evrex-mcp 0.8.0 → 0.8.1

package/dist/capture.js

@@ -180,7 +180,7 @@ function commitMeta(repoPath, sha) {
   const line = git(repoPath, [
     "show",
     "-s",
-    `--format=%H${FIELD_SEP}%an${FIELD_SEP}%ae${FIELD_SEP}%aI`,
+    `--format=%H${FIELD_SEP}%an${FIELD_SEP}%ae${FIELD_SEP}%aI${FIELD_SEP}%P`,
     sha
   ]).trim();
   const parts = line.split(FIELD_SEP);
@@ -190,6 +190,7 @@ function commitMeta(repoPath, sha) {
     author: parts[1] ?? "",
     authorEmail: parts[2] ?? "",
     ts: parts[3] ?? (/* @__PURE__ */ new Date()).toISOString(),
+    parents: (parts[4] ?? "").split(" ").filter(Boolean),
     message
   };
 }
@@ -238,6 +239,10 @@ function trailerValue(trailers, key) {
   const k = key.toLowerCase();
   return trailers.find((t) => t.key.toLowerCase() === k)?.value ?? null;
 }
+function originOf(trailers) {
+  const v = trailerValue(trailers, EVREX_ORIGIN_TRAILER_KEY);
+  return v?.trim().toLowerCase() === "human" ? "human" : null;
+}
 function agentTrailersOf(trailers) {
   const out = [];
   for (const t of trailers) {
@@ -319,15 +324,17 @@ function parseGitLog(repoPath, repoId, known) {
       authorEmail: meta.authorEmail,
       ts: meta.ts,
       message: meta.message,
+      parents: meta.parents,
       branch: commitBranch(repoPath, sha),
       evrexSessionTrailer: trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
+      origin: originOf(trailers),
       agentTrailers: agentTrailersOf(trailers),
       statedInsights: statedInsightsOf(trailers),
       files: commitFiles(repoPath, sha)
     };
   });
 }
-var DIFF_CAP, FIELD_SEP, EVREX_SESSION_TRAILER_KEY;
+var DIFF_CAP, FIELD_SEP, EVREX_SESSION_TRAILER_KEY, EVREX_ORIGIN_TRAILER_KEY;
 var init_git_history = __esm({
   "../../packages/ingest-core/src/git-history.ts"() {
     "use strict";
@@ -335,6 +342,7 @@ var init_git_history = __esm({
     DIFF_CAP = 2e4;
     FIELD_SEP = "";
     EVREX_SESSION_TRAILER_KEY = "Evrex-Session";
+    EVREX_ORIGIN_TRAILER_KEY = "Evrex-Origin";
   }
 });
 
@@ -811,6 +819,7 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
   let totalRedactions = 0;
   let branch = null;
   let agentId = null;
+  const seenIds = /* @__PURE__ */ new Set();
   for (const line of lines) {
     let record;
     try {
@@ -826,6 +835,8 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     const id = record.uuid;
     const ts = record.timestamp;
     if (!id || !ts) continue;
+    if (seenIds.has(id)) continue;
+    seenIds.add(id);
     sessionId ??= record.sessionId ?? record.session_id ?? null;
     if (typeof record.gitBranch === "string" && record.gitBranch) branch = record.gitBranch;
     if (typeof record.agentId === "string" && record.agentId) agentId ??= record.agentId;
@@ -2928,6 +2939,7 @@ var src_exports = {};
 __export(src_exports, {
   CONVERSATION_KINDS: () => CONVERSATION_KINDS,
   EMPTY_USAGE: () => EMPTY_USAGE,
+  EVREX_ORIGIN_TRAILER_KEY: () => EVREX_ORIGIN_TRAILER_KEY,
   EVREX_SESSION_TRAILER_KEY: () => EVREX_SESSION_TRAILER_KEY,
   JIRA_FIELDS: () => JIRA_FIELDS,
   JIRA_SEARCH_PATH: () => JIRA_SEARCH_PATH,
@@ -2986,6 +2998,7 @@ __export(src_exports, {
   measureLineSurvival: () => measureLineSurvival,
   normalizeRepoRemote: () => normalizeRepoRemote,
   opencodeDbPath: () => opencodeDbPath,
+  originOf: () => originOf,
   parseAllCodexSessions: () => parseAllCodexSessions,
   parseAllCopilotSessions: () => parseAllCopilotSessions,
   parseAllCursorSessions: () => parseAllCursorSessions,

package/dist/import.js

@@ -180,7 +180,7 @@ function commitMeta(repoPath, sha) {
   const line = git(repoPath, [
     "show",
     "-s",
-    `--format=%H${FIELD_SEP}%an${FIELD_SEP}%ae${FIELD_SEP}%aI`,
+    `--format=%H${FIELD_SEP}%an${FIELD_SEP}%ae${FIELD_SEP}%aI${FIELD_SEP}%P`,
     sha
   ]).trim();
   const parts = line.split(FIELD_SEP);
@@ -190,6 +190,7 @@ function commitMeta(repoPath, sha) {
     author: parts[1] ?? "",
     authorEmail: parts[2] ?? "",
     ts: parts[3] ?? (/* @__PURE__ */ new Date()).toISOString(),
+    parents: (parts[4] ?? "").split(" ").filter(Boolean),
     message
   };
 }
@@ -238,6 +239,10 @@ function trailerValue(trailers, key) {
   const k = key.toLowerCase();
   return trailers.find((t) => t.key.toLowerCase() === k)?.value ?? null;
 }
+function originOf(trailers) {
+  const v = trailerValue(trailers, EVREX_ORIGIN_TRAILER_KEY);
+  return v?.trim().toLowerCase() === "human" ? "human" : null;
+}
 function agentTrailersOf(trailers) {
   const out = [];
   for (const t of trailers) {
@@ -319,15 +324,17 @@ function parseGitLog(repoPath, repoId, known) {
       authorEmail: meta.authorEmail,
       ts: meta.ts,
       message: meta.message,
+      parents: meta.parents,
       branch: commitBranch(repoPath, sha),
       evrexSessionTrailer: trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
+      origin: originOf(trailers),
       agentTrailers: agentTrailersOf(trailers),
       statedInsights: statedInsightsOf(trailers),
       files: commitFiles(repoPath, sha)
     };
   });
 }
-var DIFF_CAP, FIELD_SEP, EVREX_SESSION_TRAILER_KEY;
+var DIFF_CAP, FIELD_SEP, EVREX_SESSION_TRAILER_KEY, EVREX_ORIGIN_TRAILER_KEY;
 var init_git_history = __esm({
   "../../packages/ingest-core/src/git-history.ts"() {
     "use strict";
@@ -335,6 +342,7 @@ var init_git_history = __esm({
     DIFF_CAP = 2e4;
     FIELD_SEP = "";
     EVREX_SESSION_TRAILER_KEY = "Evrex-Session";
+    EVREX_ORIGIN_TRAILER_KEY = "Evrex-Origin";
   }
 });
 
@@ -811,6 +819,7 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
   let totalRedactions = 0;
   let branch = null;
   let agentId = null;
+  const seenIds = /* @__PURE__ */ new Set();
   for (const line of lines) {
     let record;
     try {
@@ -826,6 +835,8 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     const id = record.uuid;
     const ts = record.timestamp;
     if (!id || !ts) continue;
+    if (seenIds.has(id)) continue;
+    seenIds.add(id);
     sessionId ??= record.sessionId ?? record.session_id ?? null;
     if (typeof record.gitBranch === "string" && record.gitBranch) branch = record.gitBranch;
     if (typeof record.agentId === "string" && record.agentId) agentId ??= record.agentId;
@@ -2928,6 +2939,7 @@ var src_exports = {};
 __export(src_exports, {
   CONVERSATION_KINDS: () => CONVERSATION_KINDS,
   EMPTY_USAGE: () => EMPTY_USAGE,
+  EVREX_ORIGIN_TRAILER_KEY: () => EVREX_ORIGIN_TRAILER_KEY,
   EVREX_SESSION_TRAILER_KEY: () => EVREX_SESSION_TRAILER_KEY,
   JIRA_FIELDS: () => JIRA_FIELDS,
   JIRA_SEARCH_PATH: () => JIRA_SEARCH_PATH,
@@ -2986,6 +2998,7 @@ __export(src_exports, {
   measureLineSurvival: () => measureLineSurvival,
   normalizeRepoRemote: () => normalizeRepoRemote,
   opencodeDbPath: () => opencodeDbPath,
+  originOf: () => originOf,
   parseAllCodexSessions: () => parseAllCodexSessions,
   parseAllCopilotSessions: () => parseAllCopilotSessions,
   parseAllCursorSessions: () => parseAllCursorSessions,
@@ -3544,7 +3557,44 @@ async function durability(target, argv) {
   const surviving = measurements.reduce((n, m) => n + m.survivingLines, 0);
   console.error(`  ${body.updated} commits recorded \xB7 ${surviving}/${added} added lines still at HEAD (${added ? Math.round(100 * surviving / added) : 0}%)`);
 }
+async function parents(target) {
+  const { deriveRepoId: deriveRepoId2 } = await Promise.resolve().then(() => (init_src(), src_exports));
+  const { execFileSync: execFileSync5 } = await import("node:child_process");
+  const { evrexApi: evrexApi2 } = await Promise.resolve().then(() => (init_client(), client_exports));
+  const token = process.env.EVREX_TOKEN ?? await new (await Promise.resolve().then(() => (init_credential_store(), credential_store_exports))).CredentialStore().retrieve().catch(() => null);
+  if (!token) {
+    console.error("evrex: this machine is not enrolled.\n  Run `npx -y evrex-mcp enrol` first, or set EVREX_TOKEN.");
+    process.exit(1);
+  }
+  const repoId = deriveRepoId2(target);
+  const rows = execFileSync5("git", ["rev-list", "--parents", "--all"], { cwd: target, maxBuffer: 64 * 1024 * 1024 }).toString("utf-8").split("\n").filter(Boolean).map((line) => {
+    const [sha, ...parents2] = line.split(" ");
+    return { sha, parents: parents2 };
+  });
+  console.error(`Reporting parents for ${rows.length} commits in ${target} (${rows.filter((r) => r.parents.length > 1).length} merges)`);
+  let updated = 0;
+  let inherited = 0;
+  for (let i = 0; i < rows.length; i += 500) {
+    const res = await fetch(`${evrexApi2.baseUrl}/ingest/commits/parents`, {
+      method: "POST",
+      headers: { "content-type": "application/json", authorization: `Bearer ${token}` },
+      body: JSON.stringify({ repoId, commits: rows.slice(i, i + 500) })
+    });
+    if (!res.ok) {
+      console.error(`  /ingest/commits/parents -> ${res.status}`);
+      process.exit(1);
+    }
+    const body = await res.json();
+    updated += body.updated;
+    inherited = body.inherited;
+  }
+  console.error(`  ${updated} commits updated \xB7 merges now inherit ${inherited} link(s)`);
+}
 async function main() {
+  if (process.argv[2] === "parents") {
+    await parents(resolve(process.argv[3] && !process.argv[3].startsWith("--") ? process.argv[3] : process.cwd()));
+    return;
+  }
   if (process.argv[2] === "durability") {
     await durability(resolve(process.argv[3] && !process.argv[3].startsWith("--") ? process.argv[3] : process.cwd()), process.argv.slice(3));
     return;

package/dist/index.js

@@ -646,6 +646,29 @@ ${request2.user}`;
 // ../../packages/llm-core/src/provider-clients.ts
 import Anthropic3 from "@anthropic-ai/sdk";
 var NOOP = { warn: () => void 0, error: () => void 0 };
+var BATCH_POLL_MS = 5e3;
+var BATCH_TIMEOUT_MS = 6 * 60 * 60 * 1e3;
+var sleep = (ms) => new Promise((r) => setTimeout(r, ms));
+function totalize(items, results) {
+  for (const item of items) if (!results.has(item.id)) results.set(item.id, null);
+  return results;
+}
+function totalizeIds(ids, results) {
+  for (const id of ids) if (!results.has(id)) results.set(id, null);
+  return results;
+}
+async function pollToCompletion(handle, resolve, sleepFn, onTimeout) {
+  const startedAt = Date.now();
+  for (; ; ) {
+    const resolved = await resolve(handle);
+    if (resolved) return resolved;
+    if (Date.now() - startedAt > BATCH_TIMEOUT_MS) {
+      onTimeout(handle.id);
+      return null;
+    }
+    await sleepFn(BATCH_POLL_MS);
+  }
+}
 var realSleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
 var MAX_RETRIES = 4;
 var RETRYABLE_STATUS = /* @__PURE__ */ new Set([429, 500, 502, 503, 504]);
@@ -653,12 +676,12 @@ function isExhaustedStatus(status, retriesSpent) {
   if (status === 401 || status === 403) return true;
   return status === 429 && retriesSpent;
 }
-async function waitBeforeRetry(response, attempt, sleep) {
+async function waitBeforeRetry(response, attempt, sleep2) {
   const header = response.headers.get("retry-after") ?? response.headers.get("x-ratelimit-reset-tokens");
   const seconds = header ? parseDuration(header) : null;
   const backoff = Math.min(8 * 2 ** attempt, 60);
   const wait = Math.min(seconds ?? backoff, MAX_RETRY_WAIT_SECONDS);
-  await sleep(wait * 1e3);
+  await sleep2(wait * 1e3);
 }
 var MAX_RETRY_WAIT_SECONDS = 90;
 function parseDuration(value) {
@@ -678,7 +701,7 @@ function parseJsonBody(text) {
     return null;
   }
 }
-var AnthropicClient = class {
+var AnthropicClient = class _AnthropicClient {
   constructor(key, model, logger) {
     this.model = model;
     this.logger = logger;
@@ -687,37 +710,47 @@ var AnthropicClient = class {
   provider = "anthropic";
   exhausted = false;
   sdk;
+  // The one place a ModelRequest becomes Anthropic's message params, so the
+  // single and batch paths cannot drift on the cache breakpoint or the
+  // json_schema wiring.
+  messageParams(request2) {
+    return {
+      model: request2.model ?? this.model,
+      max_tokens: request2.maxTokens ?? DEFAULT_MAX_TOKENS,
+      // The breakpoint sits at the end of the system prompt, which is the
+      // only part of these requests that repeats. Rendering order is
+      // tools -> system -> messages, so a marker here caches everything
+      // ahead of the transcript.
+      //
+      // Deliberately NOT on the user block. That block is a different
+      // transcript chunk on every call, so a breakpoint there would write a
+      // fresh cache entry per request and read none of them back — paying
+      // the write premium for nothing, which is worse than not caching.
+      system: [
+        {
+          type: "text",
+          text: request2.system,
+          cache_control: { type: "ephemeral" }
+        }
+      ],
+      messages: [{ role: "user", content: request2.user }],
+      ...request2.schema ? {
+        output_config: {
+          format: { type: "json_schema", schema: request2.schema }
+        }
+      } : {}
+    };
+  }
+  static messageText(message) {
+    return message.content.filter((b) => b.type === "text").map((b) => b.text).join("");
+  }
   async completeJson(request2) {
     try {
-      const message = await this.sdk.messages.create({
-        model: request2.model ?? this.model,
-        max_tokens: request2.maxTokens ?? DEFAULT_MAX_TOKENS,
-        // The breakpoint sits at the end of the system prompt, which is the
-        // only part of these requests that repeats. Rendering order is
-        // tools -> system -> messages, so a marker here caches everything
-        // ahead of the transcript.
-        //
-        // Deliberately NOT on the user block. That block is a different
-        // transcript chunk on every call, so a breakpoint there would write a
-        // fresh cache entry per request and read none of them back — paying
-        // the write premium for nothing, which is worse than not caching.
-        system: [
-          {
-            type: "text",
-            text: request2.system,
-            cache_control: { type: "ephemeral" }
-          }
-        ],
-        messages: [{ role: "user", content: request2.user }],
-        ...request2.schema ? {
-          output_config: {
-            format: { type: "json_schema", schema: request2.schema }
-          }
-        } : {}
-      });
+      const message = await this.sdk.messages.create(
+        this.messageParams(request2)
+      );
       this.reportCacheUsage(message.usage);
-      const text = message.content.filter((b) => b.type === "text").map((b) => b.text).join("");
-      return parseJsonBody(text);
+      return parseJsonBody(_AnthropicClient.messageText(message));
     } catch (err) {
       const status = err?.status;
       if (typeof status === "number" && isExhaustedStatus(status, true)) {
@@ -727,6 +760,58 @@ var AnthropicClient = class {
       return null;
     }
   }
+  async submitBatch(items) {
+    if (items.length === 0) return null;
+    try {
+      const created = await this.sdk.messages.batches.create({
+        requests: items.map((item) => ({
+          custom_id: item.id,
+          params: this.messageParams(item.request)
+        }))
+      });
+      return { provider: "anthropic", id: created.id, itemIds: items.map((i) => i.id) };
+    } catch (err) {
+      const status = err?.status;
+      if (typeof status === "number" && isExhaustedStatus(status, true)) {
+        this.exhausted = true;
+      }
+      this.logger.warn(`anthropic batch submit failed: ${describeError(err)}`);
+      return null;
+    }
+  }
+  async resolveBatch(handle) {
+    try {
+      const batch = await this.sdk.messages.batches.retrieve(handle.id);
+      if (batch.processing_status !== "ended") return null;
+      const results = /* @__PURE__ */ new Map();
+      for await (const entry of await this.sdk.messages.batches.results(
+        handle.id
+      )) {
+        results.set(
+          entry.custom_id,
+          entry.result.type === "succeeded" ? parseJsonBody(_AnthropicClient.messageText(entry.result.message)) : null
+        );
+      }
+      return totalizeIds(handle.itemIds, results);
+    } catch (err) {
+      this.logger.warn(`anthropic batch resolve failed: ${describeError(err)}`);
+      return totalizeIds(handle.itemIds, /* @__PURE__ */ new Map());
+    }
+  }
+  async completeJsonBatch(items) {
+    if (items.length === 0) return /* @__PURE__ */ new Map();
+    const handle = await this.submitBatch(items);
+    if (!handle) return totalize(items, /* @__PURE__ */ new Map());
+    const resolved = await pollToCompletion(
+      handle,
+      (h) => this.resolveBatch(h),
+      sleep,
+      (id) => this.logger.warn(
+        `anthropic batch ${id} did not finish within the cap; the sessions it covers stay unextracted until the next pass`
+      )
+    );
+    return resolved ?? totalize(items, /* @__PURE__ */ new Map());
+  }
   /**
    * Says out loud whether the cache was actually used.
    *
@@ -753,12 +838,12 @@ var AnthropicClient = class {
     );
   }
 };
-var GoogleClient = class {
-  constructor(key, model, logger, sleep = realSleep) {
+var GoogleClient = class _GoogleClient {
+  constructor(key, model, logger, sleep2 = realSleep) {
     this.key = key;
     this.model = model;
     this.logger = logger;
-    this.sleep = sleep;
+    this.sleep = sleep2;
   }
   provider = "google";
   exhausted = false;
@@ -801,56 +886,129 @@ ${JSON.stringify(request2.schema)}` : "";
       return null;
     }
   }
+  static GEMINI_BASE = "https://generativelanguage.googleapis.com/v1beta";
+  async submitBatch(items) {
+    if (items.length === 0) return null;
+    const model = items[0].request.model ?? this.model;
+    try {
+      const create = await fetch(
+        `${_GoogleClient.GEMINI_BASE}/models/${encodeURIComponent(model)}:batchGenerateContent`,
+        {
+          method: "POST",
+          headers: {
+            "Content-Type": "application/json",
+            "x-goog-api-key": this.key
+          },
+          body: JSON.stringify({
+            batch: {
+              display_name: "evrex-extraction",
+              input_config: {
+                requests: {
+                  requests: items.map((item) => {
+                    const shape = item.request.schema ? `
+
+Respond with JSON matching exactly this schema:
+${JSON.stringify(item.request.schema)}` : "";
+                    return {
+                      request: {
+                        systemInstruction: {
+                          parts: [{ text: item.request.system + shape }]
+                        },
+                        contents: [
+                          { role: "user", parts: [{ text: item.request.user }] }
+                        ],
+                        generationConfig: { responseMimeType: "application/json" }
+                      },
+                      metadata: { key: item.id }
+                    };
+                  })
+                }
+              }
+            }
+          })
+        }
+      );
+      if (!create.ok) {
+        if (isExhaustedStatus(create.status, true)) this.exhausted = true;
+        this.logger.warn(
+          `google batch submit failed: ${create.status} ${(await create.text()).slice(0, 200)}`
+        );
+        return null;
+      }
+      const name = (await create.json()).name;
+      if (!name) {
+        this.logger.warn("google batch: no job name returned");
+        return null;
+      }
+      return { provider: "google", id: name, itemIds: items.map((i) => i.id) };
+    } catch (err) {
+      this.logger.warn(`google batch submit failed: ${describeError(err)}`);
+      return null;
+    }
+  }
+  async resolveBatch(handle) {
+    try {
+      const poll = await fetch(`${_GoogleClient.GEMINI_BASE}/${handle.id}`, {
+        headers: { "x-goog-api-key": this.key }
+      });
+      if (!poll.ok) {
+        this.logger.warn(`google batch poll failed: ${poll.status}`);
+        return totalizeIds(handle.itemIds, /* @__PURE__ */ new Map());
+      }
+      const job = await poll.json();
+      const state = job.state ?? "";
+      if (state === "JOB_STATE_FAILED" || state === "JOB_STATE_CANCELLED" || state === "JOB_STATE_EXPIRED") {
+        this.logger.warn(`google batch ${handle.id} ended ${state}`);
+        return totalizeIds(handle.itemIds, /* @__PURE__ */ new Map());
+      }
+      if (state !== "JOB_STATE_SUCCEEDED") return null;
+      const inlined = job.response?.inlinedResponses ?? [];
+      const results = /* @__PURE__ */ new Map();
+      handle.itemIds.forEach((id, i) => {
+        const text = inlined[i]?.response?.candidates?.[0]?.content?.parts?.map((p) => p.text ?? "").join("");
+        results.set(id, text ? parseJsonBody(text) : null);
+      });
+      return totalizeIds(handle.itemIds, results);
+    } catch (err) {
+      this.logger.warn(`google batch resolve failed: ${describeError(err)}`);
+      return totalizeIds(handle.itemIds, /* @__PURE__ */ new Map());
+    }
+  }
+  async completeJsonBatch(items) {
+    if (items.length === 0) return /* @__PURE__ */ new Map();
+    const handle = await this.submitBatch(items);
+    if (!handle) return totalize(items, /* @__PURE__ */ new Map());
+    const resolved = await pollToCompletion(
+      handle,
+      (h) => this.resolveBatch(h),
+      this.sleep,
+      (id) => this.logger.warn(
+        `google batch ${id} did not finish within the cap; leaving its sessions for a later pass`
+      )
+    );
+    return resolved ?? totalize(items, /* @__PURE__ */ new Map());
+  }
 };
 var OpenAiCompatibleClient = class {
-  constructor(provider, baseUrl, key, model, logger, sleep = realSleep) {
+  constructor(provider, baseUrl, key, model, logger, sleep2 = realSleep) {
     this.provider = provider;
     this.baseUrl = baseUrl;
     this.key = key;
     this.model = model;
     this.logger = logger;
-    this.sleep = sleep;
+    this.sleep = sleep2;
+    if (this.provider === "openai") {
+      this.completeJsonBatch = (items) => this.runBatch(items);
+    }
   }
   exhausted = false;
+  completeJsonBatch;
   async completeJson(request2, attempt = 0) {
-    const strict = this.provider === "openai";
-    const shape = request2.schema && !strict ? `
-
-Respond with JSON matching exactly this schema:
-${JSON.stringify(request2.schema)}` : "";
     try {
       const res = await fetch(`${this.baseUrl}/chat/completions`, {
         method: "POST",
-        headers: {
-          "Content-Type": "application/json",
-          // A local server does not want one, and some reject an empty bearer
-          // token outright.
-          ...this.key ? { Authorization: `Bearer ${this.key}` } : {}
-        },
-        body: JSON.stringify({
-          model: request2.model ?? this.model,
-          messages: [
-            { role: "system", content: request2.system + shape },
-            { role: "user", content: request2.user }
-          ],
-          ...request2.schema && strict ? {
-            response_format: {
-              type: "json_schema",
-              json_schema: {
-                name: "evrex_result",
-                strict: true,
-                schema: request2.schema
-              }
-            }
-          } : { response_format: { type: "json_object" } },
-          // Extraction is a reading task with a right answer, not a writing
-          // task, and a default sampling temperature makes it a different
-          // answer each run: one model scored 53% and then 20% on the same ten
-          // sessions before this was pinned. Indexing twice must not produce
-          // two different histories.
-          temperature: 0,
-          stream: false
-        })
+        headers: this.authHeaders(),
+        body: JSON.stringify(this.chatBody(request2))
       });
       if (RETRYABLE_STATUS.has(res.status) && attempt < MAX_RETRIES) {
         await waitBeforeRetry(res, attempt, this.sleep);
@@ -871,6 +1029,147 @@ ${JSON.stringify(request2.schema)}` : "";
       return null;
     }
   }
+  authHeaders() {
+    return {
+      "Content-Type": "application/json",
+      // A local server does not want one, and some reject an empty bearer token.
+      ...this.key ? { Authorization: `Bearer ${this.key}` } : {}
+    };
+  }
+  // One place a ModelRequest becomes the chat body, so completeJson and the
+  // batch path cannot drift on response_format or the pinned temperature.
+  chatBody(request2) {
+    const strict = this.provider === "openai";
+    const shape = request2.schema && !strict ? `
+
+Respond with JSON matching exactly this schema:
+${JSON.stringify(request2.schema)}` : "";
+    return {
+      model: request2.model ?? this.model,
+      messages: [
+        { role: "system", content: request2.system + shape },
+        { role: "user", content: request2.user }
+      ],
+      ...request2.schema && strict ? {
+        response_format: {
+          type: "json_schema",
+          json_schema: {
+            name: "evrex_result",
+            strict: true,
+            schema: request2.schema
+          }
+        }
+      } : { response_format: { type: "json_object" } },
+      // A reading task with a right answer: temperature pinned so indexing
+      // twice is the same history.
+      temperature: 0,
+      stream: false
+    };
+  }
+  async runBatch(items) {
+    const results = /* @__PURE__ */ new Map();
+    if (items.length === 0) return results;
+    const auth = this.key ? { Authorization: `Bearer ${this.key}` } : {};
+    try {
+      const jsonl = items.map(
+        (item) => JSON.stringify({
+          custom_id: item.id,
+          method: "POST",
+          url: "/v1/chat/completions",
+          body: this.chatBody(item.request)
+        })
+      ).join("\n");
+      const form = new FormData();
+      form.append("purpose", "batch");
+      form.append(
+        "file",
+        new Blob([jsonl], { type: "application/jsonl" }),
+        "evrex-batch.jsonl"
+      );
+      const upload = await fetch(`${this.baseUrl}/files`, {
+        method: "POST",
+        headers: auth,
+        body: form
+      });
+      if (!upload.ok) {
+        if (isExhaustedStatus(upload.status, true)) this.exhausted = true;
+        this.logger.warn(
+          `openai batch upload failed: ${upload.status} ${(await upload.text()).slice(0, 200)}`
+        );
+        return totalize(items, results);
+      }
+      const fileId = (await upload.json()).id;
+      if (!fileId) return totalize(items, results);
+      const created = await fetch(`${this.baseUrl}/batches`, {
+        method: "POST",
+        headers: this.authHeaders(),
+        body: JSON.stringify({
+          input_file_id: fileId,
+          endpoint: "/v1/chat/completions",
+          completion_window: "24h"
+        })
+      });
+      if (!created.ok) {
+        this.logger.warn(
+          `openai batch create failed: ${created.status} ${(await created.text()).slice(0, 200)}`
+        );
+        return totalize(items, results);
+      }
+      const batchId = (await created.json()).id;
+      if (!batchId) return totalize(items, results);
+      let status = "";
+      let outputFileId;
+      const startedAt = Date.now();
+      while (status !== "completed") {
+        if (Date.now() - startedAt > BATCH_TIMEOUT_MS) {
+          this.logger.warn(
+            `openai batch ${batchId} did not finish within the cap; leaving its sessions for a later pass`
+          );
+          return totalize(items, results);
+        }
+        await this.sleep(BATCH_POLL_MS);
+        const poll = await fetch(`${this.baseUrl}/batches/${batchId}`, {
+          headers: auth
+        });
+        if (!poll.ok) {
+          this.logger.warn(`openai batch poll failed: ${poll.status}`);
+          return totalize(items, results);
+        }
+        const job = await poll.json();
+        status = job.status ?? "";
+        outputFileId = job.output_file_id;
+        if (status === "failed" || status === "expired" || status === "cancelled") {
+          this.logger.warn(`openai batch ${batchId} ended ${status}`);
+          return totalize(items, results);
+        }
+      }
+      if (!outputFileId) return totalize(items, results);
+      const out = await fetch(`${this.baseUrl}/files/${outputFileId}/content`, {
+        headers: auth
+      });
+      if (!out.ok) {
+        this.logger.warn(`openai batch results fetch failed: ${out.status}`);
+        return totalize(items, results);
+      }
+      for (const line of (await out.text()).split("\n")) {
+        if (!line.trim()) continue;
+        try {
+          const row = JSON.parse(line);
+          const content = row.response?.body?.choices?.[0]?.message?.content;
+          if (row.custom_id) {
+            results.set(
+              row.custom_id,
+              typeof content === "string" ? parseJsonBody(content) : null
+            );
+          }
+        } catch {
+        }
+      }
+    } catch (err) {
+      this.logger.warn(`openai batch failed: ${describeError(err)}`);
+    }
+    return totalize(items, results);
+  }
 };
 function describeError(err) {
   return err instanceof Error ? err.message : String(err);
@@ -882,9 +1181,9 @@ function createModelClient(key, options = {}) {
   const model = options.model ?? DEFAULT_MODELS[provider];
   const trimmed = key.trim();
   if (provider === "anthropic") return new AnthropicClient(trimmed, model, logger);
-  const sleep = options.sleep ?? realSleep;
+  const sleep2 = options.sleep ?? realSleep;
   if (provider === "google")
-    return new GoogleClient(trimmed, model, logger, sleep);
+    return new GoogleClient(trimmed, model, logger, sleep2);
   const cli = CLI_SPECS[provider];
   if (cli) return new CliClient(cli, model, logger);
   if (provider === "local") {
@@ -894,7 +1193,7 @@ function createModelClient(key, options = {}) {
       null,
       model,
       logger,
-      sleep
+      sleep2
     );
   }
   const baseUrl = options.baseUrl ?? OPENAI_COMPATIBLE_BASE_URLS[provider];
@@ -905,7 +1204,7 @@ function createModelClient(key, options = {}) {
     trimmed,
     model,
     logger,
-    sleep
+    sleep2
   );
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "evrex-mcp",
-  "version": "0.8.0",
+  "version": "0.8.1",
   "description": "MCP server that gives coding agents the recorded reasoning behind a repo: prior decisions, hard constraints, and approaches already rejected.",
   "keywords": [
     "mcp",
@@ -36,27 +36,28 @@
   "engines": {
     "node": ">=20"
   },
+  "scripts": {
+    "build": "tsc -p tsconfig.json",
+    "bundle": "node scripts/bundle.mjs",
+    "dev": "tsc -p tsconfig.json --watch",
+    "lint": "eslint .",
+    "check-types": "tsc --noEmit",
+    "test": "tsc -p tsconfig.json && node --test dist/*.test.js",
+    "prepack": "pnpm run bundle"
+  },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.72.0",
     "@modelcontextprotocol/sdk": "^1.20.1",
     "zod": "^3.24.1"
   },
   "devDependencies": {
+    "@repo/eslint-config": "workspace:*",
+    "@repo/ingest-core": "workspace:*",
+    "@repo/llm-core": "workspace:*",
+    "@repo/typescript-config": "workspace:*",
     "@types/node": "^24.0.0",
     "esbuild": "^0.25.12",
     "eslint": "^9.39.1",
-    "typescript": "^5.9.2",
-    "@repo/eslint-config": "0.0.0",
-    "@repo/ingest-core": "0.0.0",
-    "@repo/llm-core": "0.0.0",
-    "@repo/typescript-config": "0.0.0"
-  },
-  "scripts": {
-    "build": "tsc -p tsconfig.json",
-    "bundle": "node scripts/bundle.mjs",
-    "dev": "tsc -p tsconfig.json --watch",
-    "lint": "eslint .",
-    "check-types": "tsc --noEmit",
-    "test": "tsc -p tsconfig.json && node --test dist/*.test.js"
+    "typescript": "^5.9.2"
   }
-}
+}