURCES);
+    SOURCE_RANK = { git: 3, tool: 2, content: 1, model: 0 };
+    RepoRef = external_exports.object({
+      /** `github.com`, `gitlab.com`, … — null when only the slug is known. */
+      host: external_exports.string().max(80).nullable().default(null),
+      /** `org/repo`. The stable identity of the repo. */
+      slug: external_exports.string().max(200),
+      /** Every branch observed for this repo this session. */
+      branches: external_exports.array(external_exports.string().max(200)).max(50).default([]),
+      source: RefSource.default("content")
+    });
+    PullRequestRef = external_exports.object({
+      host: external_exports.string().max(80).nullable().default(null),
+      slug: external_exports.string().max(200).nullable().default(null),
+      number: external_exports.number().int().positive(),
+      url: external_exports.string().max(400).nullable().default(null),
+      source: RefSource.default("content"),
+      confidence: external_exports.number().min(0).max(1).default(0.9)
+    });
+    CommitRef = external_exports.object({
+      /** 7–40 char hex SHA. */
+      sha: external_exports.string().max(64),
+      slug: external_exports.string().max(200).nullable().default(null),
+      url: external_exports.string().max(400).nullable().default(null),
+      source: RefSource.default("content"),
+      confidence: external_exports.number().min(0).max(1).default(0.8)
+    });
+    ISSUE_PROVIDERS = [
+      "github",
+      "gitlab",
+      "bitbucket",
+      "linear",
+      "jira",
+      "other"
+    ];
+    IssueRef = external_exports.object({
+      provider: external_exports.enum(ISSUE_PROVIDERS).default("other"),
+      key: external_exports.string().max(80),
+      slug: external_exports.string().max(200).nullable().default(null),
+      url: external_exports.string().max(400).nullable().default(null),
+      source: RefSource.default("content"),
+      confidence: external_exports.number().min(0).max(1).default(0.8)
+    });
+    SessionMetadata = external_exports.object({
+      repos: external_exports.array(RepoRef).max(50).default([]),
+      pull_requests: external_exports.array(PullRequestRef).max(100).default([]),
+      commits: external_exports.array(CommitRef).max(200).default([]),
+      issues: external_exports.array(IssueRef).max(100).default([])
+    });
+    EventReferences = external_exports.object({
+      repos: external_exports.array(RepoRef).max(24).default([]),
+      pull_requests: external_exports.array(PullRequestRef).max(24).default([]),
+      commits: external_exports.array(CommitRef).max(24).default([]),
+      issues: external_exports.array(IssueRef).max(24).default([])
+    });
+    GITHUB_PR = /https?:\/\/github\.com\/([\w.-]+)\/([\w.-]+)\/pull\/(\d+)/gi;
+    GITLAB_MR = /https?:\/\/gitlab\.com\/([\w./-]+?)\/-\/merge_requests\/(\d+)/gi;
+    BITBUCKET_PR = /https?:\/\/bitbucket\.org\/([\w.-]+)\/([\w.-]+)\/pull-requests\/(\d+)/gi;
+    GITHUB_ISSUE = /https?:\/\/github\.com\/([\w.-]+)\/([\w.-]+)\/issues\/(\d+)/gi;
+    GITLAB_ISSUE = /https?:\/\/gitlab\.com\/([\w./-]+?)\/-\/issues\/(\d+)/gi;
+    GITHUB_COMMIT = /https?:\/\/github\.com\/([\w.-]+)\/([\w.-]+)\/commit\/([0-9a-f]{7,40})/gi;
+    GITLAB_COMMIT = /https?:\/\/gitlab\.com\/([\w./-]+?)\/-\/commit\/([0-9a-f]{7,40})/gi;
+    LINEAR_ISSUE = /https?:\/\/linear\.app\/[\w.-]+\/issue\/([A-Z][A-Z0-9]*-\d+)/gi;
+    JIRA_ISSUE = /https?:\/\/[\w.-]+\/browse\/([A-Z][A-Z0-9]+-\d+)/gi;
+    SLUG_HASH = /(?:^|[\s([{<])([\w.-]+\/[\w.-]+)#(\d+)\b/g;
+    BARE_TICKET = /\b([A-Z][A-Z0-9]{1,9}-\d{1,6})\b/g;
+  }
+});
+
 // ../../packages/core/src/schemas.ts
 var TokenUsage, GitContext, RawEvent, RedactionReport, TaxonomyHintRooted, Segment, ToolAction, ToolCallWire, IngestBatch, HeartbeatPayload, DeviceEnrollment, DeviceSelfRegister, DeviceClaimRequest, ProcessingMetadata, RedactionPolicy, DetectedInstallation, DetectedIdentity, DiscoveryReport, ClassificationConfidenceEnum;
 var init_schemas = __esm({
@@ -4610,6 +4991,7 @@ var init_schemas = __esm({
     "use strict";
     init_zod();
     init_enums();
+    init_session_metadata();
     TokenUsage = external_exports.object({
       input: external_exports.number().int().nonnegative().default(0),
       output: external_exports.number().int().nonnegative().default(0),
@@ -4662,6 +5044,12 @@ var init_schemas = __esm({
       // 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(),
+      // Public code references (PRs, issues, commits, repos) detected on-device
+      // from this turn's FULL text — the high-recall feed the server rolls up into
+      // SessionMetadata. Only public reference shapes (forge URLs, slugs, numbers,
+      // ticket keys) ride here, never raw text — so it is derived pre-redaction
+      // safely (same class as git.remote_slug). Optional + additive.
+      references: EventReferences.optional(),
       // Reference to originating file for reparsing
       source_file: external_exports.string().max(1024).nullable(),
       source_byte_offset: external_exports.number().int().nonnegative().nullable(),
@@ -4810,7 +5198,14 @@ var init_schemas = __esm({
        * session view on every upload, so the latest batch always carries the
        * freshest title. Absent for runtimes without a titler (older agents,
        * no-op browser summariser). */
-      session_titles: external_exports.record(external_exports.string(), external_exports.string().max(120)).optional()
+      session_titles: external_exports.record(external_exports.string(), external_exports.string().max(120)).optional(),
+      /** Optional per-session deterministic metadata — session_id →
+       * {@link SessionMetadata}: the repos, pull requests, commits, and issues the
+       * session touched, detected on-device across git context, tool calls,
+       * redacted content, and the local model (so it works for any provider).
+       * Additive — old companions omit it, old servers ignore it (the wire has no
+       * `deny_unknown_fields`). The join layer between AI spend and shipped work. */
+      session_metadata: external_exports.record(external_exports.string(), SessionMetadata).optional()
     });
     HeartbeatPayload = external_exports.object({
       device_id: external_exports.string(),
@@ -4918,6 +5313,7 @@ var init_src = __esm({
     init_redact();
     init_redact_floor();
     init_schemas();
+    init_session_metadata();
   }
 });
 
@@ -5133,6 +5529,22 @@ function extractExcerpt(content) {
   const truncated = cleaned.slice(0, 320);
   return truncated.length > 0 ? truncated : void 0;
 }
+function collectRefText(content) {
+  if (!content) return "";
+  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(" ");
+  }
+  return text.length > 64e3 ? text.slice(0, 64e3) : text;
+}
 function buildToolCallDraft(opts) {
   const { server, name } = splitObservedToolName(opts.observedName);
   const hashes = hashArgs(opts.input);
@@ -5270,6 +5682,7 @@ async function parseClaudeCodeJsonl(ctx) {
       }
       const slug = guessRepoSlugFromPath(cwd);
       const excerpt = extractExcerpt(a.message?.content);
+      const refs = detectEventReferences(collectRefText(a.message?.content));
       const aggregate = {};
       const blocks = Array.isArray(a.message?.content) ? a.message.content : [];
       let callIndex = 0;
@@ -5309,9 +5722,14 @@ async function parseClaudeCodeJsonl(ctx) {
         turn_index: null,
         parent_event_id: a.parentUuid ?? null,
         cwd,
-        git: slug ? {
+        // Emit git context whenever there is ANY signal — a repo slug OR the
+        // session's branch. Gating on `slug` alone dropped the branch (and all
+        // git context) for cwds the path heuristic doesn't match (e.g.
+        // ~/Documents/<repo>), starving the session-metadata join. The branch
+        // is the historical one Claude Code recorded for the turn.
+        git: slug || gitBranch ? {
           remote_url: null,
-          remote_host: slug.includes("/") ? "github.com" : null,
+          remote_host: slug?.includes("/") ? "github.com" : null,
           remote_slug: slug,
           branch: gitBranch,
           commit_sha: null
@@ -5327,6 +5745,7 @@ async function parseClaudeCodeJsonl(ctx) {
         tool_calls: aggregate,
         files_touched: [],
         ...excerpt ? { content_excerpt: excerpt } : {},
+        ...refs ? { references: refs } : {},
         source_file: ctx.sourceFile,
         source_byte_offset: offsetAtLineStart,
         // Files in ~/.claude/projects/ come from the Claude Code app
@@ -5361,6 +5780,7 @@ async function parseClaudeCodeJsonl(ctx) {
         continue;
       }
       const excerpt = extractExcerpt(u.message?.content);
+      const refs = detectEventReferences(collectRefText(u.message?.content));
       await emit({
         source_event_id: eventId,
         ts: u.timestamp,
@@ -5378,6 +5798,7 @@ async function parseClaudeCodeJsonl(ctx) {
         tool_calls: {},
         files_touched: [],
         ...excerpt ? { content_excerpt: excerpt } : {},
+        ...refs ? { references: refs } : {},
         source_file: ctx.sourceFile,
         source_byte_offset: offsetAtLineStart,
         billing: "subscription"
@@ -29466,8 +29887,8 @@ var require_cache3 = __commonJS({
        * @returns {requestResponseList}
        */
       #batchCacheOperations(operations) {
-        const cache = this.#relevantRequestResponseList;
-        const backupCache = [...cache];
+        const cache2 = this.#relevantRequestResponseList;
+        const backupCache = [...cache2];
         const addedItems = [];
         const resultList = [];
         try {
@@ -29494,9 +29915,9 @@ var require_cache3 = __commonJS({
                 return [];
               }
               for (const requestResponse of requestResponses) {
-                const idx = cache.indexOf(requestResponse);
+                const idx = cache2.indexOf(requestResponse);
                 assert2(idx !== -1);
-                cache.splice(idx, 1);
+                cache2.splice(idx, 1);
               }
             } else if (operation.type === "put") {
               if (operation.response == null) {
@@ -29526,11 +29947,11 @@ var require_cache3 = __commonJS({
               }
               requestResponses = this.#queryCache(operation.request);
               for (const requestResponse of requestResponses) {
-                const idx = cache.indexOf(requestResponse);
+                const idx = cache2.indexOf(requestResponse);
                 assert2(idx !== -1);
-                cache.splice(idx, 1);
+                cache2.splice(idx, 1);
               }
-              cache.push([operation.request, operation.response]);
+              cache2.push([operation.request, operation.response]);
               addedItems.push([operation.request, operation.response]);
             }
             resultList.push([operation.request, operation.response]);
@@ -29707,13 +30128,13 @@ var require_cachestorage = __commonJS({
         if (options.cacheName != null) {
           if (this.#caches.has(options.cacheName)) {
             const cacheList = this.#caches.get(options.cacheName);
-            const cache = new Cache(kConstruct, cacheList);
-            return await cache.match(request3, options);
+            const cache2 = new Cache(kConstruct, cacheList);
+            return await cache2.match(request3, options);
           }
         } else {
           for (const cacheList of this.#caches.values()) {
-            const cache = new Cache(kConstruct, cacheList);
-            const response = await cache.match(request3, options);
+            const cache2 = new Cache(kConstruct, cacheList);
+            const response = await cache2.match(request3, options);
             if (response !== void 0) {
               return response;
             }
@@ -29743,12 +30164,12 @@ var require_cachestorage = __commonJS({
         webidl.argumentLengthCheck(arguments, 1, prefix);
         cacheName = webidl.converters.DOMString(cacheName, prefix, "cacheName");
         if (this.#caches.has(cacheName)) {
-          const cache2 = this.#caches.get(cacheName);
-          return new Cache(kConstruct, cache2);
+          const cache3 = this.#caches.get(cacheName);
+          return new Cache(kConstruct, cache3);
         }
-        const cache = [];
-        this.#caches.set(cacheName, cache);
-        return new Cache(kConstruct, cache);
+        const cache2 = [];
+        this.#caches.set(cacheName, cache2);
+        return new Cache(kConstruct, cache2);
       }
       /**
        * @see https://w3c.github.io/ServiceWorker/#cache-storage-delete
@@ -43156,7 +43577,7 @@ var require_range = __commonJS({
       parseRange(range) {
         const memoOpts = (this.options.includePrerelease && FLAG_INCLUDE_PRERELEASE) | (this.options.loose && FLAG_LOOSE);
         const memoKey = memoOpts + ":" + range;
-        const cached2 = cache.get(memoKey);
+        const cached2 = cache2.get(memoKey);
         if (cached2) {
           return cached2;
         }
@@ -43190,7 +43611,7 @@ var require_range = __commonJS({
           rangeMap.delete("");
         }
         const result2 = [...rangeMap.values()];
-        cache.set(memoKey, result2);
+        cache2.set(memoKey, result2);
         return result2;
       }
       intersects(range, options) {
@@ -43229,7 +43650,7 @@ var require_range = __commonJS({
     };
     module.exports = Range;
     var LRU = require_lrucache();
-    var cache = new LRU();
+    var cache2 = new LRU();
     var parseOptions = require_parse_options();
     var Comparator = require_comparator();
     var debug = require_debug();
@@ -45301,6 +45722,38 @@ var init_prompts = __esm({
   }
 });
 
+// ../../packages/companion-core/src/pipeline/script-summary.ts
+function buildScriptSummaryUserPrompt(input) {
+  const body = input.content.slice(0, SCRIPT_SUMMARY_INPUT_MAX_CHARS);
+  return [
+    `Script reference: ${input.ref}`,
+    "Contents:",
+    "```",
+    body,
+    "```",
+    "",
+    "One sentence (\u2264200 chars): what does running this script do?"
+  ].join("\n");
+}
+var SCRIPT_SUMMARY_OUTPUT_MAX_CHARS, SCRIPT_SUMMARY_TEMPERATURE, SCRIPT_SUMMARY_MAX_TOKENS, SCRIPT_SUMMARY_INPUT_MAX_CHARS, SCRIPT_SUMMARY_SYSTEM_PROMPT;
+var init_script_summary = __esm({
+  "../../packages/companion-core/src/pipeline/script-summary.ts"() {
+    "use strict";
+    SCRIPT_SUMMARY_OUTPUT_MAX_CHARS = 200;
+    SCRIPT_SUMMARY_TEMPERATURE = 0.2;
+    SCRIPT_SUMMARY_MAX_TOKENS = 120;
+    SCRIPT_SUMMARY_INPUT_MAX_CHARS = 6e3;
+    SCRIPT_SUMMARY_SYSTEM_PROMPT = `You summarise what a script file does, for an engineer scanning a dashboard.
+
+Rules:
+- Output ONE plain sentence (at most 200 characters) stating what the script DOES when it runs.
+- Be concrete and factual: the actions it performs and the systems it touches.
+- No preamble ("This script\u2026"), no markdown, no quotes, no line breaks.
+- Do not invent behaviour that is not in the file. If the file is trivial or unreadable, say so briefly.
+- Never include secrets, tokens, passwords, or personal data, even if they appear in the file.`;
+  }
+});
+
 // ../../packages/companion-core/src/pipeline/title.ts
 function buildTitleUserPrompt(input) {
   const lines = input.abstracts.map(
@@ -45384,35 +45837,113 @@ var init_title = __esm({
   }
 });
 
-// ../../packages/companion-core/src/pipeline/script-summary.ts
-function buildScriptSummaryUserPrompt(input) {
-  const body = input.content.slice(0, SCRIPT_SUMMARY_INPUT_MAX_CHARS);
-  return [
-    `Script reference: ${input.ref}`,
-    "Contents:",
-    "```",
-    body,
-    "```",
-    "",
-    "One sentence (\u2264200 chars): what does running this script do?"
-  ].join("\n");
+// ../../packages/companion-core/src/pipeline/session-metadata.ts
+function buildLinkExtractUserPrompt(abstracts) {
+  const lines = abstracts.map((a, i) => `  [part ${i + 1}] ${a.replace(/\s+/g, " ").trim().slice(0, 240)}`).join("\n");
+  return `Summaries of the session's parts:
+${lines}
+
+List the references, one per line, or "none".`;
 }
-var SCRIPT_SUMMARY_OUTPUT_MAX_CHARS, SCRIPT_SUMMARY_TEMPERATURE, SCRIPT_SUMMARY_MAX_TOKENS, SCRIPT_SUMMARY_INPUT_MAX_CHARS, SCRIPT_SUMMARY_SYSTEM_PROMPT;
-var init_script_summary = __esm({
-  "../../packages/companion-core/src/pipeline/script-summary.ts"() {
+function groupBy(items, keyOf) {
+  const out = /* @__PURE__ */ new Map();
+  for (const item of items) {
+    const key = keyOf(item);
+    const arr = out.get(key) ?? [];
+    arr.push(item);
+    out.set(key, arr);
+  }
+  return out;
+}
+async function buildSessionMetadata(segments, events, opts = {}) {
+  const eventsBySession = groupBy(events, (e) => e.session_id);
+  const segsBySession = groupBy(segments, (s) => s.session_id);
+  const sessionIds = /* @__PURE__ */ new Set([...eventsBySession.keys(), ...segsBySession.keys()]);
+  const out = {};
+  for (const sessionId of sessionIds) {
+    try {
+      const evs = eventsBySession.get(sessionId) ?? [];
+      const segs = segsBySession.get(sessionId) ?? [];
+      const parts = [];
+      const cwds = /* @__PURE__ */ new Set();
+      for (const e of evs) {
+        if (e.cwd) cwds.add(e.cwd);
+        if (!e.git) continue;
+        const refs = emptyDetectedRefs();
+        if (e.git.remote_slug) {
+          refs.repos.push({
+            host: e.git.remote_host ?? null,
+            slug: e.git.remote_slug,
+            branches: e.git.branch ? [e.git.branch] : [],
+            source: "git"
+          });
+        }
+        if (e.git.branch) refs.issues.push(...detectBranchTickets(e.git.branch));
+        if (e.git.commit_sha) {
+          refs.commits.push({
+            sha: e.git.commit_sha,
+            slug: e.git.remote_slug ?? null,
+            url: null,
+            source: "git",
+            confidence: 0.6
+          });
+        }
+        parts.push(refs);
+      }
+      if (opts.resolveGit) {
+        for (const cwd of cwds) {
+          let g = null;
+          try {
+            g = await opts.resolveGit(cwd);
+          } catch {
+            g = null;
+          }
+          if (!g?.remote_slug) continue;
+          const refs = emptyDetectedRefs();
+          refs.repos.push({
+            host: g.remote_host ?? null,
+            slug: g.remote_slug,
+            branches: g.branch ? [g.branch] : [],
+            source: "git"
+          });
+          if (g.branch) refs.issues.push(...detectBranchTickets(g.branch));
+          parts.push(refs);
+        }
+      }
+      for (const e of evs) {
+        if (e.references) parts.push(e.references);
+      }
+      for (const e of evs) {
+        if (e.content_excerpt) parts.push(detectReferences(e.content_excerpt, "content"));
+      }
+      const abstracts = [...segs].sort((a, b) => a.started_at.localeCompare(b.started_at)).map((s) => stripCognitionSuffix(s.abstract)).filter((a) => a.length > 0);
+      for (const a of abstracts) parts.push(detectReferences(a, "content"));
+      if (opts.extractLinks && abstracts.length > 0) {
+        try {
+          const reply = await opts.extractLinks({
+            abstracts: sampleAbstracts(abstracts, LINK_EXTRACT_MAX_ABSTRACTS)
+          });
+          if (reply) parts.push(detectReferences(reply, "model"));
+        } catch {
+        }
+      }
+      const meta = dedupeSessionMetadata(parts);
+      if (!isEmptySessionMetadata(meta)) out[sessionId] = meta;
+    } catch {
+    }
+  }
+  return out;
+}
+var LINK_EXTRACT_SYSTEM_PROMPT, LINK_EXTRACT_MAX_TOKENS, LINK_EXTRACT_TEMPERATURE, LINK_EXTRACT_MAX_ABSTRACTS;
+var init_session_metadata2 = __esm({
+  "../../packages/companion-core/src/pipeline/session-metadata.ts"() {
     "use strict";
-    SCRIPT_SUMMARY_OUTPUT_MAX_CHARS = 200;
-    SCRIPT_SUMMARY_TEMPERATURE = 0.2;
-    SCRIPT_SUMMARY_MAX_TOKENS = 120;
-    SCRIPT_SUMMARY_INPUT_MAX_CHARS = 6e3;
-    SCRIPT_SUMMARY_SYSTEM_PROMPT = `You summarise what a script file does, for an engineer scanning a dashboard.
-
-Rules:
-- Output ONE plain sentence (at most 200 characters) stating what the script DOES when it runs.
-- Be concrete and factual: the actions it performs and the systems it touches.
-- No preamble ("This script\u2026"), no markdown, no quotes, no line breaks.
-- Do not invent behaviour that is not in the file. If the file is trivial or unreadable, say so briefly.
-- Never include secrets, tokens, passwords, or personal data, even if they appear in the file.`;
+    init_session_metadata();
+    init_title();
+    LINK_EXTRACT_SYSTEM_PROMPT = "You extract code-collaboration references from one-sentence summaries of an AI-coding session. Report ONLY references that explicitly appear in the text: pull/merge request URLs, issue URLs, commit URLs, the `org/repo#123` shorthand, and ticket keys like ENG-123 or PROJ-4567. Output one reference per line, verbatim, nothing else \u2014 no prose, no numbering, no markdown. If the summaries contain no such reference, reply with exactly `none`. Never invent a reference, a repo, a number, or a URL that is not present in the text.";
+    LINK_EXTRACT_MAX_TOKENS = 120;
+    LINK_EXTRACT_TEMPERATURE = 0.1;
+    LINK_EXTRACT_MAX_ABSTRACTS = 12;
   }
 });
 
@@ -45727,8 +46258,9 @@ var init_pipeline = __esm({
     init_redact();
     init_cognition();
     init_prompts();
-    init_title();
     init_script_summary();
+    init_session_metadata2();
+    init_title();
     SEGMENT_TIME_GAP_MS = 15 * 6e4;
     SEGMENT_TOPIC_THRESHOLD = 0.35;
     SEGMENT_MAX_TURNS = 100;
@@ -45875,116 +46407,6 @@ var init_file_queue_store = __esm({
   }
 });
 
-// ../../packages/companion-core/src/node/ollama.ts
-function defaultOllamaConfig() {
-  const base = globalThis.process?.env ?? {};
-  return {
-    baseUrl: base.OLLAMA_URL ?? "http://localhost:11434",
-    embedModel: base.OLLAMA_EMBED_MODEL ?? OLLAMA_EMBED_MODEL,
-    chatModel: base.OLLAMA_CHAT_MODEL ?? OLLAMA_CHAT_MODEL
-  };
-}
-function ollamaEmbed(cfg = defaultOllamaConfig()) {
-  return async (text) => {
-    const res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/embeddings`, {
-      method: "POST",
-      headers: { "content-type": "application/json" },
-      body: JSON.stringify({ model: cfg.embedModel, prompt: text })
-    });
-    if (!res.ok) {
-      throw new Error(`ollama embed failed: ${res.status} ${await res.text().catch(() => "")}`);
-    }
-    const body = await res.json();
-    const v = body.embedding ?? [];
-    let norm = 0;
-    for (const x of v) norm += x * x;
-    norm = Math.sqrt(norm) || 1;
-    return v.map((x) => x / norm);
-  };
-}
-function ollamaTokenize() {
-  return (text) => Math.max(1, Math.ceil(text.length / QWEN_CHARS_PER_TOKEN));
-}
-function ollamaSummarize(cfg = defaultOllamaConfig()) {
-  return async ({ prompt, maxTokens }) => {
-    const res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/chat`, {
-      method: "POST",
-      headers: { "content-type": "application/json" },
-      body: JSON.stringify({
-        model: cfg.chatModel,
-        stream: false,
-        // Disable reasoning. qwen3 (the default summariser family) is a
-        // thinking model: with `think` on it spends the entire
-        // `num_predict` budget on a <think> block and returns EMPTY
-        // content, so the summariser saw "" and the whole pipeline
-        // crash-looped at preflight. We only want the final terse
-        // abstract, never the chain-of-thought. Ollama ignores this
-        // field for non-thinking models, so it's safe across families.
-        think: false,
-        options: {
-          temperature: SUMMARISER_TEMPERATURE,
-          num_predict: Math.min(maxTokens, SUMMARISER_MAX_TOKENS)
-        },
-        messages: [
-          { role: "system", content: SUMMARISER_SYSTEM_PROMPT },
-          { role: "user", content: prompt }
-        ]
-      })
-    });
-    if (!res.ok) {
-      throw new Error(`ollama summarize failed: ${res.status} ${await res.text().catch(() => "")}`);
-    }
-    const body = await res.json();
-    return (body.message?.content ?? "").trim().slice(0, 240);
-  };
-}
-function ollamaCognize(cfg = defaultOllamaConfig()) {
-  return async ({ abstract }) => {
-    if (!abstract || abstract.trim().length < 12) return null;
-    let res;
-    try {
-      res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/chat`, {
-        method: "POST",
-        headers: { "content-type": "application/json" },
-        body: JSON.stringify({
-          model: cfg.chatModel,
-          stream: false,
-          format: "json",
-          // Same reason as the summariser: no thinking budget, just the
-          // JSON cognition tags. Thinking models otherwise emit a long
-          // <think> block and return empty content.
-          think: false,
-          options: {
-            temperature: COGNITION_TEMPERATURE,
-            num_predict: COGNITION_MAX_TOKENS
-          },
-          messages: [
-            { role: "system", content: COGNITION_SYSTEM_PROMPT },
-            { role: "user", content: buildCognitionUserPrompt(abstract) }
-          ]
-        })
-      });
-    } catch {
-      return null;
-    }
-    if (!res.ok) return null;
-    let body;
-    try {
-      body = await res.json();
-    } catch {
-      return null;
-    }
-    return parseCognitionReply(body.message?.content ?? "");
-  };
-}
-var init_ollama = __esm({
-  "../../packages/companion-core/src/node/ollama.ts"() {
-    "use strict";
-    init_prompts();
-    init_cognition();
-  }
-});
-
 // ../../packages/companion-core/src/node/llama.ts
 import { existsSync as existsSync7 } from "fs";
 import { mkdir } from "fs/promises";
@@ -46113,7 +46535,14 @@ async function loadOnce(cfg) {
       contextSequence: scriptContext.getSequence(),
       systemPrompt: SCRIPT_SUMMARY_SYSTEM_PROMPT
     });
-    loaded = { summarizer, cognizer, entitler, scriptSummarizer: scriptSummarizer2 };
+    const linkExtractContext = await model.createContext({
+      contextSize: cfg.contextSize
+    });
+    const linkExtractor = new llamaMod.LlamaChatSession({
+      contextSequence: linkExtractContext.getSequence(),
+      systemPrompt: LINK_EXTRACT_SYSTEM_PROMPT
+    });
+    loaded = { summarizer, cognizer, entitler, scriptSummarizer: scriptSummarizer2, linkExtractor };
     return loaded;
   })();
   try {
@@ -46232,6 +46661,34 @@ function llamaScriptSummarize(cfg = defaultLlamaConfig()) {
     }
   };
 }
+function llamaExtractLinks(cfg = defaultLlamaConfig()) {
+  return async ({ abstracts }) => {
+    if (abstracts.length === 0) return null;
+    let loadedSessions;
+    try {
+      loadedSessions = await loadOnce(cfg);
+    } catch {
+      return null;
+    }
+    const { linkExtractor } = loadedSessions;
+    const run = inflight.then(async () => {
+      linkExtractor.resetChatHistory();
+      const raw = await linkExtractor.prompt(buildLinkExtractUserPrompt(abstracts), {
+        temperature: LINK_EXTRACT_TEMPERATURE,
+        // Same thinking-budget rationale as cognition/title: the answer is a
+        // few short lines but Qwen3.5 reasons first.
+        maxTokens: LINK_EXTRACT_MAX_TOKENS + 400
+      });
+      return stripThinking(raw ?? "") || null;
+    });
+    inflight = run.catch(() => void 0);
+    try {
+      return await run;
+    } catch {
+      return null;
+    }
+  };
+}
 var DEFAULT_LLAMA_MODEL_URL, LLAMA_MAX_TOKENS, loaded, loadPromise, inflight;
 var init_llama = __esm({
   "../../packages/companion-core/src/node/llama.ts"() {
@@ -46239,6 +46696,7 @@ var init_llama = __esm({
     init_cognition();
     init_prompts();
     init_script_summary();
+    init_session_metadata2();
     init_title();
     DEFAULT_LLAMA_MODEL_URL = "https://huggingface.co/lmstudio-community/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf";
     LLAMA_MAX_TOKENS = 1024;
@@ -46248,6 +46706,116 @@ var init_llama = __esm({
   }
 });
 
+// ../../packages/companion-core/src/node/ollama.ts
+function defaultOllamaConfig() {
+  const base = globalThis.process?.env ?? {};
+  return {
+    baseUrl: base.OLLAMA_URL ?? "http://localhost:11434",
+    embedModel: base.OLLAMA_EMBED_MODEL ?? OLLAMA_EMBED_MODEL,
+    chatModel: base.OLLAMA_CHAT_MODEL ?? OLLAMA_CHAT_MODEL
+  };
+}
+function ollamaEmbed(cfg = defaultOllamaConfig()) {
+  return async (text) => {
+    const res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/embeddings`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({ model: cfg.embedModel, prompt: text })
+    });
+    if (!res.ok) {
+      throw new Error(`ollama embed failed: ${res.status} ${await res.text().catch(() => "")}`);
+    }
+    const body = await res.json();
+    const v = body.embedding ?? [];
+    let norm = 0;
+    for (const x of v) norm += x * x;
+    norm = Math.sqrt(norm) || 1;
+    return v.map((x) => x / norm);
+  };
+}
+function ollamaTokenize() {
+  return (text) => Math.max(1, Math.ceil(text.length / QWEN_CHARS_PER_TOKEN));
+}
+function ollamaSummarize(cfg = defaultOllamaConfig()) {
+  return async ({ prompt, maxTokens }) => {
+    const res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/chat`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        model: cfg.chatModel,
+        stream: false,
+        // Disable reasoning. qwen3 (the default summariser family) is a
+        // thinking model: with `think` on it spends the entire
+        // `num_predict` budget on a <think> block and returns EMPTY
+        // content, so the summariser saw "" and the whole pipeline
+        // crash-looped at preflight. We only want the final terse
+        // abstract, never the chain-of-thought. Ollama ignores this
+        // field for non-thinking models, so it's safe across families.
+        think: false,
+        options: {
+          temperature: SUMMARISER_TEMPERATURE,
+          num_predict: Math.min(maxTokens, SUMMARISER_MAX_TOKENS)
+        },
+        messages: [
+          { role: "system", content: SUMMARISER_SYSTEM_PROMPT },
+          { role: "user", content: prompt }
+        ]
+      })
+    });
+    if (!res.ok) {
+      throw new Error(`ollama summarize failed: ${res.status} ${await res.text().catch(() => "")}`);
+    }
+    const body = await res.json();
+    return (body.message?.content ?? "").trim().slice(0, 240);
+  };
+}
+function ollamaCognize(cfg = defaultOllamaConfig()) {
+  return async ({ abstract }) => {
+    if (!abstract || abstract.trim().length < 12) return null;
+    let res;
+    try {
+      res = await fetch(`${cfg.baseUrl.replace(/\/+$/, "")}/api/chat`, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({
+          model: cfg.chatModel,
+          stream: false,
+          format: "json",
+          // Same reason as the summariser: no thinking budget, just the
+          // JSON cognition tags. Thinking models otherwise emit a long
+          // <think> block and return empty content.
+          think: false,
+          options: {
+            temperature: COGNITION_TEMPERATURE,
+            num_predict: COGNITION_MAX_TOKENS
+          },
+          messages: [
+            { role: "system", content: COGNITION_SYSTEM_PROMPT },
+            { role: "user", content: buildCognitionUserPrompt(abstract) }
+          ]
+        })
+      });
+    } catch {
+      return null;
+    }
+    if (!res.ok) return null;
+    let body;
+    try {
+      body = await res.json();
+    } catch {
+      return null;
+    }
+    return parseCognitionReply(body.message?.content ?? "");
+  };
+}
+var init_ollama = __esm({
+  "../../packages/companion-core/src/node/ollama.ts"() {
+    "use strict";
+    init_prompts();
+    init_cognition();
+  }
+});
+
 // ../../packages/companion-core/src/optional-module.ts
 function isMissingOptionalModuleError(err) {
   const code = err?.code;
@@ -46346,6 +46914,7 @@ __export(node_exports, {
   ensureLlamaModel: () => ensureLlamaModel,
   llamaCognize: () => llamaCognize,
   llamaEntitle: () => llamaEntitle,
+  llamaExtractLinks: () => llamaExtractLinks,
   llamaScriptSummarize: () => llamaScriptSummarize,
   llamaSummarize: () => llamaSummarize,
   ollamaCognize: () => ollamaCognize,
@@ -46357,9 +46926,8 @@ var init_node2 = __esm({
   "../../packages/companion-core/src/node/index.ts"() {
     "use strict";
     init_file_queue_store();
-    init_file_queue_store();
-    init_ollama();
     init_llama();
+    init_ollama();
     init_transformersjs_embed();
   }
 });
@@ -46548,6 +47116,7 @@ var init_enrich_scripts = __esm({
 var pipeline_exports = {};
 __export(pipeline_exports, {
   buildSegments: () => buildSegments,
+  buildSessionMetadata: () => buildSessionMetadata2,
   buildSessionTitles: () => buildSessionTitles2,
   enrichScripts: () => enrichScripts,
   preflightSummariser: () => preflightSummariser
@@ -46571,6 +47140,13 @@ async function bundledAdapters() {
     // sessions-list title in the dashboard). Best-effort like cognize:
     // failures fall back to a deterministic title in buildSessionTitles.
     entitle: llamaEntitle(llamaCfg),
+    // Link-extraction pass — same bundled model, fifth chat session with
+    // LINK_EXTRACT_SYSTEM_PROMPT. One short call per session that surfaces
+    // PR/issue/commit references from the redacted abstracts, so detection
+    // works even for clients whose logs carry no git data. Best-effort:
+    // failures fall back to the deterministic git + content channels in
+    // buildSessionMetadata.
+    extractLinks: llamaExtractLinks(llamaCfg),
     // Model-based PII redactor (OpenAI Privacy Filter via
     // transformers.js / ONNX). Runs locally on CPU after the regex
     // pass in packages/core/redact.ts. ~1 GB model downloaded on
@@ -46602,6 +47178,13 @@ async function buildSessionTitles2(segments) {
   const a = await getAdapters();
   return buildSessionTitles(segments, a.entitle);
 }
+async function buildSessionMetadata2(segments, events) {
+  const a = await getAdapters();
+  return buildSessionMetadata(segments, events, {
+    resolveGit: resolveGitContext,
+    extractLinks: a.extractLinks
+  });
+}
 async function enrichScripts(drafts, contexts = []) {
   if (contexts.length === 0 || drafts.length === 0) return;
   await getAdapters();
@@ -46635,6 +47218,7 @@ var init_pipeline2 = __esm({
     init_node2();
     init_pipeline();
     init_privacy_filter();
+    init_src2();
     init_enrich_scripts();
     adapters = null;
     MAX_SCRIPT_READ_BYTES = 64 * 1024;
@@ -46702,6 +47286,14 @@ async function scanAll(cb = {}) {
   } catch (e) {
     console.warn("codex scan skipped:", e.message);
   }
+  const ordered = (await Promise.all(
+    jobs.map(async (j) => ({
+      job: j,
+      mtimeMs: (await stat2(j.path).catch(() => null))?.mtimeMs ?? 0
+    }))
+  )).sort((a, b) => b.mtimeMs - a.mtimeMs).map((x) => x.job);
+  const MAX_FILES_PER_SCAN = 12;
+  let morePending = false;
   let filesScanned = 0;
   let filesUnchanged = 0;
   let batchesUploaded = 0;
@@ -46730,6 +47322,15 @@ async function scanAll(cb = {}) {
     } catch (e) {
       console.warn("session titling failed \u2014 shipping batch untitled:", e.message);
     }
+    let sessionMetadata = {};
+    try {
+      sessionMetadata = await buildSessionMetadata2(titleInput, events);
+    } catch (e) {
+      console.warn(
+        "session metadata detection failed \u2014 shipping batch without it:",
+        e.message
+      );
+    }
     const callSegmentByEvent = /* @__PURE__ */ new Map();
     for (const sessionId of new Set(toolCallBuffer.map((c) => c.session_id))) {
       for (const seg of runSegmentsBySession.get(sessionId) ?? []) {
@@ -46743,7 +47344,8 @@ async function scanAll(cb = {}) {
       events,
       segments,
       tool_calls: attachSegmentIdsByMap(toolCallBuffer, callSegmentByEvent),
-      ...Object.keys(sessionTitles).length ? { session_titles: sessionTitles } : {}
+      ...Object.keys(sessionTitles).length ? { session_titles: sessionTitles } : {},
+      ...Object.keys(sessionMetadata).length ? { session_metadata: sessionMetadata } : {}
     };
     cb.onUpload?.({ events: events.length, segments: segments.length });
     const res = await uploadBatch(batch);
@@ -46773,9 +47375,9 @@ async function scanAll(cb = {}) {
       toolCallBuffer.push(c);
     }
   };
-  for (let i = 0; i < jobs.length; i++) {
-    const job = jobs[i];
-    cb.onFile?.(job.path, i, jobs.length);
+  for (let i = 0; i < ordered.length; i++) {
+    const job = ordered[i];
+    cb.onFile?.(job.path, i, ordered.length);
     const cur = state.getCursor(job.path);
     const cs = await quickChecksum(job.path).catch(() => null);
     if (cs && cur && cur.size === cs.size && cur.tailHash === cs.tailHash) {
@@ -46798,9 +47400,20 @@ async function scanAll(cb = {}) {
     } catch (e) {
       console.warn(`  ! parse failed for ${job.path}:`, e.message);
     }
+    if (filesScanned >= MAX_FILES_PER_SCAN) {
+      morePending = true;
+      break;
+    }
   }
   await flushBatch();
-  return { filesScanned, filesUnchanged, batchesUploaded, eventsUploaded, segmentsUploaded };
+  return {
+    filesScanned,
+    filesUnchanged,
+    batchesUploaded,
+    eventsUploaded,
+    segmentsUploaded,
+    morePending
+  };
 }
 var AGENT_VERSION, BATCH_MAX_EVENTS, BATCH_MAX_TOOL_CALLS, BATCH_BUFFER_HARD_CAP, ZERO_TOKENS;
 var init_scan = __esm({
@@ -46813,7 +47426,7 @@ var init_scan = __esm({
     init_api();
     init_config2();
     init_pipeline2();
-    AGENT_VERSION = true ? "agent-0.0.52" : "agent-dev";
+    AGENT_VERSION = true ? "agent-0.1.0" : "agent-dev";
     BATCH_MAX_EVENTS = INGEST_BATCH_MAX_EVENTS;
     BATCH_MAX_TOOL_CALLS = 2e4;
     BATCH_BUFFER_HARD_CAP = BATCH_MAX_EVENTS * 2;
@@ -49177,8 +49790,13 @@ async function runScanCycle(reason) {
     });
     bumpStat("files_scanned", r.filesScanned);
     bumpStat("files_unchanged", r.filesUnchanged);
-    setPhase("watching", "Waiting for new events");
-    setProgress(0, 0);
+    if (r.morePending) {
+      setPhase("processing", "Catching up on history\u2026");
+      setTimeout(() => void requestScan("backfill"), 250);
+    } else {
+      setPhase("watching", "Waiting for new events");
+      setProgress(0, 0);
+    }
   } catch (e) {
     setStat("segments_sending", 0);
     setPhase("offline", `Upload failed: ${describeErrorWithCause(e)}`);
@@ -49313,7 +49931,7 @@ var init_daemon = __esm({
     init_machine_key();
     init_scan();
     init_single_flight();
-    AGENT_VERSION2 = true ? "agent-0.0.52" : "agent-dev";
+    AGENT_VERSION2 = true ? "agent-0.1.0" : "agent-dev";
     HEARTBEAT_INTERVAL_MS = 1e4;
     SCAN_INTERVAL_MS = 5 * 60 * 1e3;
     DISCOVERY_INTERVAL_MS = 6e4;
@@ -49527,6 +50145,8 @@ function installNativeRuntime(sourceCli) {
   const childEnv = { ...process.env };
   delete childEnv.npm_config_global;
   delete childEnv.npm_config_prefix;
+  process.stderr.write(`  \xB7 staging summariser runtime (node-llama-cpp@${version})\u2026
+`);
   const r = spawnSync2(
     "npm",
     [
@@ -49538,6 +50158,13 @@ function installNativeRuntime(sourceCli) {
       "--omit=dev",
       "--no-audit",
       "--no-fund",
+      // Prefer the npm cache the current install already populated (node-llama-cpp
+      // is a direct dep, so the platform prebuilt is cached) — an offline ~3s copy
+      // instead of a redundant network re-fetch. Only a genuine cache miss touches
+      // the network, and a capped per-request timeout makes that fail fast (the
+      // daemon's summariser preflight re-stages) rather than hang indefinitely.
+      "--prefer-offline",
+      "--fetch-timeout=60000",
       "--loglevel=error",
       `node-llama-cpp@${version}`
     ],
@@ -49906,7 +50533,7 @@ function tryOpenBrowser(url) {
     return false;
   }
 }
-var AGENT_VERSION3 = true ? "agent-0.0.52" : "agent-dev";
+var AGENT_VERSION3 = true ? "agent-0.1.0" : "agent-dev";
 function osFamily() {
   const p = platform5();
   if (p === "darwin") return "macos";