pullfrog 0.1.102 → 0.1.104

package/dist/cli.mjs

@@ -112919,7 +112919,7 @@ var import_picocolors = __toESM(require_picocolors(), 1);
 // package.json
 var package_default = {
   name: "pullfrog",
-  version: "0.1.102",
+  version: "0.1.104",
   cliContractVersion: "1",
   type: "module",
   bin: {
@@ -117052,7 +117052,7 @@ var providers2 = {
       },
       // Zen's FREE open-weights preview (listed 2026-10-05), the eighth promo.
       // Zen's docs: no training, 30-day retention for safety. hidden until a
-      // published release carries the alias.
+      // published release carries the alias; 0.1.102 is the first.
       beam: {
         displayName: "Beam",
         description: "Preview, free for a limited time",
@@ -117060,8 +117060,7 @@ var providers2 = {
         effort: ["low", "medium", "high", "xhigh", "max"],
         // free to run, still gated on the provider's own OPENCODE_API_KEY —
         // see the big-pickle note above (#1077).
-        isFree: true,
-        hidden: true
+        isFree: true
       },
       // Zen's live free MiMo, and the second free row in a menu that big-pickle
       // was alone in since `mimo-v2-pro-free` lost its model.
@@ -117434,10 +117433,12 @@ var providers2 = {
         openRouterResolve: "openrouter/openai/gpt-5.4",
         fallback: "openrouter/gpt-sol"
       },
+      // OpenRouter dropped `openai/o4-mini` on 2026-10-28 (only `:batch` remains) — folds stored pins to Luna.
       "o4-mini": {
         displayName: "O4 Mini",
         resolve: "openrouter/openai/o4-mini",
-        openRouterResolve: "openrouter/openai/o4-mini"
+        openRouterResolve: "openrouter/openai/o4-mini",
+        fallback: "openrouter/gpt-luna"
       },
       "gemini-pro": {
         displayName: "Gemini Pro",
@@ -119564,7 +119565,7 @@ ${PR_SUMMARY_FORMAT}`
 
 2. **checkout**: call \`${t2("checkout_pr")}\` \u2014 this returns PR metadata, \`diffPath\` (full diff), \`incrementalDiffPath\` (changes since last reviewed version, if available), and a supplemental \`impactPath\` when change-impact extraction is enabled.
 
-3. **incremental scope**: if \`incrementalDiffPath\` is present, read it FIRST to see what changed since the last review. this is a range-diff that isolates the net changes, filtering out base branch noise. then read the authoritative full diff end-to-end, beginning with the TOC and using its line ranges as your coverage checklist. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after establishing that authoritative scope and completing raw-diff coverage, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
+3. **incremental scope**: if \`incrementalDiffPath\` is present, it is your coverage checklist: read it FIRST and in full. it is a range-diff of the net changes since the last review, free of base-branch noise. read only the \`diffPath\` ranges the delta needs for context \u2014 do not re-read the whole PR. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after that, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
 
 4. **prior feedback \u2014 read AND retire it**: fetch previous reviews via \`${t2("list_pull_request_reviews")}\`, then call \`${t2("get_review_comments")}\` on each prior Pullfrog review. Each thread renders as a section whose first line is a fenced tag \`comment author=<login> id=<fullDatabaseId> review=<reviewId> thread=<graphqlId>\`; section headers carry \`[RESOLVED]\` / \`[OUTDATED]\` when relevant. For every **open, Pullfrog-originated** thread, decide and act:
 
@@ -119580,13 +119581,13 @@ ${PR_SUMMARY_FORMAT}`
 
    skip the deeper pass and jump to step 10's non-substantive path (do NOT submit a review) only when the incremental changes have **no behavioral surface at all** \u2014 formatting, comment tweaks, import reordering, lockfile regen, a mechanical rename of import paths. line count is not the signal: a one-line change to auth, money, SQL, a comparison operator, a redirect, or a config default is not trivial.
 
-6. **specialist decision**: after covering the incremental and full diffs, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
+6. **specialist decision**: after covering the delta, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
 
    **There is NO one-specialist cap or fixed maximum.** cover every orthogonal question that remains; do not collapse several real questions into one broad prompt just to reduce the count. there is no file-count, line-count, or budget threshold either \u2014 diff size is not a proxy for review uncertainty.
 
    frame each question through the lens that primes the right failure modes. for high-stakes subsystems, lead with the **domain** ("the billing lens", "the auth lens", "the schema-migration lens") rather than the generic equivalent \u2014 the domain framing makes the subagent recall failure modes a generic lens misses.
 
-   you remain the synthesizer: reading the complete raw full diff plus the incremental diff, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
+   you remain the synthesizer: reading the delta and the context it needs, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
 
 7. **dispatch specialists (only if step 6 found unresolved questions)**: for 2+ questions, emit every Task tool_use block **IN A SINGLE ASSISTANT TURN** before reading any result, so the investigations run in parallel rather than serially. your own \`read\` / \`grep\` / \`webfetch\` calls can ride in that same turn.
 
@@ -192058,7 +192059,9 @@ function parseDiffTocEntries(params) {
   const lines = params.toc.split("\n");
   const entries = [];
   for (const line of lines) {
-    const match3 = line.match(/^- (.+) (?:→|->) lines (\d+)-(\d+)(?: · diff-[0-9a-f]+)?$/);
+    const match3 = line.match(
+      /^- (.+) (?:→|->) lines (\d+)-(\d+)(?: · diff-[0-9a-f]+)?(?: · generated)?$/
+    );
     if (!match3) continue;
     const startLine = Number.parseInt(match3[2], 10);
     const endLine = Number.parseInt(match3[3], 10);
@@ -192067,11 +192070,24 @@ function parseDiffTocEntries(params) {
   }
   return entries;
 }
+function parseIncrementalDiffEntries(params) {
+  const lines = params.content.split("\n");
+  const entries = [];
+  for (const [index, line] of lines.entries()) {
+    const match3 = line.match(/^@@ (.+)$|^[ +-]## (.+) ##$/);
+    if (!match3) continue;
+    const previous = entries.at(-1);
+    if (previous) previous.endLine = index;
+    const filename = (match3[1] ?? match3[2]).replace(/(: | \().*$/, "");
+    entries.push({ filename, startLine: index + 1, endLine: lines.length });
+  }
+  return entries;
+}
 function createDiffCoverageState(params) {
   return {
     diffPath: params.diffPath,
     totalLines: params.totalLines,
-    tocEntries: parseDiffTocEntries({ toc: params.toc }),
+    tocEntries: params.tocEntries,
     coveredRanges: [],
     // carry forward across checkout_pr refreshes so the nudge stays "once per
     // review session". coveredRanges are intentionally not carried because
@@ -192286,6 +192302,9 @@ function readNumber(params) {
   return void 0;
 }
 
+// utils/generatedFiles.ts
+var GENERATED = /(^|\/)(package-lock\.json|pnpm-lock\.yaml|yarn\.lock|bun\.lockb?|Cargo\.lock|go\.sum|poetry\.lock|Pipfile\.lock|Gemfile\.lock|composer\.lock|.*\.snap|.*\.min\.(js|css)|.*\.map|.*\.lock|.*\.(pb\.go|_pb2\.py|g\.dart)|.*\.(generated|gen)\.[^/]+|.*\.(svg|png|jpe?g|gif|ico|webp|woff2?|ttf|pdf))$|(^|\/)(dist|generated|__generated__|node_modules|\.yarn|__snapshots__|coverage|\.next|\.turbo|drizzle\/meta)\//i;
+
 // utils/gitAuth.ts
 import { execSync } from "node:child_process";
 import { createHash as createHash2 } from "node:crypto";
@@ -193628,14 +193647,18 @@ async function runGitLines(args2) {
 // utils/rangeDiff.ts
 function computeIncrementalDiff(params) {
   try {
+    const tips = [`origin/${params.baseBranch}`, ...params.formerBaseTips];
+    const oldBase = versionBase({ sha: params.beforeSha, tips });
+    const newBase = versionBase({ sha: params.headSha, tips });
     const raw2 = $(
       "sh",
       [
         "-c",
-        'old_base=$(git merge-base "$1" "origin/$2") && new_base=$(git merge-base "$3" "origin/$2") && git range-diff --no-color "$old_base..$(git commit-tree "$1^{tree}" -p "$old_base" -m x)" "$new_base..$(git commit-tree "$3^{tree}" -p "$new_base" -m x)"',
+        'git range-diff --no-color --creation-factor=1000 "$1..$(git commit-tree "$2^{tree}" -p "$1" -m x)" "$3..$(git commit-tree "$4^{tree}" -p "$3" -m x)"',
         "--",
+        oldBase,
         params.beforeSha,
-        params.baseBranch,
+        newBase,
         params.headSha
       ],
       { log: false }
@@ -193648,6 +193671,27 @@ function computeIncrementalDiff(params) {
     return { status: "unavailable", reason };
   }
 }
+function git(args2) {
+  return $("git", args2, { log: false }).trim();
+}
+function versionBase(params) {
+  let best;
+  for (const tip of params.tips) {
+    let sha;
+    try {
+      sha = git(["merge-base", params.sha, tip]);
+    } catch {
+      continue;
+    }
+    const commits = Number(git(["rev-list", "--count", `${sha}..${params.sha}`]));
+    if (!best || commits < best.commits) best = { tip, sha, commits };
+  }
+  if (!best) throw new Error(`no merge base between ${params.sha} and the base branch`);
+  if (best.tip !== params.tips[0]) {
+    log2.info(`\xBB incremental diff: ${params.sha.slice(0, 7)} sits on former base tip ${best.tip}`);
+  }
+  return best.sha;
+}
 function isDiffPrefix(ch) {
   return ch === " " || ch === "+" || ch === "-";
 }
@@ -193655,6 +193699,8 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
   if (!raw2.trim()) return null;
   if (/^\d+:\s+\w+\s+=\s+\d+:/m.test(raw2)) return null;
   const beforeBuf = [];
+  let lastOuterHdr = null;
+  let outerHdrEmitted = true;
   let lastFileHdr = null;
   let lastHunkHdr = null;
   let fileHdrEmitted = true;
@@ -193668,17 +193714,33 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
     if (lastEmittedSeq >= 0 && line.seq > lastEmittedSeq + 1) out += (out ? "\n" : "") + "...";
     out += (out ? "\n" : "") + line.prefix + raw2.slice(line.from, line.to);
     lastEmittedSeq = line.seq;
+    if (lastOuterHdr?.seq === line.seq) outerHdrEmitted = true;
     if (lastFileHdr?.seq === line.seq) fileHdrEmitted = true;
     if (lastHunkHdr?.seq === line.seq) hunkHdrEmitted = true;
   }
   function flushBefore() {
-    if (lastFileHdr && !fileHdrEmitted) emit2(lastFileHdr);
-    if (lastHunkHdr && !hunkHdrEmitted) emit2(lastHunkHdr);
-    for (const line of beforeBuf) {
-      if (line.seq > lastEmittedSeq) emit2(line);
+    const headers = [
+      outerHdrEmitted ? null : lastOuterHdr,
+      fileHdrEmitted ? null : lastFileHdr,
+      hunkHdrEmitted ? null : lastHunkHdr
+    ].filter((line) => line !== null);
+    const pending = [...headers, ...beforeBuf.filter((line) => line.seq > lastEmittedSeq)].sort(
+      (a2, b) => a2.seq - b.seq
+    );
+    for (const [i2, line] of pending.entries()) {
+      if (pending[i2 - 1]?.seq !== line.seq) emit2(line);
     }
     beforeBuf.length = 0;
   }
+  function isFileHeader(pos) {
+    return raw2.startsWith("## ", pos) && !raw2.startsWith("## Commit message", pos);
+  }
+  function markFileHeader(line, emitted) {
+    lastFileHdr = line;
+    fileHdrEmitted = emitted;
+    lastHunkHdr = null;
+    hunkHdrEmitted = true;
+  }
   let cursor3 = 0;
   while (cursor3 < raw2.length) {
     const eol = raw2.indexOf("\n", cursor3);
@@ -193687,7 +193749,18 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
       cursor3 = lineEnd + 1;
       continue;
     }
-    if (lineEnd - cursor3 >= 5 && raw2.startsWith("    ", cursor3)) {
+    if (raw2.startsWith("    @@", cursor3)) {
+      const label = raw2.slice(cursor3 + 4, lineEnd);
+      lastOuterHdr = label === "@@ Metadata" || label === "@@ Commit message" ? null : { prefix: "", from: cursor3 + 4, to: lineEnd, seq };
+      outerHdrEmitted = false;
+      lastFileHdr = null;
+      fileHdrEmitted = true;
+      lastHunkHdr = null;
+      hunkHdrEmitted = true;
+      beforeBuf.length = 0;
+      afterRemaining = 0;
+      seq++;
+    } else if (lineEnd - cursor3 >= 5 && raw2.startsWith("    ", cursor3)) {
       const prefix = raw2[cursor3 + 4];
       if (isDiffPrefix(prefix)) {
         const contentPos = cursor3 + 5;
@@ -193699,13 +193772,13 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
         } else if (isDiffPrefix(raw2[contentPos])) {
           isChange = isOuterChange;
           line = { prefix, from: contentPos + 1, to: lineEnd, seq };
+          if (raw2[contentPos] === " " && isFileHeader(contentPos + 1)) {
+            markFileHeader(line, isChange);
+          }
         } else {
           line = { prefix, from: contentPos, to: lineEnd, seq };
-          if (raw2.startsWith("## ", contentPos) && !raw2.startsWith("## Commit message", contentPos)) {
-            lastFileHdr = line;
-            fileHdrEmitted = false;
-            lastHunkHdr = null;
-            hunkHdrEmitted = true;
+          if (isFileHeader(contentPos)) {
+            markFileHeader(line, false);
           } else if (raw2.startsWith("@@", contentPos) && !raw2.startsWith("@@ Metadata", contentPos)) {
             lastHunkHdr = line;
             hunkHdrEmitted = false;
@@ -199698,6 +199771,7 @@ var STRING_KEYS = [
   "prNodeId",
   "issueNodeId",
   "reviewNodeId",
+  "reviewedCommitSha",
   "planCommentNodeId",
   "summarySnapshot",
   "model",
@@ -200862,7 +200936,10 @@ async function createAndSubmitWithFooter(ctx, params, opts) {
   }
 }
 async function reportReviewNodeId(ctx, params) {
-  await patchWorkflowRunFields(ctx, { reviewNodeId: params.nodeId });
+  await patchWorkflowRunFields(ctx, {
+    reviewNodeId: params.nodeId,
+    reviewedCommitSha: params.reviewedSha
+  });
 }
 
 // utils/setup.ts
@@ -201132,8 +201209,9 @@ function formatFilesWithLineNumbers(files) {
   const tocLines = [`## Files (${files.length})`];
   for (const entry of tocEntries) {
     const anchor2 = createHash3("sha256").update(entry.filename).digest("hex");
+    const generated = GENERATED.test(entry.filename) ? " \xB7 generated" : "";
     tocLines.push(
-      `- ${entry.filename} \u2192 lines ${entry.startLine}-${entry.endLine} \xB7 diff-${anchor2}`
+      `- ${entry.filename} \u2192 lines ${entry.startLine}-${entry.endLine} \xB7 diff-${anchor2}${generated}`
     );
   }
   tocLines.push("");
@@ -201455,6 +201533,34 @@ function describeHead(h2) {
   if (h2.kind === "branch") return `branch \`${h2.name}\``;
   return `detached HEAD \`${h2.sha}\``;
 }
+var FORMER_BASE_TIPS_QUERY = `
+  query ($owner: String!, $repo: String!, $number: Int!) {
+    repository(owner: $owner, name: $repo) {
+      pullRequest(number: $number) {
+        timelineItems(itemTypes: [BASE_REF_FORCE_PUSHED_EVENT], last: 100) {
+          nodes { ... on BaseRefForcePushedEvent { beforeCommit { oid } } }
+        }
+      }
+    }
+  }
+`;
+async function getFormerBaseTips(ctx, pullNumber) {
+  try {
+    const response = await ctx.octokit.graphql(FORMER_BASE_TIPS_QUERY, {
+      owner: ctx.repo.owner,
+      repo: ctx.repo.name,
+      number: pullNumber
+    });
+    return response.repository.pullRequest.timelineItems.nodes.flatMap(
+      (node3) => node3.beforeCommit ? [node3.beforeCommit.oid] : []
+    );
+  } catch (error63) {
+    log2.warning(
+      `\xBB former base tips unavailable: ${error63 instanceof Error ? error63.message : String(error63)}`
+    );
+    return [];
+  }
+}
 function CheckoutPrTool(ctx) {
   const runCheckout = async (pull_number, signal) => {
     const prResponse = await ctx.octokit.rest.pulls.get({
@@ -201498,6 +201604,8 @@ function CheckoutPrTool(ctx) {
     if (!checkoutSha) throw new Error("checkout completed without a resolved HEAD SHA");
     const headShort = checkoutSha.slice(0, 7);
     let incrementalDiffPath;
+    let incrementalDiff;
+    let incrementalEmpty = false;
     let incrementalUnavailable = false;
     let incrementalUnavailableReason;
     if (primary.beforeSha) {
@@ -201505,7 +201613,8 @@ function CheckoutPrTool(ctx) {
       const incremental = computeIncrementalDiff({
         baseBranch: pr.baseRef,
         beforeSha: primary.beforeSha,
-        headSha: checkoutSha
+        headSha: checkoutSha,
+        formerBaseTips: await getFormerBaseTips(ctx, pull_number)
       });
       if (incremental.status === "ok") {
         incrementalDiffPath = join17(
@@ -201513,9 +201622,12 @@ function CheckoutPrTool(ctx) {
           `pr-${pull_number}-${beforeShort}-${headShort}-incremental.diff`
         );
         writeFileSync11(incrementalDiffPath, incremental.diff);
+        incrementalDiff = incremental.diff;
         log2.info(
           `\xBB incremental diff computed (${incremental.diff.length} bytes) \u2192 ${incrementalDiffPath}`
         );
+      } else if (incremental.status === "empty") {
+        incrementalEmpty = true;
       } else if (incremental.status === "unavailable") {
         incrementalUnavailable = true;
         incrementalUnavailableReason = incremental.reason;
@@ -201528,14 +201640,16 @@ ${diffPreview}`);
     const diffPath = join17(tempDir, `pr-${pull_number}-${headShort}.diff`);
     writeFileSync11(diffPath, formatResult.content);
     log2.debug(`wrote diff to ${diffPath} (${formatResult.content.length} bytes)`);
-    primary.diffCoverage = createDiffCoverageState({
-      diffPath,
-      totalLines: countLines({ content: formatResult.content }),
-      toc: formatResult.toc,
+    const coverageEntries = incrementalDiff ? parseIncrementalDiffEntries({ content: incrementalDiff }) : parseDiffTocEntries({ toc: formatResult.toc });
+    primary.diffCoverage = incrementalEmpty ? void 0 : createDiffCoverageState({
+      diffPath: incrementalDiffPath ?? diffPath,
+      totalLines: countLines({ content: incrementalDiff ?? formatResult.content }),
+      tocEntries: coverageEntries.filter((entry) => !GENERATED.test(entry.filename)),
       previous: primary.diffCoverage
     });
+    const coverage = primary.diffCoverage;
     log2.debug(
-      `\xBB diff coverage initialized: diffPath=${diffPath}, totalLines=${primary.diffCoverage.totalLines}, tocEntries=${primary.diffCoverage.tocEntries.length}`
+      coverage ? `\xBB diff coverage initialized: diffPath=${coverage.diffPath}, totalLines=${coverage.totalLines}, tocEntries=${coverage.tocEntries.length}` : "\xBB diff coverage skipped: no changes since the last reviewed version"
     );
     let impactPath;
     if (process.env.PULLFROG_DISABLE_CHANGE_IMPACT === "1") {
@@ -201586,7 +201700,7 @@ ${diffPreview}`);
     primary.commentableLinesByFile = cached4;
     primary.commentableLinesPullNumber = pull_number;
     primary.commentableLinesCheckoutSha = checkoutSha;
-    const incrementalInstructions = incrementalDiffPath ? `IMPORTANT: incrementalDiffPath contains ONLY the changes since the last reviewed version (computed via range-diff). you MUST read incrementalDiffPath FIRST to understand what changed, then use diffPath for full PR context. do NOT skip the incremental diff. ` : incrementalUnavailable ? (
+    const incrementalInstructions = incrementalDiffPath ? `IMPORTANT: incrementalDiffPath contains ONLY the changes since the last reviewed version (computed via range-diff), and it \u2014 not diffPath's TOC \u2014 is this review's coverage checklist. read it FIRST and in full, then read only the diffPath ranges the delta needs for context. ` : incrementalEmpty ? `NOTE: the PR's net changes are identical to the last reviewed version (e.g. a rebase with no content change), so there is no new diff to cover. ` : incrementalUnavailable ? (
       // say it FAILED rather than leaving the agent to infer "nothing changed"
       // from an absent field. only suggest diff-tree when the previous
       // revision is actually PRESENT locally — if the fetch never landed it,
@@ -201637,7 +201751,7 @@ ${diffPreview}`);
       commitLogTruncated,
       commitLogUnavailable,
       hookWarning: checkoutResult.hookWarning,
-      instructions: incrementalInstructions + `the diff file at diffPath contains a table of contents (TOC) at the top listing every changed file with its line range. use the TOC line ranges as your checklist and read specific files from the diff instead of reading the entire file. for example, if the TOC says "src/foo.ts \u2192 lines 5-42", read lines 5-42 from diffPath to see that file's changes. review files selectively based on relevance rather than reading everything sequentially. to inspect the PR's changed files, use diffPath \u2014 do NOT run \`git diff\` to re-derive what's already in diffPath. the formatted diff with line numbers is authoritative. if you ever do need a branch-vs-base diff via the git tool, use \`git diff --merge-base <base>\` (single call, includes uncommitted edits) or three-dot \`git diff <base>...HEAD\` (committed-only). bare \`<base>\` and two-dot \`<base>..HEAD\` are symmetric and pull in the inverse of every commit landed on \`<base>\` since the branch forked \u2014 the git tool will reject those forms when the divergence is detected. \`$(...)\` subshells are NOT expanded by the git tool. \`git log\` and \`git diff --stat\` are fine for commit-range overview, and \`git diff\` / \`git diff --cached\` are fine for inspecting *your own* uncommitted changes \u2014 but PR review content MUST come from diffPath. before your review is submitted, a one-time coverage pre-flight may error listing unread TOC regions. retry the same create_pull_request_review call to proceed \u2014 optionally after reading the listed ranges. the pre-flight will not block again this session. the local branch is 'localBranch' (pr-{number}), not the remote branch name. when pushing, omit branchName to use the current branch. do not use remoteBranch as a local branch name.` + impactInstructions + hookWarningInstructions + commitLogInstructions
+      instructions: incrementalInstructions + `the diff file at diffPath contains a table of contents (TOC) at the top listing every changed file with its line range. absent incrementalDiffPath, use the TOC line ranges as your checklist (entries marked \`generated\` are off it) and read specific files from the diff instead of reading the entire file. for example, if the TOC says "src/foo.ts \u2192 lines 5-42", read lines 5-42 from diffPath to see that file's changes. review files selectively based on relevance rather than reading everything sequentially. to inspect the PR's changed files, use diffPath \u2014 do NOT run \`git diff\` to re-derive what's already in diffPath. the formatted diff with line numbers is authoritative. if you ever do need a branch-vs-base diff via the git tool, use \`git diff --merge-base <base>\` (single call, includes uncommitted edits) or three-dot \`git diff <base>...HEAD\` (committed-only). bare \`<base>\` and two-dot \`<base>..HEAD\` are symmetric and pull in the inverse of every commit landed on \`<base>\` since the branch forked \u2014 the git tool will reject those forms when the divergence is detected. \`$(...)\` subshells are NOT expanded by the git tool. \`git log\` and \`git diff --stat\` are fine for commit-range overview, and \`git diff\` / \`git diff --cached\` are fine for inspecting *your own* uncommitted changes \u2014 but PR review content MUST come from diffPath. before your review is submitted, a one-time coverage pre-flight may error listing unread TOC regions. retry the same create_pull_request_review call to proceed \u2014 optionally after reading the listed ranges. the pre-flight will not block again this session. the local branch is 'localBranch' (pr-{number}), not the remote branch name. when pushing, omit branchName to use the current branch. do not use remoteBranch as a local branch name.` + impactInstructions + hookWarningInstructions + commitLogInstructions
     };
   };
   return tool({
@@ -208095,7 +208209,7 @@ async function postReviewCleanup(ctx) {
   if (!review) {
     const checkoutSha = primaryRepoState(ctx.toolState).checkoutSha;
     if (ctx.toolState.selectedMode === "IncrementalReview" && checkoutSha) {
-      await patchWorkflowRunFields(ctx, { reviewSkipped: true });
+      await patchWorkflowRunFields(ctx, { reviewSkipped: true, reviewedCommitSha: checkoutSha });
       await bestEffort(
         () => dispatchFollowUpReReview(ctx, checkoutSha),
         "follow-up re-review dispatch"
@@ -208104,7 +208218,7 @@ async function postReviewCleanup(ctx) {
     return;
   }
   delete ctx.toolState.review;
-  await bestEffort(() => reportReviewNodeId(ctx, { nodeId: review.nodeId }), "reportReviewNodeId");
+  await bestEffort(() => reportReviewNodeId(ctx, review), "reportReviewNodeId");
   if (review.reviewedSha) {
     await bestEffort(
       () => dispatchFollowUpReReview(ctx, review.reviewedSha),
@@ -209874,7 +209988,7 @@ function resolveTarget2(repo) {
   return { owner: match3[1], repo: match3[2] };
 }
 function cliVersion() {
-  const parts = /^(\d+)\.(\d+)\.(\d+)$/.exec("0.1.102");
+  const parts = /^(\d+)\.(\d+)\.(\d+)$/.exec("0.1.104");
   if (!parts) return "0.0.0";
   return `${Number(parts[1])}.${Number(parts[2])}.${Number(parts[3])}`;
 }
@@ -210041,7 +210155,7 @@ async function runCli8(input2) {
 }
 
 // cli.ts
-var VERSION10 = "0.1.102";
+var VERSION10 = "0.1.104";
 var bin = basename2(process.argv[1] || "");
 var PROG = bin === "pf" || bin === "pullfrog" ? bin : "pullfrog";
 var rawArgs = process.argv.slice(2);

package/dist/index.js

@@ -109128,7 +109128,7 @@ var providers = {
       },
       // Zen's FREE open-weights preview (listed 2026-10-05), the eighth promo.
       // Zen's docs: no training, 30-day retention for safety. hidden until a
-      // published release carries the alias.
+      // published release carries the alias; 0.1.102 is the first.
       beam: {
         displayName: "Beam",
         description: "Preview, free for a limited time",
@@ -109136,8 +109136,7 @@ var providers = {
         effort: ["low", "medium", "high", "xhigh", "max"],
         // free to run, still gated on the provider's own OPENCODE_API_KEY —
         // see the big-pickle note above (#1077).
-        isFree: true,
-        hidden: true
+        isFree: true
       },
       // Zen's live free MiMo, and the second free row in a menu that big-pickle
       // was alone in since `mimo-v2-pro-free` lost its model.
@@ -109510,10 +109509,12 @@ var providers = {
         openRouterResolve: "openrouter/openai/gpt-5.4",
         fallback: "openrouter/gpt-sol"
       },
+      // OpenRouter dropped `openai/o4-mini` on 2026-10-28 (only `:batch` remains) — folds stored pins to Luna.
       "o4-mini": {
         displayName: "O4 Mini",
         resolve: "openrouter/openai/o4-mini",
-        openRouterResolve: "openrouter/openai/o4-mini"
+        openRouterResolve: "openrouter/openai/o4-mini",
+        fallback: "openrouter/gpt-luna"
       },
       "gemini-pro": {
         displayName: "Gemini Pro",
@@ -112140,7 +112141,7 @@ var import_semver = __toESM(require_semver2(), 1);
 // package.json
 var package_default = {
   name: "pullfrog",
-  version: "0.1.102",
+  version: "0.1.104",
   cliContractVersion: "1",
   type: "module",
   bin: {
@@ -113089,7 +113090,7 @@ ${PR_SUMMARY_FORMAT}`
 
 2. **checkout**: call \`${t("checkout_pr")}\` \u2014 this returns PR metadata, \`diffPath\` (full diff), \`incrementalDiffPath\` (changes since last reviewed version, if available), and a supplemental \`impactPath\` when change-impact extraction is enabled.
 
-3. **incremental scope**: if \`incrementalDiffPath\` is present, read it FIRST to see what changed since the last review. this is a range-diff that isolates the net changes, filtering out base branch noise. then read the authoritative full diff end-to-end, beginning with the TOC and using its line ranges as your coverage checklist. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after establishing that authoritative scope and completing raw-diff coverage, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
+3. **incremental scope**: if \`incrementalDiffPath\` is present, it is your coverage checklist: read it FIRST and in full. it is a range-diff of the net changes since the last review, free of base-branch noise. read only the \`diffPath\` ranges the delta needs for context \u2014 do not re-read the whole PR. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after that, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
 
 4. **prior feedback \u2014 read AND retire it**: fetch previous reviews via \`${t("list_pull_request_reviews")}\`, then call \`${t("get_review_comments")}\` on each prior Pullfrog review. Each thread renders as a section whose first line is a fenced tag \`comment author=<login> id=<fullDatabaseId> review=<reviewId> thread=<graphqlId>\`; section headers carry \`[RESOLVED]\` / \`[OUTDATED]\` when relevant. For every **open, Pullfrog-originated** thread, decide and act:
 
@@ -113105,13 +113106,13 @@ ${PR_SUMMARY_FORMAT}`
 
    skip the deeper pass and jump to step 10's non-substantive path (do NOT submit a review) only when the incremental changes have **no behavioral surface at all** \u2014 formatting, comment tweaks, import reordering, lockfile regen, a mechanical rename of import paths. line count is not the signal: a one-line change to auth, money, SQL, a comparison operator, a redirect, or a config default is not trivial.
 
-6. **specialist decision**: after covering the incremental and full diffs, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
+6. **specialist decision**: after covering the delta, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
 
    **There is NO one-specialist cap or fixed maximum.** cover every orthogonal question that remains; do not collapse several real questions into one broad prompt just to reduce the count. there is no file-count, line-count, or budget threshold either \u2014 diff size is not a proxy for review uncertainty.
 
    frame each question through the lens that primes the right failure modes. for high-stakes subsystems, lead with the **domain** ("the billing lens", "the auth lens", "the schema-migration lens") rather than the generic equivalent \u2014 the domain framing makes the subagent recall failure modes a generic lens misses.
 
-   you remain the synthesizer: reading the complete raw full diff plus the incremental diff, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
+   you remain the synthesizer: reading the delta and the context it needs, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
 
 7. **dispatch specialists (only if step 6 found unresolved questions)**: for 2+ questions, emit every Task tool_use block **IN A SINGLE ASSISTANT TURN** before reading any result, so the investigations run in parallel rather than serially. your own \`read\` / \`grep\` / \`webfetch\` calls can ride in that same turn.
 
@@ -188540,7 +188541,9 @@ function parseDiffTocEntries(params) {
   const lines = params.toc.split("\n");
   const entries = [];
   for (const line of lines) {
-    const match3 = line.match(/^- (.+) (?:→|->) lines (\d+)-(\d+)(?: · diff-[0-9a-f]+)?$/);
+    const match3 = line.match(
+      /^- (.+) (?:→|->) lines (\d+)-(\d+)(?: · diff-[0-9a-f]+)?(?: · generated)?$/
+    );
     if (!match3) continue;
     const startLine = Number.parseInt(match3[2], 10);
     const endLine = Number.parseInt(match3[3], 10);
@@ -188549,11 +188552,24 @@ function parseDiffTocEntries(params) {
   }
   return entries;
 }
+function parseIncrementalDiffEntries(params) {
+  const lines = params.content.split("\n");
+  const entries = [];
+  for (const [index, line] of lines.entries()) {
+    const match3 = line.match(/^@@ (.+)$|^[ +-]## (.+) ##$/);
+    if (!match3) continue;
+    const previous = entries.at(-1);
+    if (previous) previous.endLine = index;
+    const filename = (match3[1] ?? match3[2]).replace(/(: | \().*$/, "");
+    entries.push({ filename, startLine: index + 1, endLine: lines.length });
+  }
+  return entries;
+}
 function createDiffCoverageState(params) {
   return {
     diffPath: params.diffPath,
     totalLines: params.totalLines,
-    tocEntries: parseDiffTocEntries({ toc: params.toc }),
+    tocEntries: params.tocEntries,
     coveredRanges: [],
     // carry forward across checkout_pr refreshes so the nudge stays "once per
     // review session". coveredRanges are intentionally not carried because
@@ -188768,6 +188784,9 @@ function readNumber(params) {
   return void 0;
 }
 
+// utils/generatedFiles.ts
+var GENERATED = /(^|\/)(package-lock\.json|pnpm-lock\.yaml|yarn\.lock|bun\.lockb?|Cargo\.lock|go\.sum|poetry\.lock|Pipfile\.lock|Gemfile\.lock|composer\.lock|.*\.snap|.*\.min\.(js|css)|.*\.map|.*\.lock|.*\.(pb\.go|_pb2\.py|g\.dart)|.*\.(generated|gen)\.[^/]+|.*\.(svg|png|jpe?g|gif|ico|webp|woff2?|ttf|pdf))$|(^|\/)(dist|generated|__generated__|node_modules|\.yarn|__snapshots__|coverage|\.next|\.turbo|drizzle\/meta)\//i;
+
 // utils/gitAuth.ts
 import { execSync } from "node:child_process";
 import { createHash } from "node:crypto";
@@ -190110,14 +190129,18 @@ async function runGitLines(args2) {
 // utils/rangeDiff.ts
 function computeIncrementalDiff(params) {
   try {
+    const tips = [`origin/${params.baseBranch}`, ...params.formerBaseTips];
+    const oldBase = versionBase({ sha: params.beforeSha, tips });
+    const newBase = versionBase({ sha: params.headSha, tips });
     const raw2 = $(
       "sh",
       [
         "-c",
-        'old_base=$(git merge-base "$1" "origin/$2") && new_base=$(git merge-base "$3" "origin/$2") && git range-diff --no-color "$old_base..$(git commit-tree "$1^{tree}" -p "$old_base" -m x)" "$new_base..$(git commit-tree "$3^{tree}" -p "$new_base" -m x)"',
+        'git range-diff --no-color --creation-factor=1000 "$1..$(git commit-tree "$2^{tree}" -p "$1" -m x)" "$3..$(git commit-tree "$4^{tree}" -p "$3" -m x)"',
         "--",
+        oldBase,
         params.beforeSha,
-        params.baseBranch,
+        newBase,
         params.headSha
       ],
       { log: false }
@@ -190130,6 +190153,27 @@ function computeIncrementalDiff(params) {
     return { status: "unavailable", reason };
   }
 }
+function git(args2) {
+  return $("git", args2, { log: false }).trim();
+}
+function versionBase(params) {
+  let best;
+  for (const tip of params.tips) {
+    let sha;
+    try {
+      sha = git(["merge-base", params.sha, tip]);
+    } catch {
+      continue;
+    }
+    const commits = Number(git(["rev-list", "--count", `${sha}..${params.sha}`]));
+    if (!best || commits < best.commits) best = { tip, sha, commits };
+  }
+  if (!best) throw new Error(`no merge base between ${params.sha} and the base branch`);
+  if (best.tip !== params.tips[0]) {
+    log.info(`\xBB incremental diff: ${params.sha.slice(0, 7)} sits on former base tip ${best.tip}`);
+  }
+  return best.sha;
+}
 function isDiffPrefix(ch) {
   return ch === " " || ch === "+" || ch === "-";
 }
@@ -190137,6 +190181,8 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
   if (!raw2.trim()) return null;
   if (/^\d+:\s+\w+\s+=\s+\d+:/m.test(raw2)) return null;
   const beforeBuf = [];
+  let lastOuterHdr = null;
+  let outerHdrEmitted = true;
   let lastFileHdr = null;
   let lastHunkHdr = null;
   let fileHdrEmitted = true;
@@ -190150,17 +190196,33 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
     if (lastEmittedSeq >= 0 && line.seq > lastEmittedSeq + 1) out += (out ? "\n" : "") + "...";
     out += (out ? "\n" : "") + line.prefix + raw2.slice(line.from, line.to);
     lastEmittedSeq = line.seq;
+    if (lastOuterHdr?.seq === line.seq) outerHdrEmitted = true;
     if (lastFileHdr?.seq === line.seq) fileHdrEmitted = true;
     if (lastHunkHdr?.seq === line.seq) hunkHdrEmitted = true;
   }
   function flushBefore() {
-    if (lastFileHdr && !fileHdrEmitted) emit(lastFileHdr);
-    if (lastHunkHdr && !hunkHdrEmitted) emit(lastHunkHdr);
-    for (const line of beforeBuf) {
-      if (line.seq > lastEmittedSeq) emit(line);
+    const headers = [
+      outerHdrEmitted ? null : lastOuterHdr,
+      fileHdrEmitted ? null : lastFileHdr,
+      hunkHdrEmitted ? null : lastHunkHdr
+    ].filter((line) => line !== null);
+    const pending = [...headers, ...beforeBuf.filter((line) => line.seq > lastEmittedSeq)].sort(
+      (a, b) => a.seq - b.seq
+    );
+    for (const [i, line] of pending.entries()) {
+      if (pending[i - 1]?.seq !== line.seq) emit(line);
     }
     beforeBuf.length = 0;
   }
+  function isFileHeader(pos) {
+    return raw2.startsWith("## ", pos) && !raw2.startsWith("## Commit message", pos);
+  }
+  function markFileHeader(line, emitted) {
+    lastFileHdr = line;
+    fileHdrEmitted = emitted;
+    lastHunkHdr = null;
+    hunkHdrEmitted = true;
+  }
   let cursor = 0;
   while (cursor < raw2.length) {
     const eol = raw2.indexOf("\n", cursor);
@@ -190169,7 +190231,18 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
       cursor = lineEnd + 1;
       continue;
     }
-    if (lineEnd - cursor >= 5 && raw2.startsWith("    ", cursor)) {
+    if (raw2.startsWith("    @@", cursor)) {
+      const label = raw2.slice(cursor + 4, lineEnd);
+      lastOuterHdr = label === "@@ Metadata" || label === "@@ Commit message" ? null : { prefix: "", from: cursor + 4, to: lineEnd, seq };
+      outerHdrEmitted = false;
+      lastFileHdr = null;
+      fileHdrEmitted = true;
+      lastHunkHdr = null;
+      hunkHdrEmitted = true;
+      beforeBuf.length = 0;
+      afterRemaining = 0;
+      seq++;
+    } else if (lineEnd - cursor >= 5 && raw2.startsWith("    ", cursor)) {
       const prefix = raw2[cursor + 4];
       if (isDiffPrefix(prefix)) {
         const contentPos = cursor + 5;
@@ -190181,13 +190254,13 @@ function postProcessRangeDiff(raw2, contextLines = 3) {
         } else if (isDiffPrefix(raw2[contentPos])) {
           isChange = isOuterChange;
           line = { prefix, from: contentPos + 1, to: lineEnd, seq };
+          if (raw2[contentPos] === " " && isFileHeader(contentPos + 1)) {
+            markFileHeader(line, isChange);
+          }
         } else {
           line = { prefix, from: contentPos, to: lineEnd, seq };
-          if (raw2.startsWith("## ", contentPos) && !raw2.startsWith("## Commit message", contentPos)) {
-            lastFileHdr = line;
-            fileHdrEmitted = false;
-            lastHunkHdr = null;
-            hunkHdrEmitted = true;
+          if (isFileHeader(contentPos)) {
+            markFileHeader(line, false);
           } else if (raw2.startsWith("@@", contentPos) && !raw2.startsWith("@@ Metadata", contentPos)) {
             lastHunkHdr = line;
             hunkHdrEmitted = false;
@@ -196180,6 +196253,7 @@ var STRING_KEYS = [
   "prNodeId",
   "issueNodeId",
   "reviewNodeId",
+  "reviewedCommitSha",
   "planCommentNodeId",
   "summarySnapshot",
   "model",
@@ -197344,7 +197418,10 @@ async function createAndSubmitWithFooter(ctx, params, opts) {
   }
 }
 async function reportReviewNodeId(ctx, params) {
-  await patchWorkflowRunFields(ctx, { reviewNodeId: params.nodeId });
+  await patchWorkflowRunFields(ctx, {
+    reviewNodeId: params.nodeId,
+    reviewedCommitSha: params.reviewedSha
+  });
 }
 
 // utils/setup.ts
@@ -197614,8 +197691,9 @@ function formatFilesWithLineNumbers(files) {
   const tocLines = [`## Files (${files.length})`];
   for (const entry of tocEntries) {
     const anchor2 = createHash2("sha256").update(entry.filename).digest("hex");
+    const generated = GENERATED.test(entry.filename) ? " \xB7 generated" : "";
     tocLines.push(
-      `- ${entry.filename} \u2192 lines ${entry.startLine}-${entry.endLine} \xB7 diff-${anchor2}`
+      `- ${entry.filename} \u2192 lines ${entry.startLine}-${entry.endLine} \xB7 diff-${anchor2}${generated}`
     );
   }
   tocLines.push("");
@@ -197937,6 +198015,34 @@ function describeHead(h) {
   if (h.kind === "branch") return `branch \`${h.name}\``;
   return `detached HEAD \`${h.sha}\``;
 }
+var FORMER_BASE_TIPS_QUERY = `
+  query ($owner: String!, $repo: String!, $number: Int!) {
+    repository(owner: $owner, name: $repo) {
+      pullRequest(number: $number) {
+        timelineItems(itemTypes: [BASE_REF_FORCE_PUSHED_EVENT], last: 100) {
+          nodes { ... on BaseRefForcePushedEvent { beforeCommit { oid } } }
+        }
+      }
+    }
+  }
+`;
+async function getFormerBaseTips(ctx, pullNumber) {
+  try {
+    const response = await ctx.octokit.graphql(FORMER_BASE_TIPS_QUERY, {
+      owner: ctx.repo.owner,
+      repo: ctx.repo.name,
+      number: pullNumber
+    });
+    return response.repository.pullRequest.timelineItems.nodes.flatMap(
+      (node3) => node3.beforeCommit ? [node3.beforeCommit.oid] : []
+    );
+  } catch (error63) {
+    log.warning(
+      `\xBB former base tips unavailable: ${error63 instanceof Error ? error63.message : String(error63)}`
+    );
+    return [];
+  }
+}
 function CheckoutPrTool(ctx) {
   const runCheckout = async (pull_number, signal) => {
     const prResponse = await ctx.octokit.rest.pulls.get({
@@ -197980,6 +198086,8 @@ function CheckoutPrTool(ctx) {
     if (!checkoutSha) throw new Error("checkout completed without a resolved HEAD SHA");
     const headShort = checkoutSha.slice(0, 7);
     let incrementalDiffPath;
+    let incrementalDiff;
+    let incrementalEmpty = false;
     let incrementalUnavailable = false;
     let incrementalUnavailableReason;
     if (primary.beforeSha) {
@@ -197987,7 +198095,8 @@ function CheckoutPrTool(ctx) {
       const incremental = computeIncrementalDiff({
         baseBranch: pr.baseRef,
         beforeSha: primary.beforeSha,
-        headSha: checkoutSha
+        headSha: checkoutSha,
+        formerBaseTips: await getFormerBaseTips(ctx, pull_number)
       });
       if (incremental.status === "ok") {
         incrementalDiffPath = join16(
@@ -197995,9 +198104,12 @@ function CheckoutPrTool(ctx) {
           `pr-${pull_number}-${beforeShort}-${headShort}-incremental.diff`
         );
         writeFileSync10(incrementalDiffPath, incremental.diff);
+        incrementalDiff = incremental.diff;
         log.info(
           `\xBB incremental diff computed (${incremental.diff.length} bytes) \u2192 ${incrementalDiffPath}`
         );
+      } else if (incremental.status === "empty") {
+        incrementalEmpty = true;
       } else if (incremental.status === "unavailable") {
         incrementalUnavailable = true;
         incrementalUnavailableReason = incremental.reason;
@@ -198010,14 +198122,16 @@ ${diffPreview}`);
     const diffPath = join16(tempDir, `pr-${pull_number}-${headShort}.diff`);
     writeFileSync10(diffPath, formatResult.content);
     log.debug(`wrote diff to ${diffPath} (${formatResult.content.length} bytes)`);
-    primary.diffCoverage = createDiffCoverageState({
-      diffPath,
-      totalLines: countLines({ content: formatResult.content }),
-      toc: formatResult.toc,
+    const coverageEntries = incrementalDiff ? parseIncrementalDiffEntries({ content: incrementalDiff }) : parseDiffTocEntries({ toc: formatResult.toc });
+    primary.diffCoverage = incrementalEmpty ? void 0 : createDiffCoverageState({
+      diffPath: incrementalDiffPath ?? diffPath,
+      totalLines: countLines({ content: incrementalDiff ?? formatResult.content }),
+      tocEntries: coverageEntries.filter((entry) => !GENERATED.test(entry.filename)),
       previous: primary.diffCoverage
     });
+    const coverage = primary.diffCoverage;
     log.debug(
-      `\xBB diff coverage initialized: diffPath=${diffPath}, totalLines=${primary.diffCoverage.totalLines}, tocEntries=${primary.diffCoverage.tocEntries.length}`
+      coverage ? `\xBB diff coverage initialized: diffPath=${coverage.diffPath}, totalLines=${coverage.totalLines}, tocEntries=${coverage.tocEntries.length}` : "\xBB diff coverage skipped: no changes since the last reviewed version"
     );
     let impactPath;
     if (process.env.PULLFROG_DISABLE_CHANGE_IMPACT === "1") {
@@ -198068,7 +198182,7 @@ ${diffPreview}`);
     primary.commentableLinesByFile = cached4;
     primary.commentableLinesPullNumber = pull_number;
     primary.commentableLinesCheckoutSha = checkoutSha;
-    const incrementalInstructions = incrementalDiffPath ? `IMPORTANT: incrementalDiffPath contains ONLY the changes since the last reviewed version (computed via range-diff). you MUST read incrementalDiffPath FIRST to understand what changed, then use diffPath for full PR context. do NOT skip the incremental diff. ` : incrementalUnavailable ? (
+    const incrementalInstructions = incrementalDiffPath ? `IMPORTANT: incrementalDiffPath contains ONLY the changes since the last reviewed version (computed via range-diff), and it \u2014 not diffPath's TOC \u2014 is this review's coverage checklist. read it FIRST and in full, then read only the diffPath ranges the delta needs for context. ` : incrementalEmpty ? `NOTE: the PR's net changes are identical to the last reviewed version (e.g. a rebase with no content change), so there is no new diff to cover. ` : incrementalUnavailable ? (
       // say it FAILED rather than leaving the agent to infer "nothing changed"
       // from an absent field. only suggest diff-tree when the previous
       // revision is actually PRESENT locally — if the fetch never landed it,
@@ -198119,7 +198233,7 @@ ${diffPreview}`);
       commitLogTruncated,
       commitLogUnavailable,
       hookWarning: checkoutResult.hookWarning,
-      instructions: incrementalInstructions + `the diff file at diffPath contains a table of contents (TOC) at the top listing every changed file with its line range. use the TOC line ranges as your checklist and read specific files from the diff instead of reading the entire file. for example, if the TOC says "src/foo.ts \u2192 lines 5-42", read lines 5-42 from diffPath to see that file's changes. review files selectively based on relevance rather than reading everything sequentially. to inspect the PR's changed files, use diffPath \u2014 do NOT run \`git diff\` to re-derive what's already in diffPath. the formatted diff with line numbers is authoritative. if you ever do need a branch-vs-base diff via the git tool, use \`git diff --merge-base <base>\` (single call, includes uncommitted edits) or three-dot \`git diff <base>...HEAD\` (committed-only). bare \`<base>\` and two-dot \`<base>..HEAD\` are symmetric and pull in the inverse of every commit landed on \`<base>\` since the branch forked \u2014 the git tool will reject those forms when the divergence is detected. \`$(...)\` subshells are NOT expanded by the git tool. \`git log\` and \`git diff --stat\` are fine for commit-range overview, and \`git diff\` / \`git diff --cached\` are fine for inspecting *your own* uncommitted changes \u2014 but PR review content MUST come from diffPath. before your review is submitted, a one-time coverage pre-flight may error listing unread TOC regions. retry the same create_pull_request_review call to proceed \u2014 optionally after reading the listed ranges. the pre-flight will not block again this session. the local branch is 'localBranch' (pr-{number}), not the remote branch name. when pushing, omit branchName to use the current branch. do not use remoteBranch as a local branch name.` + impactInstructions + hookWarningInstructions + commitLogInstructions
+      instructions: incrementalInstructions + `the diff file at diffPath contains a table of contents (TOC) at the top listing every changed file with its line range. absent incrementalDiffPath, use the TOC line ranges as your checklist (entries marked \`generated\` are off it) and read specific files from the diff instead of reading the entire file. for example, if the TOC says "src/foo.ts \u2192 lines 5-42", read lines 5-42 from diffPath to see that file's changes. review files selectively based on relevance rather than reading everything sequentially. to inspect the PR's changed files, use diffPath \u2014 do NOT run \`git diff\` to re-derive what's already in diffPath. the formatted diff with line numbers is authoritative. if you ever do need a branch-vs-base diff via the git tool, use \`git diff --merge-base <base>\` (single call, includes uncommitted edits) or three-dot \`git diff <base>...HEAD\` (committed-only). bare \`<base>\` and two-dot \`<base>..HEAD\` are symmetric and pull in the inverse of every commit landed on \`<base>\` since the branch forked \u2014 the git tool will reject those forms when the divergence is detected. \`$(...)\` subshells are NOT expanded by the git tool. \`git log\` and \`git diff --stat\` are fine for commit-range overview, and \`git diff\` / \`git diff --cached\` are fine for inspecting *your own* uncommitted changes \u2014 but PR review content MUST come from diffPath. before your review is submitted, a one-time coverage pre-flight may error listing unread TOC regions. retry the same create_pull_request_review call to proceed \u2014 optionally after reading the listed ranges. the pre-flight will not block again this session. the local branch is 'localBranch' (pr-{number}), not the remote branch name. when pushing, omit branchName to use the current branch. do not use remoteBranch as a local branch name.` + impactInstructions + hookWarningInstructions + commitLogInstructions
     };
   };
   return tool({
@@ -204620,7 +204734,7 @@ async function postReviewCleanup(ctx) {
   if (!review) {
     const checkoutSha = primaryRepoState(ctx.toolState).checkoutSha;
     if (ctx.toolState.selectedMode === "IncrementalReview" && checkoutSha) {
-      await patchWorkflowRunFields(ctx, { reviewSkipped: true });
+      await patchWorkflowRunFields(ctx, { reviewSkipped: true, reviewedCommitSha: checkoutSha });
       await bestEffort(
         () => dispatchFollowUpReReview(ctx, checkoutSha),
         "follow-up re-review dispatch"
@@ -204629,7 +204743,7 @@ async function postReviewCleanup(ctx) {
     return;
   }
   delete ctx.toolState.review;
-  await bestEffort(() => reportReviewNodeId(ctx, { nodeId: review.nodeId }), "reportReviewNodeId");
+  await bestEffort(() => reportReviewNodeId(ctx, review), "reportReviewNodeId");
   if (review.reviewedSha) {
     await bestEffort(
       () => dispatchFollowUpReReview(ctx, review.reviewedSha),

package/dist/internal/index.d.ts

@@ -17,6 +17,7 @@ export { decodeJwtExpMs, OAuthInvalidGrantError, parseCodexAuthBody, refreshCode
 export type { CredentialVerdict } from "../utils/credentialCheck.ts";
 export { isCredentialProbeable, verifyCredential } from "../utils/credentialCheck.ts";
 export { renderFailure } from "../utils/failureBody.ts";
+export { GENERATED } from "../utils/generatedFiles.ts";
 export type { ResourceUsage, UsageSummary } from "../utils/github.ts";
 export { isPullfrog } from "../utils/isPullfrog.ts";
 export { isLeapingIntoActionCommentBody, isLiveChecklistCommentBody, LEAPING_INTO_ACTION_PREFIX, } from "../utils/leapingComment.ts";

package/dist/internal.js

@@ -8,7 +8,7 @@ var __export = (target, all) => {
 // package.json
 var package_default = {
   name: "pullfrog",
-  version: "0.1.102",
+  version: "0.1.104",
   cliContractVersion: "1",
   type: "module",
   bin: {
@@ -21026,7 +21026,7 @@ var providers = {
       },
       // Zen's FREE open-weights preview (listed 2026-10-05), the eighth promo.
       // Zen's docs: no training, 30-day retention for safety. hidden until a
-      // published release carries the alias.
+      // published release carries the alias; 0.1.102 is the first.
       beam: {
         displayName: "Beam",
         description: "Preview, free for a limited time",
@@ -21034,8 +21034,7 @@ var providers = {
         effort: ["low", "medium", "high", "xhigh", "max"],
         // free to run, still gated on the provider's own OPENCODE_API_KEY —
         // see the big-pickle note above (#1077).
-        isFree: true,
-        hidden: true
+        isFree: true
       },
       // Zen's live free MiMo, and the second free row in a menu that big-pickle
       // was alone in since `mimo-v2-pro-free` lost its model.
@@ -21408,10 +21407,12 @@ var providers = {
         openRouterResolve: "openrouter/openai/gpt-5.4",
         fallback: "openrouter/gpt-sol"
       },
+      // OpenRouter dropped `openai/o4-mini` on 2026-10-28 (only `:batch` remains) — folds stored pins to Luna.
       "o4-mini": {
         displayName: "O4 Mini",
         resolve: "openrouter/openai/o4-mini",
-        openRouterResolve: "openrouter/openai/o4-mini"
+        openRouterResolve: "openrouter/openai/o4-mini",
+        fallback: "openrouter/gpt-luna"
       },
       "gemini-pro": {
         displayName: "Gemini Pro",
@@ -22133,7 +22134,7 @@ ${PR_SUMMARY_FORMAT}`
 
 2. **checkout**: call \`${t("checkout_pr")}\` \u2014 this returns PR metadata, \`diffPath\` (full diff), \`incrementalDiffPath\` (changes since last reviewed version, if available), and a supplemental \`impactPath\` when change-impact extraction is enabled.
 
-3. **incremental scope**: if \`incrementalDiffPath\` is present, read it FIRST to see what changed since the last review. this is a range-diff that isolates the net changes, filtering out base branch noise. then read the authoritative full diff end-to-end, beginning with the TOC and using its line ranges as your coverage checklist. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after establishing that authoritative scope and completing raw-diff coverage, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
+3. **incremental scope**: if \`incrementalDiffPath\` is present, it is your coverage checklist: read it FIRST and in full. it is a range-diff of the net changes since the last review, free of base-branch noise. read only the \`diffPath\` ranges the delta needs for context \u2014 do not re-read the whole PR. if no incremental diff is present, start with the full-diff TOC, determine what changed since Pullfrog's most recent review, and complete the raw-diff read. only after that, use \`impactPath\` as an explicitly incomplete list of reference leads; it never replaces raw-diff reading or establishes coverage.
 
 4. **prior feedback \u2014 read AND retire it**: fetch previous reviews via \`${t("list_pull_request_reviews")}\`, then call \`${t("get_review_comments")}\` on each prior Pullfrog review. Each thread renders as a section whose first line is a fenced tag \`comment author=<login> id=<fullDatabaseId> review=<reviewId> thread=<graphqlId>\`; section headers carry \`[RESOLVED]\` / \`[OUTDATED]\` when relevant. For every **open, Pullfrog-originated** thread, decide and act:
 
@@ -22149,13 +22150,13 @@ ${PR_SUMMARY_FORMAT}`
 
    skip the deeper pass and jump to step 10's non-substantive path (do NOT submit a review) only when the incremental changes have **no behavioral surface at all** \u2014 formatting, comment tweaks, import reordering, lockfile regen, a mechanical rename of import paths. line count is not the signal: a one-line change to auth, money, SQL, a comparison operator, a redirect, or a config default is not trivial.
 
-6. **specialist decision**: after covering the incremental and full diffs, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
+6. **specialist decision**: after covering the delta, name the questions about the new changes that you still cannot answer confidently yourself, and dispatch one \`${REVIEWER_AGENT_NAME}\` specialist per question. a question qualifies only when a specialist could return evidence that **changes your disposition** on the PR \u2014 generic requests for another look, extra confidence, or polish do not. most incremental reviews need zero or one, especially thread-reply re-reviews; some need several.
 
    **There is NO one-specialist cap or fixed maximum.** cover every orthogonal question that remains; do not collapse several real questions into one broad prompt just to reduce the count. there is no file-count, line-count, or budget threshold either \u2014 diff size is not a proxy for review uncertainty.
 
    frame each question through the lens that primes the right failure modes. for high-stakes subsystems, lead with the **domain** ("the billing lens", "the auth lens", "the schema-migration lens") rather than the generic equivalent \u2014 the domain framing makes the subagent recall failure modes a generic lens misses.
 
-   you remain the synthesizer: reading the complete raw full diff plus the incremental diff, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
+   you remain the synthesizer: reading the delta and the context it needs, investigating surrounding code, validating every returned finding, and writing the review are yours. specialist reads supplement that work; they never satisfy your own coverage obligation.
 
 7. **dispatch specialists (only if step 6 found unresolved questions)**: for 2+ questions, emit every Task tool_use block **IN A SINGLE ASSISTANT TURN** before reading any result, so the investigations run in parallel rather than serially. your own \`read\` / \`grep\` / \`webfetch\` calls can ride in that same turn.
 
@@ -22633,6 +22634,9 @@ async function verifyCredential(params) {
   }
 }
 
+// utils/generatedFiles.ts
+var GENERATED = /(^|\/)(package-lock\.json|pnpm-lock\.yaml|yarn\.lock|bun\.lockb?|Cargo\.lock|go\.sum|poetry\.lock|Pipfile\.lock|Gemfile\.lock|composer\.lock|.*\.snap|.*\.min\.(js|css)|.*\.map|.*\.lock|.*\.(pb\.go|_pb2\.py|g\.dart)|.*\.(generated|gen)\.[^/]+|.*\.(svg|png|jpe?g|gif|ico|webp|woff2?|ttf|pdf))$|(^|\/)(dist|generated|__generated__|node_modules|\.yarn|__snapshots__|coverage|\.next|\.turbo|drizzle\/meta)\//i;
+
 // utils/isPullfrog.ts
 function isPullfrog(actor) {
   actor = actor?.toLowerCase().replace("[bot]", "");
@@ -23480,6 +23484,7 @@ export {
   DEFAULT_EFFORT_POSITION,
   DEFAULT_PROXY_MODEL,
   EFFORT_ALIASES,
+  GENERATED,
   LEAPING_INTO_ACTION_PREFIX,
   MAX_LEARNINGS_LENGTH,
   OAuthInvalidGrantError,

package/dist/mcp/review.d.ts

@@ -263,4 +263,5 @@ export declare function createAndSubmitWithFooter(ctx: ToolContext, params: Rest
  */
 export declare function reportReviewNodeId(ctx: ToolContext, params: {
     nodeId: string;
+    reviewedSha: string | undefined;
 }): Promise<void>;

package/dist/utils/diffCoverage.d.ts

@@ -38,10 +38,18 @@ export declare function countLines(params: {
 export declare function parseDiffTocEntries(params: {
     toc: string;
 }): DiffTocEntry[];
+/**
+ * one entry per section of an incremental (range-diff) diff, which opens on an
+ * `@@ <file>[: <function>]` hunk header or a `## <file>[ (new)] ##` file header,
+ * either of which may be the only line naming its file.
+ */
+export declare function parseIncrementalDiffEntries(params: {
+    content: string;
+}): DiffTocEntry[];
 export declare function createDiffCoverageState(params: {
     diffPath: string;
     totalLines: number;
-    toc: string;
+    tocEntries: DiffTocEntry[];
     previous?: DiffCoverageState | undefined;
 }): DiffCoverageState;
 export declare function recordDiffReadFromToolUse(params: {

package/dist/utils/generatedFiles.d.ts

@@ -0,0 +1,10 @@
+/**
+ * files whose hunks carry nothing a reviewer weighs — lockfiles, minified and
+ * generated output, snapshots, images and fonts, built trees, migration
+ * metadata. the router's scorer reads a one-line stub in their place and its
+ * heuristic skips their lines; a review's diff keeps them but leaves them off
+ * its coverage checklist. `build/` and `vendor/` are deliberately not here: the
+ * router eval set has source under both (`src/build/expo-brand.ts`, a vendored
+ * crate patched in the PR), and hiding a patch is worse than reading a copy.
+ */
+export declare const GENERATED: RegExp;

package/dist/utils/patchWorkflowRunFields.d.ts

@@ -4,6 +4,8 @@ import type { ToolContext } from "../mcp/server.ts";
  * String-valued PATCH fields (all serialized identically on the wire):
  *  - artifact node IDs (`*NodeId`, `summarySnapshot`) — PATCHed incrementally
  *    by MCP tools as GitHub entities are created during the run.
+ *  - `reviewedCommitSha` — the PR head a review covered, sent with `reviewNodeId`
+ *    or `reviewSkipped`; the next re-review's delta starts there.
  *  - `model` — the resolved/effective model the run actually ran on (proxy spec
  *    for router/oss, post-fallback slug otherwise; NOT the configured
  *    `Repo.model` slug), PATCHed once at end-of-run so per-model cost analytics
@@ -13,7 +15,7 @@ import type { ToolContext } from "../mcp/server.ts";
  * workflow can start the action several times in one run, and the server keeps
  * one row per start, keyed by `START_KEY`, and sums them onto the run.
  */
-declare const STRING_KEYS: readonly ["prNodeId", "issueNodeId", "reviewNodeId", "planCommentNodeId", "summarySnapshot", "model", "agent", "credential"];
+declare const STRING_KEYS: readonly ["prNodeId", "issueNodeId", "reviewNodeId", "reviewedCommitSha", "planCommentNodeId", "summarySnapshot", "model", "agent", "credential"];
 /**
  * Number-valued usage fields — aggregated across all agent calls and PATCHed
  * once at end-of-run. Token counts are Int4 on the DB side (ample for any
@@ -23,7 +25,7 @@ declare const STRING_KEYS: readonly ["prNodeId", "issueNodeId", "reviewNodeId",
 declare const NUMBER_KEYS: readonly ["inputTokens", "outputTokens", "cacheReadTokens", "cacheWriteTokens", "costUsd"];
 /** one-way flags, only ever sent as `true`. Keep in sync with `BOOLEAN_FIELDS` in the server route. */
 declare const FLAG_KEYS: readonly ["reviewSkipped"];
-export type WorkflowRunPatch = Partial<Record<(typeof STRING_KEYS)[number], string>> & Partial<Record<(typeof NUMBER_KEYS)[number], number>> & Partial<Record<(typeof FLAG_KEYS)[number], true>> & {
+export type WorkflowRunPatch = Partial<Record<(typeof STRING_KEYS)[number], string | undefined>> & Partial<Record<(typeof NUMBER_KEYS)[number], number>> & Partial<Record<(typeof FLAG_KEYS)[number], true>> & {
     succeeded?: boolean;
 };
 /** PATCH workflow-run fields (Pullfrog JWT, not GitHub). */

package/dist/utils/rangeDiff.d.ts

@@ -2,6 +2,8 @@ type ComputeIncrementalDiffParams = {
     baseBranch: string;
     beforeSha: string;
     headSha: string;
+    /** tips the base branch was force-pushed away from, which either version may still be built on */
+    formerBaseTips: string[];
 };
 /**
  * computes the incremental diff between two versions of a PR using range-diff
@@ -64,7 +66,9 @@ export declare function computeIncrementalDiff(params: ComputeIncrementalDiffPar
  *
  * uses a streaming approach: a ring buffer of before-context lines is flushed when
  * a change is hit, then afterCount lines of after-context are emitted directly.
- * nearest preceding ## / @@ headers are force-included when outside the context window.
+ * nearest preceding ## / @@ headers are force-included when outside the context window,
+ * and so is the outer `@@ <file>[: <function>]` hunk header, which is often the only
+ * line naming the file a change belongs to.
  */
 export declare function postProcessRangeDiff(raw: string, contextLines?: number): string | null;
 export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pullfrog",
-  "version": "0.1.102",
+  "version": "0.1.104",
   "cliContractVersion": "1",
   "type": "module",
   "bin": {