evrex-mcp 0.8.1 → 0.8.2

package/dist/account.js

@@ -49,7 +49,7 @@ var evrexApi = {
   // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
   // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
   // which wants ranked hits fast, not a synthesized paragraph.
-  search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+  search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
   // Everything that happened in a repo, newest first, bounded by days — the
   // same query the desktop Timeline screen makes. Sessions and commits
   // interleaved, each with the handle evrex_expand takes.

package/dist/capture.js

@@ -184,14 +184,34 @@ function commitMeta(repoPath, sha) {
     sha
   ]).trim();
   const parts = line.split(FIELD_SEP);
-  const message = git(repoPath, ["show", "-s", "--format=%B", sha]).replace(/\n+$/, "");
+  const body = git(repoPath, ["show", "-s", "--notes", `--format=%B${FIELD_SEP}%N`, sha]);
+  const sep2 = body.indexOf(FIELD_SEP);
+  let message = (sep2 === -1 ? body : body.slice(0, sep2)).replace(/\n+$/, "");
+  const note = sep2 === -1 ? "" : body.slice(sep2 + FIELD_SEP.length);
+  const noted = /^Evrex-Session:\s*(\S+)\s*$/m.exec(note);
+  if (noted && !/^Evrex-Session:/m.test(message)) {
+    message = `${message}
+
+${noted[0].trim()}`;
+  }
   return {
     sha: parts[0] ?? sha,
     author: parts[1] ?? "",
     authorEmail: parts[2] ?? "",
     ts: parts[3] ?? (/* @__PURE__ */ new Date()).toISOString(),
     parents: (parts[4] ?? "").split(" ").filter(Boolean),
-    message
+    message,
+    // The note's session, exposed on its own as well as folded above. When
+    // the message ALREADY carries a trailer, the fold declines (never rewrite
+    // what was stamped at commit time) — but the parse must still prefer the
+    // note: it exists only because an operator pressed "verify" on a commit
+    // whose message trailer names a session no ingest can resolve. Commit
+    // 843dabb3 is the standing example — its at-commit-time trailer names a
+    // session with no transcript anywhere, so the folding guard alone left it
+    // permanently un-verifiable: every promote re-ingested the dangling id,
+    // bestLink demoted it below the resolvable match, and the button offered
+    // the same commit again.
+    notedSession: noted?.[1] ?? null
   };
 }
 function parseTrailers(repoPath, message) {
@@ -312,6 +332,9 @@ function commitFiles(repoPath, sha) {
 }
 function parseGitLog(repoPath, repoId, known) {
   const shas = listShas(repoPath).filter((sha) => !known?.has(sha));
+  return parseCommitsBySha(repoPath, shas, repoId);
+}
+function parseCommitsBySha(repoPath, shas, repoId) {
   const id = repoId ?? deriveRepoId(repoPath);
   return shas.map((sha) => {
     const meta = commitMeta(repoPath, sha);
@@ -326,7 +349,9 @@ function parseGitLog(repoPath, repoId, known) {
       message: meta.message,
       parents: meta.parents,
       branch: commitBranch(repoPath, sha),
-      evrexSessionTrailer: trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
+      // A note is the operator's explicit later correction, so where both
+      // exist it names the session — see commitMeta's notedSession.
+      evrexSessionTrailer: meta.notedSession ?? trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
       origin: originOf(trailers),
       agentTrailers: agentTrailersOf(trailers),
       statedInsights: statedInsightsOf(trailers),
@@ -407,7 +432,7 @@ var PARSER_VERSION;
 var init_incremental = __esm({
   "../../packages/ingest-core/src/incremental.ts"() {
     "use strict";
-    PARSER_VERSION = 4;
+    PARSER_VERSION = 7;
   }
 });
 
@@ -522,11 +547,54 @@ var init_redact = __esm({
 
 // ../../packages/ingest-core/src/subject-linking.ts
 function matchCommitToSession(commit, sessions) {
+  const saw = sessions.filter(
+    (s) => (s.committedShas ?? []).some(
+      (short) => short.length >= 7 && commit.sha.startsWith(short)
+    )
+  );
+  if (saw.length > 0) {
+    const oneLineage2 = saw.every((s) => s.startedAt === saw[0].startedAt);
+    if (!oneLineage2) {
+      return { sha: commit.sha, sessionId: null, reason: "ambiguous" };
+    }
+    const original2 = [...saw].sort((a, b) => a.endedAt - b.endedAt)[0];
+    return {
+      sha: commit.sha,
+      sessionId: original2.sessionId,
+      evidence: "saw-the-sha",
+      confidence: SHA_CONFIDENCE
+    };
+  }
   const subject = commit.subject.trim();
+  const prMatch = /\(#(\d+)\)$/.exec(subject) ?? /^Merge pull request #(\d+)\b/.exec(subject);
+  if (prMatch) {
+    const pr = Number(prMatch[1]);
+    const mergedIt = sessions.filter(
+      (s) => (s.mergedPrNumbers ?? []).includes(pr)
+    );
+    if (mergedIt.length > 0) {
+      const oneLineage2 = mergedIt.every(
+        (s) => s.startedAt === mergedIt[0].startedAt
+      );
+      if (!oneLineage2) {
+        return { sha: commit.sha, sessionId: null, reason: "ambiguous" };
+      }
+      const original2 = [...mergedIt].sort((a, b) => a.endedAt - b.endedAt)[0];
+      return {
+        sha: commit.sha,
+        sessionId: original2.sessionId,
+        evidence: "merged-the-pr",
+        confidence: SHA_CONFIDENCE
+      };
+    }
+  }
   if (subject.length === 0) {
     return { sha: commit.sha, sessionId: null, reason: "absent" };
   }
-  const ran = sessions.filter((s) => s.committedSubjects.includes(subject));
+  const unsquashed = subject.replace(/\s+\(#\d+\)$/, "");
+  const ran = sessions.filter(
+    (s) => s.committedSubjects.includes(subject) || unsquashed !== subject && s.committedSubjects.includes(unsquashed)
+  );
   if (ran.length === 0) {
     return { sha: commit.sha, sessionId: null, reason: "absent" };
   }
@@ -580,15 +648,61 @@ function committedSubjects(command) {
     if (inline) {
       const subject = inline[2].split("\n")[0].trim();
       if (subject) out.push(subject);
+      continue;
+    }
+    const marginal = /^[^\n]*?-am\s*(['"])([\s\S]*?)\1/.exec(rest) ?? /^[^\n]*?--message[= ]\s*(['"])([\s\S]*?)\1/.exec(rest) ?? /^[^\n]*?-m(['"])([\s\S]*?)\1/.exec(rest);
+    if (marginal) {
+      const subject = marginal[2].split("\n")[0].trim();
+      if (subject) out.push(subject);
+      continue;
+    }
+    const fromFile = /^[^\n]*?(?:-F|--file)[= ]\s*(\S+)/.exec(rest);
+    if (fromFile) {
+      const file = fromFile[1].replace(/^["']|["']$/g, "");
+      const escaped = file.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+      const wrote = new RegExp(
+        `(?:cat\\s*>>?|tee\\s+(?:-a\\s+)?)\\s*["']?${escaped}["']?[^\\n]*<<-?\\s*(['"]?)(\\w+)\\1[^\\n]*\\r?\\n([^\\n]*)`
+      ).exec(command.slice(0, found.index));
+      if (wrote) {
+        const subject = wrote[3].trim();
+        if (subject) out.push(subject);
+      }
     }
   }
   return out;
 }
-var INFERRED_CONFIDENCE;
+function mergeResponseShas(output) {
+  const out = /* @__PURE__ */ new Set();
+  const line = /"sha"\s*:\s*"([0-9a-f]{40})"/g;
+  for (let m = line.exec(output); m !== null; m = line.exec(output)) {
+    out.add(m[1]);
+  }
+  return [...out];
+}
+function commitShaOutputs(output) {
+  const out = /* @__PURE__ */ new Set();
+  const line = /\[[\w./+-]+ (?:\(root-commit\) )?([0-9a-f]{7,12})\]/g;
+  for (let m = line.exec(output); m !== null; m = line.exec(output)) {
+    out.add(m[1]);
+  }
+  return [...out];
+}
+function mergedPrFromCommand(command) {
+  const api = /\/pulls\/(\d+)\/merge\b/.exec(command);
+  if (api) return Number(api[1]);
+  const gh = /\bgh\s+pr\s+merge\s+(?:.*?\/pull\/(\d+)\b|(\d+)\b)/.exec(command);
+  if (gh) return Number(gh[1] ?? gh[2]);
+  return null;
+}
+function mergeSucceeded(output) {
+  return /"merged"\s*:\s*true/.test(output) || /(Squashed and |Rebased and )?[Mm]erged pull request #\d+/.test(output);
+}
+var INFERRED_CONFIDENCE, SHA_CONFIDENCE;
 var init_subject_linking = __esm({
   "../../packages/ingest-core/src/subject-linking.ts"() {
     "use strict";
     INFERRED_CONFIDENCE = 0.9;
+    SHA_CONFIDENCE = 0.95;
   }
 });
 
@@ -890,6 +1004,7 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
   const subagent = agentId ? { agentId, agentType: meta?.agentType ?? null, description: meta?.description ?? null } : null;
   if (agentId && !aiTitle && meta?.description) aiTitle = meta.description;
   const sortedTs = turns.map((t) => t.ts).sort();
+  const commitEvidence = collectCommitEvidence(lines);
   const rawContent = capped ? "" : lines.map((line) => {
     try {
       return JSON.stringify(redactJsonValue(JSON.parse(line)).value);
@@ -913,7 +1028,9 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     // filled in by the caller — see getGitUserName in git-history.ts
     sourceFile: filePath,
     redactionCount: totalRedactions,
-    committedSubjects: collectCommittedSubjects(lines),
+    committedSubjects: commitEvidence.subjects,
+    committedShas: commitEvidence.shas,
+    mergedPrNumbers: commitEvidence.prNumbers,
     branch,
     parentSessionId,
     subagent,
@@ -922,10 +1039,14 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     turns
   };
 }
-function collectCommittedSubjects(lines) {
+function collectCommitEvidence(lines) {
   const subjects = /* @__PURE__ */ new Set();
+  const shas = /* @__PURE__ */ new Set();
+  const prNumbers = /* @__PURE__ */ new Set();
+  const commitCallIds = /* @__PURE__ */ new Set();
+  const mergeCallIds = /* @__PURE__ */ new Map();
   for (const line of lines) {
-    if (!line.includes("git commit")) continue;
+    if (!line.includes("git commit") && !line.includes("/merge") && !line.includes("tool_use_id")) continue;
     let record;
     try {
       record = JSON.parse(line);
@@ -935,13 +1056,40 @@ function collectCommittedSubjects(lines) {
     const content = record?.message?.content;
     if (!Array.isArray(content)) continue;
     for (const block of content) {
-      if (block?.type !== "tool_use") continue;
-      const command = block?.input?.command;
-      if (typeof command !== "string") continue;
-      for (const subject of committedSubjects(command)) subjects.add(subject);
+      if (block?.type === "tool_use") {
+        const command = block?.input?.command;
+        if (typeof command !== "string") continue;
+        const found = committedSubjects(command);
+        for (const subject of found) subjects.add(subject);
+        if (/\bgit\s+(?:-C\s+\S+\s+)?commit\b/.test(command) && typeof block?.id === "string") {
+          commitCallIds.add(block.id);
+        }
+        if ((/\/pulls\/\d+\/merge\b/.test(command) || /\bgh\s+pr\s+merge\b/.test(command)) && typeof block?.id === "string") {
+          mergeCallIds.set(block.id, mergedPrFromCommand(command));
+        }
+        continue;
+      }
+      if (block?.type === "tool_result" && mergeCallIds.has(block?.tool_use_id)) {
+        const raw = block?.content;
+        const text = typeof raw === "string" ? raw : Array.isArray(raw) ? raw.map(
+          (x) => typeof x?.text === "string" ? x.text : ""
+        ).join(" ") : "";
+        for (const sha of mergeResponseShas(text)) shas.add(sha);
+        if (mergeSucceeded(text)) {
+          const pr = mergeCallIds.get(block?.tool_use_id);
+          if (typeof pr === "number") prNumbers.add(pr);
+        }
+      }
+      if (block?.type === "tool_result" && commitCallIds.has(block?.tool_use_id)) {
+        const raw = block?.content;
+        const text = typeof raw === "string" ? raw : Array.isArray(raw) ? raw.map(
+          (x) => typeof x?.text === "string" ? x.text : ""
+        ).join(" ") : "";
+        for (const sha of commitShaOutputs(text)) shas.add(sha);
+      }
     }
   }
-  return [...subjects];
+  return { subjects: [...subjects], shas: [...shas], prNumbers: [...prNumbers] };
 }
 function parseAllSessions(repoPath, repoId, cursors = {}) {
   const id = repoId ?? deriveRepoId(repoPath);
@@ -1907,16 +2055,58 @@ function addUsage(into, next) {
     model: into.model ?? next.model
   };
 }
-function committedSubjectsFromToolCalls(lines) {
+function commitEvidenceFromToolCalls(lines) {
   const subjects = /* @__PURE__ */ new Set();
+  const shas = /* @__PURE__ */ new Set();
+  const prNumbers = /* @__PURE__ */ new Set();
+  const commitCallIds = /* @__PURE__ */ new Set();
+  const mergeCallIds = /* @__PURE__ */ new Map();
   for (const line of lines) {
     if (line.type !== "response_item") continue;
-    if (payloadType(line) !== "custom_tool_call") continue;
-    const input = asString(line.payload?.input);
-    if (!input) continue;
-    for (const subject of committedSubjects(input)) subjects.add(subject);
+    const kind = payloadType(line);
+    if (kind === "custom_tool_call") {
+      const input = asString(line.payload?.input);
+      if (!input) continue;
+      const cmd = execCmdFrom(input) ?? input;
+      const callId = asString(line.payload?.call_id);
+      for (const subject of committedSubjects(cmd)) subjects.add(subject);
+      if (/\bgit\s+(?:-C\s+\S+\s+)?commit\b/.test(cmd) && callId) {
+        commitCallIds.add(callId);
+      }
+      if ((/\/pulls\/\d+\/merge\b/.test(cmd) || /\bgh\s+pr\s+merge\b/.test(cmd)) && callId) {
+        mergeCallIds.set(callId, mergedPrFromCommand(cmd));
+      }
+      continue;
+    }
+    if (kind === "custom_tool_call_output") {
+      const callId = asString(line.payload?.call_id);
+      if (!callId) continue;
+      const raw = line.payload?.output;
+      const text = Array.isArray(raw) ? raw.map(
+        (x) => typeof x?.text === "string" ? x.text : ""
+      ).join(" ") : asString(raw) ?? "";
+      if (commitCallIds.has(callId)) {
+        for (const sha of commitShaOutputs(text)) shas.add(sha);
+      }
+      if (mergeCallIds.has(callId)) {
+        for (const sha of mergeResponseShas(text)) shas.add(sha);
+        if (mergeSucceeded(text)) {
+          const pr = mergeCallIds.get(callId);
+          if (typeof pr === "number") prNumbers.add(pr);
+        }
+      }
+    }
+  }
+  return { subjects: [...subjects], shas: [...shas], prNumbers: [...prNumbers] };
+}
+function execCmdFrom(input) {
+  const m = /cmd\s*:\s*"((?:[^"\\]|\\.)*)"/.exec(input);
+  if (!m) return null;
+  try {
+    return JSON.parse(`"${m[1]}"`);
+  } catch {
+    return null;
   }
-  return [...subjects];
 }
 function filesFromToolCalls(lines) {
   const byPath = /* @__PURE__ */ new Map();
@@ -1975,20 +2165,28 @@ function parseCodexSessionFile(filePath, repoPath, repoId) {
   const { content: rawContent, count: rawRedactions } = redactRollout(raw);
   redactionCount += rawRedactions;
   const firstUser = events.find((e) => e.role === "user");
+  const codexCommitEvidence = commitEvidenceFromToolCalls(lines);
   return {
     id: sessionId,
     agentKind: "codex",
     branch: null,
     parentSessionId: null,
     subagent: null,
-    // Codex records the remote itself, so identity survives a moved or deleted
-    // checkout. Falls back to the caller's derivation when it is absent.
-    repoId: repoIdFromMeta(meta) ?? repoId,
+    // The caller's derivation first: it reads the checkout's remote NOW,
+    // which is the identity every other ingest path agrees on. The remote
+    // recorded in session_meta is a historical fact — after the repo's
+    // GitHub transfer, month-old rollouts kept re-minting the dead
+    // jeffcheema id on every re-index, resurrecting a "third evrex-app"
+    // as fast as migrations folded it. Meta only serves a rollout whose
+    // checkout is gone, where a stale identity beats none.
+    repoId: repoId.startsWith("path:") ? repoIdFromMeta(meta) ?? repoId : repoId,
     cwd,
     startedAt: asString(meta.timestamp) ?? events[0]?.ts ?? null,
     endedAt: events[events.length - 1]?.ts ?? null,
     turnCount: turns.length,
-    committedSubjects: committedSubjectsFromToolCalls(lines),
+    committedSubjects: codexCommitEvidence.subjects,
+    committedShas: codexCommitEvidence.shas,
+    mergedPrNumbers: codexCommitEvidence.prNumbers,
     aiTitle: firstUser ? firstUser.text.slice(0, 120).trim() : null,
     author: null,
     sourceFile: filePath,
@@ -2962,6 +3160,7 @@ __export(src_exports, {
   changedSessions: () => changedSessions,
   codexSessionsDir: () => codexSessionsDir,
   collectRepoData: () => collectRepoData,
+  commitShaOutputs: () => commitShaOutputs,
   commitsOlderThan: () => commitsOlderThan,
   committedSubjects: () => committedSubjects,
   copilotSessionsDir: () => copilotSessionsDir,
@@ -2996,6 +3195,9 @@ __export(src_exports, {
   loadComposers: () => loadComposers,
   matchCommitToSession: () => matchCommitToSession,
   measureLineSurvival: () => measureLineSurvival,
+  mergeResponseShas: () => mergeResponseShas,
+  mergeSucceeded: () => mergeSucceeded,
+  mergedPrFromCommand: () => mergedPrFromCommand,
   normalizeRepoRemote: () => normalizeRepoRemote,
   opencodeDbPath: () => opencodeDbPath,
   originOf: () => originOf,
@@ -3005,6 +3207,7 @@ __export(src_exports, {
   parseAllOpenCodeSessions: () => parseAllOpenCodeSessions,
   parseAllSessions: () => parseAllSessions,
   parseCodexSessionFile: () => parseCodexSessionFile,
+  parseCommitsBySha: () => parseCommitsBySha,
   parseCopilotSessionDir: () => parseCopilotSessionDir,
   parseCopilotWorkspace: () => parseCopilotWorkspace,
   parseCursorComposer: () => parseCursorComposer,
@@ -3171,7 +3374,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.

package/dist/continuity.js

@@ -51,7 +51,7 @@ var evrexApi = {
   // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
   // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
   // which wants ranked hits fast, not a synthesized paragraph.
-  search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+  search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
   // Everything that happened in a repo, newest first, bounded by days — the
   // same query the desktop Timeline screen makes. Sessions and commits
   // interleaved, each with the handle evrex_expand takes.

package/dist/hook.js

@@ -49,7 +49,7 @@ var evrexApi = {
   // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
   // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
   // which wants ranked hits fast, not a synthesized paragraph.
-  search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+  search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
   // Everything that happened in a repo, newest first, bounded by days — the
   // same query the desktop Timeline screen makes. Sessions and commits
   // interleaved, each with the handle evrex_expand takes.
@@ -245,7 +245,10 @@ async function main() {
   if (!repoId) return;
   const remaining = DEADLINE_MS - (Date.now() - started);
   if (remaining < 500) return;
-  const result = await withDeadline(evrexApi.search(repoId, prompt), remaining);
+  const result = await withDeadline(
+    evrexApi.search(repoId, prompt, void 0, false),
+    remaining
+  );
   if (!result?.evidence?.length) return;
   const sessionId = input.session_id ?? "";
   const path = statePath();

package/dist/import.js

@@ -184,14 +184,34 @@ function commitMeta(repoPath, sha) {
     sha
   ]).trim();
   const parts = line.split(FIELD_SEP);
-  const message = git(repoPath, ["show", "-s", "--format=%B", sha]).replace(/\n+$/, "");
+  const body = git(repoPath, ["show", "-s", "--notes", `--format=%B${FIELD_SEP}%N`, sha]);
+  const sep2 = body.indexOf(FIELD_SEP);
+  let message = (sep2 === -1 ? body : body.slice(0, sep2)).replace(/\n+$/, "");
+  const note = sep2 === -1 ? "" : body.slice(sep2 + FIELD_SEP.length);
+  const noted = /^Evrex-Session:\s*(\S+)\s*$/m.exec(note);
+  if (noted && !/^Evrex-Session:/m.test(message)) {
+    message = `${message}
+
+${noted[0].trim()}`;
+  }
   return {
     sha: parts[0] ?? sha,
     author: parts[1] ?? "",
     authorEmail: parts[2] ?? "",
     ts: parts[3] ?? (/* @__PURE__ */ new Date()).toISOString(),
     parents: (parts[4] ?? "").split(" ").filter(Boolean),
-    message
+    message,
+    // The note's session, exposed on its own as well as folded above. When
+    // the message ALREADY carries a trailer, the fold declines (never rewrite
+    // what was stamped at commit time) — but the parse must still prefer the
+    // note: it exists only because an operator pressed "verify" on a commit
+    // whose message trailer names a session no ingest can resolve. Commit
+    // 843dabb3 is the standing example — its at-commit-time trailer names a
+    // session with no transcript anywhere, so the folding guard alone left it
+    // permanently un-verifiable: every promote re-ingested the dangling id,
+    // bestLink demoted it below the resolvable match, and the button offered
+    // the same commit again.
+    notedSession: noted?.[1] ?? null
   };
 }
 function parseTrailers(repoPath, message) {
@@ -312,6 +332,9 @@ function commitFiles(repoPath, sha) {
 }
 function parseGitLog(repoPath, repoId, known) {
   const shas = listShas(repoPath).filter((sha) => !known?.has(sha));
+  return parseCommitsBySha(repoPath, shas, repoId);
+}
+function parseCommitsBySha(repoPath, shas, repoId) {
   const id = repoId ?? deriveRepoId(repoPath);
   return shas.map((sha) => {
     const meta = commitMeta(repoPath, sha);
@@ -326,7 +349,9 @@ function parseGitLog(repoPath, repoId, known) {
       message: meta.message,
       parents: meta.parents,
       branch: commitBranch(repoPath, sha),
-      evrexSessionTrailer: trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
+      // A note is the operator's explicit later correction, so where both
+      // exist it names the session — see commitMeta's notedSession.
+      evrexSessionTrailer: meta.notedSession ?? trailerValue(trailers, EVREX_SESSION_TRAILER_KEY),
       origin: originOf(trailers),
       agentTrailers: agentTrailersOf(trailers),
       statedInsights: statedInsightsOf(trailers),
@@ -407,7 +432,7 @@ var PARSER_VERSION;
 var init_incremental = __esm({
   "../../packages/ingest-core/src/incremental.ts"() {
     "use strict";
-    PARSER_VERSION = 4;
+    PARSER_VERSION = 7;
   }
 });
 
@@ -522,11 +547,54 @@ var init_redact = __esm({
 
 // ../../packages/ingest-core/src/subject-linking.ts
 function matchCommitToSession(commit, sessions) {
+  const saw = sessions.filter(
+    (s) => (s.committedShas ?? []).some(
+      (short) => short.length >= 7 && commit.sha.startsWith(short)
+    )
+  );
+  if (saw.length > 0) {
+    const oneLineage2 = saw.every((s) => s.startedAt === saw[0].startedAt);
+    if (!oneLineage2) {
+      return { sha: commit.sha, sessionId: null, reason: "ambiguous" };
+    }
+    const original2 = [...saw].sort((a, b) => a.endedAt - b.endedAt)[0];
+    return {
+      sha: commit.sha,
+      sessionId: original2.sessionId,
+      evidence: "saw-the-sha",
+      confidence: SHA_CONFIDENCE
+    };
+  }
   const subject = commit.subject.trim();
+  const prMatch = /\(#(\d+)\)$/.exec(subject) ?? /^Merge pull request #(\d+)\b/.exec(subject);
+  if (prMatch) {
+    const pr = Number(prMatch[1]);
+    const mergedIt = sessions.filter(
+      (s) => (s.mergedPrNumbers ?? []).includes(pr)
+    );
+    if (mergedIt.length > 0) {
+      const oneLineage2 = mergedIt.every(
+        (s) => s.startedAt === mergedIt[0].startedAt
+      );
+      if (!oneLineage2) {
+        return { sha: commit.sha, sessionId: null, reason: "ambiguous" };
+      }
+      const original2 = [...mergedIt].sort((a, b) => a.endedAt - b.endedAt)[0];
+      return {
+        sha: commit.sha,
+        sessionId: original2.sessionId,
+        evidence: "merged-the-pr",
+        confidence: SHA_CONFIDENCE
+      };
+    }
+  }
   if (subject.length === 0) {
     return { sha: commit.sha, sessionId: null, reason: "absent" };
   }
-  const ran = sessions.filter((s) => s.committedSubjects.includes(subject));
+  const unsquashed = subject.replace(/\s+\(#\d+\)$/, "");
+  const ran = sessions.filter(
+    (s) => s.committedSubjects.includes(subject) || unsquashed !== subject && s.committedSubjects.includes(unsquashed)
+  );
   if (ran.length === 0) {
     return { sha: commit.sha, sessionId: null, reason: "absent" };
   }
@@ -580,15 +648,61 @@ function committedSubjects(command) {
     if (inline) {
       const subject = inline[2].split("\n")[0].trim();
       if (subject) out.push(subject);
+      continue;
+    }
+    const marginal = /^[^\n]*?-am\s*(['"])([\s\S]*?)\1/.exec(rest) ?? /^[^\n]*?--message[= ]\s*(['"])([\s\S]*?)\1/.exec(rest) ?? /^[^\n]*?-m(['"])([\s\S]*?)\1/.exec(rest);
+    if (marginal) {
+      const subject = marginal[2].split("\n")[0].trim();
+      if (subject) out.push(subject);
+      continue;
+    }
+    const fromFile = /^[^\n]*?(?:-F|--file)[= ]\s*(\S+)/.exec(rest);
+    if (fromFile) {
+      const file = fromFile[1].replace(/^["']|["']$/g, "");
+      const escaped = file.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+      const wrote = new RegExp(
+        `(?:cat\\s*>>?|tee\\s+(?:-a\\s+)?)\\s*["']?${escaped}["']?[^\\n]*<<-?\\s*(['"]?)(\\w+)\\1[^\\n]*\\r?\\n([^\\n]*)`
+      ).exec(command.slice(0, found.index));
+      if (wrote) {
+        const subject = wrote[3].trim();
+        if (subject) out.push(subject);
+      }
     }
   }
   return out;
 }
-var INFERRED_CONFIDENCE;
+function mergeResponseShas(output) {
+  const out = /* @__PURE__ */ new Set();
+  const line = /"sha"\s*:\s*"([0-9a-f]{40})"/g;
+  for (let m = line.exec(output); m !== null; m = line.exec(output)) {
+    out.add(m[1]);
+  }
+  return [...out];
+}
+function commitShaOutputs(output) {
+  const out = /* @__PURE__ */ new Set();
+  const line = /\[[\w./+-]+ (?:\(root-commit\) )?([0-9a-f]{7,12})\]/g;
+  for (let m = line.exec(output); m !== null; m = line.exec(output)) {
+    out.add(m[1]);
+  }
+  return [...out];
+}
+function mergedPrFromCommand(command) {
+  const api = /\/pulls\/(\d+)\/merge\b/.exec(command);
+  if (api) return Number(api[1]);
+  const gh = /\bgh\s+pr\s+merge\s+(?:.*?\/pull\/(\d+)\b|(\d+)\b)/.exec(command);
+  if (gh) return Number(gh[1] ?? gh[2]);
+  return null;
+}
+function mergeSucceeded(output) {
+  return /"merged"\s*:\s*true/.test(output) || /(Squashed and |Rebased and )?[Mm]erged pull request #\d+/.test(output);
+}
+var INFERRED_CONFIDENCE, SHA_CONFIDENCE;
 var init_subject_linking = __esm({
   "../../packages/ingest-core/src/subject-linking.ts"() {
     "use strict";
     INFERRED_CONFIDENCE = 0.9;
+    SHA_CONFIDENCE = 0.95;
   }
 });
 
@@ -890,6 +1004,7 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
   const subagent = agentId ? { agentId, agentType: meta?.agentType ?? null, description: meta?.description ?? null } : null;
   if (agentId && !aiTitle && meta?.description) aiTitle = meta.description;
   const sortedTs = turns.map((t) => t.ts).sort();
+  const commitEvidence = collectCommitEvidence(lines);
   const rawContent = capped ? "" : lines.map((line) => {
     try {
       return JSON.stringify(redactJsonValue(JSON.parse(line)).value);
@@ -913,7 +1028,9 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     // filled in by the caller — see getGitUserName in git-history.ts
     sourceFile: filePath,
     redactionCount: totalRedactions,
-    committedSubjects: collectCommittedSubjects(lines),
+    committedSubjects: commitEvidence.subjects,
+    committedShas: commitEvidence.shas,
+    mergedPrNumbers: commitEvidence.prNumbers,
     branch,
     parentSessionId,
     subagent,
@@ -922,10 +1039,14 @@ function parseSessionFile(filePath, repoPath, repoId = deriveRepoId(repoPath), a
     turns
   };
 }
-function collectCommittedSubjects(lines) {
+function collectCommitEvidence(lines) {
   const subjects = /* @__PURE__ */ new Set();
+  const shas = /* @__PURE__ */ new Set();
+  const prNumbers = /* @__PURE__ */ new Set();
+  const commitCallIds = /* @__PURE__ */ new Set();
+  const mergeCallIds = /* @__PURE__ */ new Map();
   for (const line of lines) {
-    if (!line.includes("git commit")) continue;
+    if (!line.includes("git commit") && !line.includes("/merge") && !line.includes("tool_use_id")) continue;
     let record;
     try {
       record = JSON.parse(line);
@@ -935,13 +1056,40 @@ function collectCommittedSubjects(lines) {
     const content = record?.message?.content;
     if (!Array.isArray(content)) continue;
     for (const block of content) {
-      if (block?.type !== "tool_use") continue;
-      const command = block?.input?.command;
-      if (typeof command !== "string") continue;
-      for (const subject of committedSubjects(command)) subjects.add(subject);
+      if (block?.type === "tool_use") {
+        const command = block?.input?.command;
+        if (typeof command !== "string") continue;
+        const found = committedSubjects(command);
+        for (const subject of found) subjects.add(subject);
+        if (/\bgit\s+(?:-C\s+\S+\s+)?commit\b/.test(command) && typeof block?.id === "string") {
+          commitCallIds.add(block.id);
+        }
+        if ((/\/pulls\/\d+\/merge\b/.test(command) || /\bgh\s+pr\s+merge\b/.test(command)) && typeof block?.id === "string") {
+          mergeCallIds.set(block.id, mergedPrFromCommand(command));
+        }
+        continue;
+      }
+      if (block?.type === "tool_result" && mergeCallIds.has(block?.tool_use_id)) {
+        const raw = block?.content;
+        const text = typeof raw === "string" ? raw : Array.isArray(raw) ? raw.map(
+          (x) => typeof x?.text === "string" ? x.text : ""
+        ).join(" ") : "";
+        for (const sha of mergeResponseShas(text)) shas.add(sha);
+        if (mergeSucceeded(text)) {
+          const pr = mergeCallIds.get(block?.tool_use_id);
+          if (typeof pr === "number") prNumbers.add(pr);
+        }
+      }
+      if (block?.type === "tool_result" && commitCallIds.has(block?.tool_use_id)) {
+        const raw = block?.content;
+        const text = typeof raw === "string" ? raw : Array.isArray(raw) ? raw.map(
+          (x) => typeof x?.text === "string" ? x.text : ""
+        ).join(" ") : "";
+        for (const sha of commitShaOutputs(text)) shas.add(sha);
+      }
     }
   }
-  return [...subjects];
+  return { subjects: [...subjects], shas: [...shas], prNumbers: [...prNumbers] };
 }
 function parseAllSessions(repoPath, repoId, cursors = {}) {
   const id = repoId ?? deriveRepoId(repoPath);
@@ -1907,16 +2055,58 @@ function addUsage(into, next) {
     model: into.model ?? next.model
   };
 }
-function committedSubjectsFromToolCalls(lines) {
+function commitEvidenceFromToolCalls(lines) {
   const subjects = /* @__PURE__ */ new Set();
+  const shas = /* @__PURE__ */ new Set();
+  const prNumbers = /* @__PURE__ */ new Set();
+  const commitCallIds = /* @__PURE__ */ new Set();
+  const mergeCallIds = /* @__PURE__ */ new Map();
   for (const line of lines) {
     if (line.type !== "response_item") continue;
-    if (payloadType(line) !== "custom_tool_call") continue;
-    const input = asString(line.payload?.input);
-    if (!input) continue;
-    for (const subject of committedSubjects(input)) subjects.add(subject);
+    const kind = payloadType(line);
+    if (kind === "custom_tool_call") {
+      const input = asString(line.payload?.input);
+      if (!input) continue;
+      const cmd = execCmdFrom(input) ?? input;
+      const callId = asString(line.payload?.call_id);
+      for (const subject of committedSubjects(cmd)) subjects.add(subject);
+      if (/\bgit\s+(?:-C\s+\S+\s+)?commit\b/.test(cmd) && callId) {
+        commitCallIds.add(callId);
+      }
+      if ((/\/pulls\/\d+\/merge\b/.test(cmd) || /\bgh\s+pr\s+merge\b/.test(cmd)) && callId) {
+        mergeCallIds.set(callId, mergedPrFromCommand(cmd));
+      }
+      continue;
+    }
+    if (kind === "custom_tool_call_output") {
+      const callId = asString(line.payload?.call_id);
+      if (!callId) continue;
+      const raw = line.payload?.output;
+      const text = Array.isArray(raw) ? raw.map(
+        (x) => typeof x?.text === "string" ? x.text : ""
+      ).join(" ") : asString(raw) ?? "";
+      if (commitCallIds.has(callId)) {
+        for (const sha of commitShaOutputs(text)) shas.add(sha);
+      }
+      if (mergeCallIds.has(callId)) {
+        for (const sha of mergeResponseShas(text)) shas.add(sha);
+        if (mergeSucceeded(text)) {
+          const pr = mergeCallIds.get(callId);
+          if (typeof pr === "number") prNumbers.add(pr);
+        }
+      }
+    }
+  }
+  return { subjects: [...subjects], shas: [...shas], prNumbers: [...prNumbers] };
+}
+function execCmdFrom(input) {
+  const m = /cmd\s*:\s*"((?:[^"\\]|\\.)*)"/.exec(input);
+  if (!m) return null;
+  try {
+    return JSON.parse(`"${m[1]}"`);
+  } catch {
+    return null;
   }
-  return [...subjects];
 }
 function filesFromToolCalls(lines) {
   const byPath = /* @__PURE__ */ new Map();
@@ -1975,20 +2165,28 @@ function parseCodexSessionFile(filePath, repoPath, repoId) {
   const { content: rawContent, count: rawRedactions } = redactRollout(raw);
   redactionCount += rawRedactions;
   const firstUser = events.find((e) => e.role === "user");
+  const codexCommitEvidence = commitEvidenceFromToolCalls(lines);
   return {
     id: sessionId,
     agentKind: "codex",
     branch: null,
     parentSessionId: null,
     subagent: null,
-    // Codex records the remote itself, so identity survives a moved or deleted
-    // checkout. Falls back to the caller's derivation when it is absent.
-    repoId: repoIdFromMeta(meta) ?? repoId,
+    // The caller's derivation first: it reads the checkout's remote NOW,
+    // which is the identity every other ingest path agrees on. The remote
+    // recorded in session_meta is a historical fact — after the repo's
+    // GitHub transfer, month-old rollouts kept re-minting the dead
+    // jeffcheema id on every re-index, resurrecting a "third evrex-app"
+    // as fast as migrations folded it. Meta only serves a rollout whose
+    // checkout is gone, where a stale identity beats none.
+    repoId: repoId.startsWith("path:") ? repoIdFromMeta(meta) ?? repoId : repoId,
     cwd,
     startedAt: asString(meta.timestamp) ?? events[0]?.ts ?? null,
     endedAt: events[events.length - 1]?.ts ?? null,
     turnCount: turns.length,
-    committedSubjects: committedSubjectsFromToolCalls(lines),
+    committedSubjects: codexCommitEvidence.subjects,
+    committedShas: codexCommitEvidence.shas,
+    mergedPrNumbers: codexCommitEvidence.prNumbers,
     aiTitle: firstUser ? firstUser.text.slice(0, 120).trim() : null,
     author: null,
     sourceFile: filePath,
@@ -2962,6 +3160,7 @@ __export(src_exports, {
   changedSessions: () => changedSessions,
   codexSessionsDir: () => codexSessionsDir,
   collectRepoData: () => collectRepoData,
+  commitShaOutputs: () => commitShaOutputs,
   commitsOlderThan: () => commitsOlderThan,
   committedSubjects: () => committedSubjects,
   copilotSessionsDir: () => copilotSessionsDir,
@@ -2996,6 +3195,9 @@ __export(src_exports, {
   loadComposers: () => loadComposers,
   matchCommitToSession: () => matchCommitToSession,
   measureLineSurvival: () => measureLineSurvival,
+  mergeResponseShas: () => mergeResponseShas,
+  mergeSucceeded: () => mergeSucceeded,
+  mergedPrFromCommand: () => mergedPrFromCommand,
   normalizeRepoRemote: () => normalizeRepoRemote,
   opencodeDbPath: () => opencodeDbPath,
   originOf: () => originOf,
@@ -3005,6 +3207,7 @@ __export(src_exports, {
   parseAllOpenCodeSessions: () => parseAllOpenCodeSessions,
   parseAllSessions: () => parseAllSessions,
   parseCodexSessionFile: () => parseCodexSessionFile,
+  parseCommitsBySha: () => parseCommitsBySha,
   parseCopilotSessionDir: () => parseCopilotSessionDir,
   parseCopilotWorkspace: () => parseCopilotWorkspace,
   parseCursorComposer: () => parseCursorComposer,
@@ -3171,7 +3374,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.

package/dist/index.js

@@ -63,7 +63,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.
@@ -525,6 +525,29 @@ async function synthesizeAnswer(question, evidence, client2, options = {}) {
   };
 }
 
+// ../../packages/llm-core/src/retrieval-boost.ts
+var CORPUS = `The corpus is an engineering-memory index built from a team's coding-agent conversations and git history, retrieved with Postgres full-text OR-matching plus BM25 (exact word stems match; meanings do not). Searchable documents:
+- Conversation turns: prose by the person and the agent, plus tool-call lines like "[tool_call: Edit] apps/backend/src/query/query.service.ts" \u2014 so file paths, function names, error strings and identifiers appear verbatim.
+- Extracted insights: short titled records of decisions made, approaches REJECTED (with the reason), constraints, and procedures.
+- Commit messages: conventional-commit subjects and explanatory bodies.`;
+var EXPAND_SYSTEM = `You expand a search query into terms that will match an engineering record.
+
+${CORPUS}
+
+Given a person's question, produce the search terms most likely to appear VERBATIM in that corpus but absent from the question itself: the identifiers an engineer would have typed (function, constant, env var, file names), the error text the failure would have printed, and the words the same idea takes in commit-message register ("fix the flaky test" \u2192 "flaky", "deflake", "retry", "timeout"). Prefer specific over general \u2014 "RLS" and "row level security" over "database permissions". Never invent identifiers you are not confident exist in this kind of codebase; a wrong specific term matches nothing and costs nothing, but fill the list with your best evidence-seeking guesses.
+
+Respond with JSON only: {"terms": ["...", ...]} \u2014 3 to 8 terms, each 1 to 4 words, no duplicates of the question's own words.`;
+var RERANK_SYSTEM = `You order search results for an engineer's question about their own project's history.
+
+${CORPUS}
+
+You are given the question and a numbered list of candidate results that lexical retrieval fetched. Order the candidate ids best-first by how much each would actually help ANSWER the question:
+- Recorded reasoning outranks mechanical record: a decision, a rejected approach with its reason, a constraint, or an explanation beats a log line or a file listing that merely shares words with the question.
+- A candidate about the question's subject beats one that only quotes its vocabulary; watch for results that repeat the question's words while being about something else entirely.
+- Prefer the origin of an idea over a later mention of it; among near-duplicates prefer the more complete one; use the date only to break ties.
+
+Respond with JSON only: {"ranked": ["id", ...]} \u2014 every listed id at most once, ids from the list only, best first. Include every id you consider relevant; omit only what would waste the reader's attention.`;
+
 // ../../packages/llm-core/src/cli-client.ts
 import { execFile } from "node:child_process";
 var TIMEOUT_MS = 12e4;

package/dist/pretool.js

@@ -49,7 +49,7 @@ var evrexApi = {
   // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
   // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
   // which wants ranked hits fast, not a synthesized paragraph.
-  search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+  search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
   // Everything that happened in a repo, newest first, bounded by days — the
   // same query the desktop Timeline screen makes. Sessions and commits
   // interleaved, each with the handle evrex_expand takes.

package/dist/tickets.js

@@ -63,7 +63,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "evrex-mcp",
-  "version": "0.8.1",
+  "version": "0.8.2",
   "description": "MCP server that gives coding agents the recorded reasoning behind a repo: prior decisions, hard constraints, and approaches already rejected.",
   "keywords": [
     "mcp",