@iceinvein/agent-skills 0.1.40 → 0.3.0

package/skills/magpie/scripts/gh.ts

@@ -1,16 +1,21 @@
 import { writeFile } from 'node:fs/promises'
 import { join } from 'node:path'
+import { localPrDiff } from './git-diff.ts'
 
 export type FetchPrInput = {
   ghBin: string
   prNumber: number
   runDir: string
-  /** Working directory for the gh invocation. Defaults to process.cwd(). */
+  /** Working directory for the gh and git invocations. Defaults to process.cwd(). */
   cwd?: string
+  /** git binary for the local-diff fallback. Defaults to `git` on PATH. */
+  gitBin?: string
+  /** Extra environment for the gh invocation. Test seam for the fake gh fixtures. */
+  env?: Record<string, string>
 }
 
 export type FetchPrResult =
-  | { ok: true; prJsonPath: string; diffPath: string }
+  | { ok: true; prJsonPath: string; diffPath: string; source: 'gh' | 'git'; mergeBase?: string }
   | { ok: false; error: string }
 
 export const PR_VIEW_FIELDS = [
@@ -31,9 +36,23 @@ export const PR_VIEW_FIELDS = [
   'closingIssuesReferences',
 ].join(',')
 
+function prFileCount(viewStdout: string): number {
+  try {
+    const parsed = JSON.parse(viewStdout) as { files?: unknown }
+    return Array.isArray(parsed.files) ? parsed.files.length : 0
+  } catch {
+    return 0
+  }
+}
+
 export async function fetchPr(input: FetchPrInput): Promise<FetchPrResult> {
   const { ghBin, prNumber, runDir, cwd } = input
-  const spawnOpts = { stdout: 'pipe' as const, stderr: 'pipe' as const, ...(cwd ? { cwd } : {}) }
+  const spawnOpts = {
+    stdout: 'pipe' as const,
+    stderr: 'pipe' as const,
+    ...(cwd ? { cwd } : {}),
+    ...(input.env ? { env: { ...process.env, ...input.env } } : {}),
+  }
   const view = Bun.spawn(
     [ghBin, 'pr', 'view', String(prNumber), '--json', PR_VIEW_FIELDS],
     spawnOpts,
@@ -48,20 +67,52 @@ export async function fetchPr(input: FetchPrInput): Promise<FetchPrResult> {
     }
   }
 
+  const prJsonPath = join(runDir, 'pr.json')
+  const diffPath = join(runDir, 'diff.patch')
+  await writeFile(prJsonPath, viewStdout)
+
   const diff = Bun.spawn([ghBin, 'pr', 'diff', String(prNumber)], spawnOpts)
   const diffStdout = await new Response(diff.stdout).text()
   const diffStderr = await new Response(diff.stderr).text()
   const diffExit = await diff.exited
-  if (diffExit !== 0) {
-    return {
-      ok: false,
-      error: `gh pr diff exit ${diffExit}: ${diffStderr.trim()}`,
-    }
+
+  // GitHub refuses the `.diff` media type above roughly 300 files (HTTP 406),
+  // and a truncating proxy can answer 200 with nothing at all. Reviewing an
+  // empty diff as though it were the PR is the worse failure of the two, so
+  // both conditions fall through to the local clone.
+  const emptyForNonEmptyPr = diffStdout.trim() === '' && prFileCount(viewStdout) > 0
+  if (diffExit === 0 && !emptyForNonEmptyPr) {
+    await writeFile(diffPath, diffStdout)
+    return { ok: true, prJsonPath, diffPath, source: 'gh' }
   }
 
-  const prJsonPath = join(runDir, 'pr.json')
-  const diffPath = join(runDir, 'diff.patch')
-  await writeFile(prJsonPath, viewStdout)
-  await writeFile(diffPath, diffStdout)
-  return { ok: true, prJsonPath, diffPath }
+  const ghError =
+    diffExit !== 0
+      ? `gh pr diff exit ${diffExit}: ${diffStderr.trim()}`
+      : 'gh pr diff returned an empty diff for a PR with changed files'
+
+  let pr: { baseRefName?: unknown; headRefOid?: unknown }
+  try {
+    pr = JSON.parse(viewStdout) as { baseRefName?: unknown; headRefOid?: unknown }
+  } catch (err) {
+    return { ok: false, error: `${ghError}; pr.json is unparseable: ${String(err)}` }
+  }
+  const baseRefName = typeof pr.baseRefName === 'string' ? pr.baseRefName : ''
+  const headRefOid = typeof pr.headRefOid === 'string' ? pr.headRefOid : ''
+  if (!baseRefName || !headRefOid) {
+    return { ok: false, error: `${ghError}; pr.json has no baseRefName/headRefOid to diff from` }
+  }
+
+  const local = await localPrDiff({
+    gitBin: input.gitBin ?? 'git',
+    repoPath: cwd ?? process.cwd(),
+    prNumber,
+    baseRefName,
+    headRefOid,
+  })
+  if (!local.ok) {
+    return { ok: false, error: `${ghError}; local fallback failed: ${local.error}` }
+  }
+  await writeFile(diffPath, local.diff)
+  return { ok: true, prJsonPath, diffPath, source: 'git', mergeBase: local.mergeBase }
 }

package/skills/magpie/scripts/git-diff.ts

@@ -0,0 +1,111 @@
+export type LocalDiffInput = {
+  gitBin: string
+  repoPath: string
+  prNumber: number
+  baseRefName: string
+  headRefOid: string
+}
+
+export type LocalDiffResult =
+  | { ok: true; diff: string; mergeBase: string }
+  | { ok: false; error: string }
+
+async function git(
+  gitBin: string,
+  cwd: string,
+  args: string[],
+): Promise<{ exit: number; stdout: string; stderr: string }> {
+  const proc = Bun.spawn([gitBin, ...args], { cwd, stdout: 'pipe', stderr: 'pipe' })
+  const stdout = await new Response(proc.stdout).text()
+  const stderr = await new Response(proc.stderr).text()
+  const exit = await proc.exited
+  return { exit, stdout, stderr }
+}
+
+async function resolveCommit(gitBin: string, cwd: string, ref: string): Promise<string | null> {
+  const r = await git(gitBin, cwd, ['rev-parse', '--verify', '--quiet', `${ref}^{commit}`])
+  const sha = r.stdout.trim()
+  return r.exit === 0 && sha.length > 0 ? sha : null
+}
+
+/**
+ * Render a fetch's outcome for an error message, so a genuine failure (wrong
+ * remote, auth failure, rate limiting) can be told apart from a benign one
+ * (no origin, offline, non-GitHub remote) when resolution ends up failing.
+ * Empty when the fetch succeeded, since a successful fetch explains nothing.
+ */
+function fetchDiagnostic(result: { exit: number; stderr: string }): string {
+  if (result.exit === 0) return ''
+  const stderr = result.stderr.trim()
+  return stderr ? `; git fetch: ${stderr}` : `; git fetch exited ${result.exit}`
+}
+
+/**
+ * Reproduce `gh pr diff` from the local clone. GitHub refuses the `.diff` media
+ * type above roughly 300 files, and a review that cannot get a diff is worth
+ * less than one built from local objects.
+ *
+ * Diffs from the merge base rather than the base tip so the output matches
+ * `gh pr diff`'s three-dot semantics: the PR's own changes, not every commit
+ * the base branch gained since the branch point.
+ */
+export async function localPrDiff(input: LocalDiffInput): Promise<LocalDiffResult> {
+  const { gitBin, repoPath, prNumber, baseRefName, headRefOid } = input
+  const prRef = `refs/magpie/pr-${prNumber}`
+
+  // Best effort. `pull/<n>/head` is a GitHub convention and resolves fork PRs
+  // that `headRefName` cannot, but a repo with no `origin`, an offline run, or
+  // a non-GitHub remote must still work when the commits are already local.
+  // Each fetch's own result is kept (not discarded) so, if resolution below
+  // still fails, the error can say why the fetch didn't help.
+  const headFetch = await git(gitBin, repoPath, [
+    'fetch',
+    'origin',
+    `pull/${prNumber}/head:${prRef}`,
+  ])
+  const baseFetch = await git(gitBin, repoPath, ['fetch', 'origin', baseRefName])
+
+  let head: string | null = null
+  for (const ref of [prRef, headRefOid]) {
+    head = await resolveCommit(gitBin, repoPath, ref)
+    if (head) break
+  }
+  if (!head) {
+    return {
+      ok: false,
+      error: `cannot resolve PR head ${headRefOid} locally (tried ${prRef} and the SHA); fetch the PR branch and retry${fetchDiagnostic(headFetch)}`,
+    }
+  }
+  // Reviewing a stale local ref and reporting it as a review of the PR is worse
+  // than failing, so a mismatch is fatal rather than a warning.
+  if (head !== headRefOid) {
+    return {
+      ok: false,
+      error: `local PR head ${head} does not match the PR head ${headRefOid}; the local ref is stale`,
+    }
+  }
+
+  let base: string | null = null
+  for (const ref of [`origin/${baseRefName}`, baseRefName]) {
+    base = await resolveCommit(gitBin, repoPath, ref)
+    if (base) break
+  }
+  if (!base) {
+    return {
+      ok: false,
+      error: `cannot resolve base branch ${baseRefName} locally (tried origin/${baseRefName} and ${baseRefName})${fetchDiagnostic(baseFetch)}`,
+    }
+  }
+
+  const mb = await git(gitBin, repoPath, ['merge-base', base, head])
+  if (mb.exit !== 0) {
+    return { ok: false, error: `git merge-base exit ${mb.exit}: ${mb.stderr.trim()}` }
+  }
+  const mergeBase = mb.stdout.trim()
+
+  const d = await git(gitBin, repoPath, ['diff', '--no-color', '--find-renames', mergeBase, head])
+  if (d.exit !== 0) {
+    return { ok: false, error: `git diff exit ${d.exit}: ${d.stderr.trim()}` }
+  }
+  return { ok: true, diff: d.stdout, mergeBase }
+}

package/skills/magpie/scripts/path-filter.ts

@@ -1,6 +1,7 @@
 import { existsSync } from 'node:fs'
 import { readFile } from 'node:fs/promises'
 import { join } from 'node:path'
+import { chunkPath, splitRawChunks } from './diff-chunks.ts'
 
 export type PathFilterConfig = {
   exclude: string[]
@@ -43,6 +44,12 @@ export const DEFAULT_EXCLUDES: readonly string[] = [
   '**/*.min.js',
   '**/*.min.css',
   '**/*.map',
+  // Generated .NET / C#
+  '**/*.Designer.cs',
+  '**/*ModelSnapshot.cs',
+  '**/*.g.cs',
+  '**/*.g.i.cs',
+  '**/*.designer.vb',
   // Snapshot fixtures
   '**/*.snap',
   '**/__snapshots__/**',
@@ -113,11 +120,9 @@ export async function loadPathFilterConfig(cwd: string): Promise<PathFilterConfi
 export type ExcludedFile = { path: string; pattern: string }
 export type FilterDiffResult = { filtered: string; excluded: ExcludedFile[] }
 
-const FILE_HEADER = /^diff --git a\/(.+?) b\/(.+?)$/m
-
 export function filterDiff(rawDiff: string, config: PathFilterConfig): FilterDiffResult {
   if (!rawDiff) return { filtered: '', excluded: [] }
-  const chunks = rawDiff.split(/^(?=diff --git )/m)
+  const chunks = splitRawChunks(rawDiff)
   const kept: string[] = []
   const excluded: ExcludedFile[] = []
   for (const chunk of chunks) {
@@ -125,8 +130,7 @@ export function filterDiff(rawDiff: string, config: PathFilterConfig): FilterDif
       if (chunk.trim()) kept.push(chunk)
       continue
     }
-    const m = chunk.match(FILE_HEADER)
-    const path = m?.[2] ?? m?.[1]
+    const path = chunkPath(chunk)
     if (!path) {
       kept.push(chunk)
       continue

package/skills/magpie/scripts/refresh.ts

@@ -98,6 +98,13 @@ export async function refreshFindings(runDir: string): Promise<RefreshResult> {
 
   const diff = await readFile(join(runDir, 'diff.patch'), 'utf8').catch(() => '')
 
+  let diffSource: { source: 'gh' | 'git'; mergeBase: string | null } | undefined
+  try {
+    diffSource = (await Bun.file(join(runDir, 'diff-source.json')).json()) as typeof diffSource
+  } catch {
+    diffSource = undefined
+  }
+
   await renderFindingsToDisk(
     {
       findings,
@@ -108,6 +115,7 @@ export async function refreshFindings(runDir: string): Promise<RefreshResult> {
       diff,
       brief: brief ?? undefined,
       issues,
+      diffSource,
     },
     join(screenDir, 'findings.html'),
   )

package/skills/magpie/scripts/render-cmd.ts

@@ -48,12 +48,15 @@ function summarizeStages(log: Array<Record<string, unknown>>): {
     report: 'pending',
     post: 'pending',
   }
-  const specialistCounts: Record<string, number> = {
-    security: 0,
-    bugs: 0,
-    performance: 0,
-    'code-smells': 0,
-    architecture: 0,
+  // focus -> shard key -> count. Keyed by shard so a re-dispatched (focus, shard)
+  // pair replaces its own count rather than doubling it, while distinct shards
+  // of the same focus add up.
+  const perShard: Record<string, Record<string, number>> = {
+    security: {},
+    bugs: {},
+    performance: {},
+    'code-smells': {},
+    architecture: {},
   }
   for (const entry of log) {
     const stage = entry.stage as string
@@ -63,11 +66,16 @@ function summarizeStages(log: Array<Record<string, unknown>>): {
     if (stage in stages && status === 'error') stages[stage] = 'error'
     if (stage in stages && status === 'skipped') stages[stage] = 'skipped'
     if (stage === 'specialist' && typeof entry.focus === 'string') {
-      if (entry.focus in specialistCounts && typeof entry.findings === 'number') {
-        specialistCounts[entry.focus] = entry.findings
+      const bucket = perShard[entry.focus]
+      if (bucket && typeof entry.findings === 'number') {
+        bucket[String(entry.shard ?? 'all')] = entry.findings
       }
     }
   }
+  const specialistCounts: Record<string, number> = {}
+  for (const [focus, bucket] of Object.entries(perShard)) {
+    specialistCounts[focus] = Object.values(bucket).reduce((a, b) => a + b, 0)
+  }
   return { stages, specialistCounts }
 }
 
@@ -81,6 +89,11 @@ export async function runRender(runDir: string, page: 'progress' | 'findings'):
     })
     const log = await readLog(runDir)
     const summary = summarizeStages(log)
+    const manifest = await readJson<{ shards?: unknown[] }>(
+      join(runDir, 'shards', 'manifest.json'),
+      {},
+    )
+    const shardCount = Array.isArray(manifest.shards) ? manifest.shards.length : 1
     const outPath = await nextVersionedPath(screenDir, 'progress')
     await renderProgressToDisk(
       {
@@ -89,6 +102,7 @@ export async function runRender(runDir: string, page: 'progress' | 'findings'):
         branch: String(prJson.headRefName ?? '?'),
         stages: summary.stages as never,
         specialistCounts: summary.specialistCounts,
+        shardCount,
       },
       outPath,
     )
@@ -124,9 +138,14 @@ export async function runRender(runDir: string, page: 'progress' | 'findings'):
   // is simply omitted.
   const brief = parseBrief(await readJson<unknown>(join(runDir, 'brief.json'), null)) ?? undefined
   const issues = parseClosingIssues(prJson)
+  const diffSource =
+    (await readJson<{ source: 'gh' | 'git'; mergeBase: string | null } | null>(
+      join(runDir, 'diff-source.json'),
+      null,
+    )) ?? undefined
   const outPath = await nextVersionedPath(screenDir, 'findings')
   await renderFindingsToDisk(
-    { findings, postStatus, runId: basename(runDir), pr, files, diff, brief, issues },
+    { findings, postStatus, runId: basename(runDir), pr, files, diff, brief, issues, diffSource },
     outPath,
   )
   return 0

package/skills/magpie/scripts/render-findings.ts

@@ -59,6 +59,8 @@ export type RenderFindingsInput = {
   issues?: BriefIssue[]
   /** Shiki highlighter, prepared by the caller. */
   highlighter: Highlighter
+  /** Where diff.patch came from. Absent means gh, which needs no note. */
+  diffSource?: { source: 'gh' | 'git'; mergeBase: string | null }
 }
 
 function esc(s: string): string {
@@ -86,13 +88,21 @@ function brandBlock(): string {
 
 function prHeader(input: RenderFindingsInput): string {
   const pr = input.pr
+  const src = input.diffSource
+  const diffNote =
+    src?.source === 'git'
+      ? `<span class="diff-source" title="gh pr diff could not serve this PR, so the diff was rebuilt from the local clone">diff from local clone${
+          src.mergeBase ? ` at <code>${esc(src.mergeBase.slice(0, 12))}</code>` : ''
+        }</span>`
+      : ''
   const prMeta = pr
     ? `<div class="pr-meta">
         <span class="pr-number">PR #${pr.number}</span>
         <span class="pr-branch">${esc(pr.branch)}</span>
         <code>${esc(pr.headSha.slice(0, 12))}</code>
+        ${diffNote}
       </div>`
-    : `<div class="pr-meta"></div>`
+    : `<div class="pr-meta">${diffNote}</div>`
   const total = input.findings.length
   return `<header class="pr-header">
     ${brandBlock()}

package/skills/magpie/scripts/render-progress.ts

@@ -41,6 +41,8 @@ export type RenderProgressInput = {
   branch: string
   stages: Record<StageId, StageStatus>
   specialistCounts: Record<string, number>
+  /** Shard count from shards/manifest.json. Absent or 1 means unsharded. */
+  shardCount?: number
 }
 
 function esc(s: string): string {
@@ -82,7 +84,10 @@ function progressPane(input: RenderProgressInput): string {
   const cur = current(input.stages)
   let lead: string
   if (cur.status === 'running') {
-    lead = STAGE_NOW_DOING[cur.id]
+    lead =
+      cur.id === 'specialists' && (input.shardCount ?? 1) > 1
+        ? `Five reviewers across ${input.shardCount} shards`
+        : STAGE_NOW_DOING[cur.id]
   } else if (cur.status === 'error') {
     lead = `${STAGE_NOW_DOING[cur.id]} stalled`
   } else if (cur.status === 'pending') {

package/skills/magpie/scripts/setup-cmd.ts

@@ -4,6 +4,7 @@ import { fetchPr } from './gh.ts'
 import { buildIncrementalContext, findPreviousRun } from './incremental.ts'
 import { filterDiff, loadPathFilterConfig } from './path-filter.ts'
 import { type Deps, defaultDeps, preflight, renderInstallHint } from './preflight.ts'
+import { shardDiff } from './shard.ts'
 import { detectMissingTests } from './tests-check.ts'
 import { createWorktree } from './worktree.ts'
 
@@ -47,6 +48,7 @@ export async function runSetup(input: RunSetupInput): Promise<number> {
 
   const fetched = await fetchPr({
     ghBin: deps.gh,
+    gitBin: deps.git,
     prNumber: input.prNumber,
     runDir: input.runDir,
     cwd: input.repoPath,
@@ -57,10 +59,31 @@ export async function runSetup(input: RunSetupInput): Promise<number> {
     await cleanup(input.runDir)
     return 4
   }
-  await logLine(input.runDir, { stage: 'fetch-pr', status: 'done' })
+  await logLine(input.runDir, {
+    stage: 'fetch-pr',
+    status: 'done',
+    source: fetched.source,
+    ...(fetched.mergeBase ? { mergeBase: fetched.mergeBase } : {}),
+  })
+  // Sidecar so the report can say where the diff came from without re-parsing
+  // log.jsonl, the same way incremental.json carries the incremental context.
+  await writeFile(
+    join(input.runDir, 'diff-source.json'),
+    `${JSON.stringify({ source: fetched.source, mergeBase: fetched.mergeBase ?? null }, null, 2)}\n`,
+  )
 
   await applyPathFilter(input.runDir, input.repoPath)
 
+  try {
+    await runShard(input.runDir)
+  } catch (err) {
+    const message = err instanceof Error ? err.message : String(err)
+    await logLine(input.runDir, { stage: 'shard', status: 'error', error: message })
+    process.stderr.write(`magpie: ${message}\n`)
+    await cleanup(input.runDir)
+    return 6
+  }
+
   await runTestsCheck(input.runDir)
 
   await detectIncrementalReview(input.runDir, input.prNumber)
@@ -90,7 +113,9 @@ async function cleanup(runDir: string): Promise<void> {
     'pr.json',
     'diff.patch',
     'diff.full.patch',
+    'diff-source.json',
     'excluded-files.json',
+    'shards',
     'incremental.json',
     'findings',
     'screen',
@@ -100,6 +125,18 @@ async function cleanup(runDir: string): Promise<void> {
   }
 }
 
+async function runShard(runDir: string): Promise<void> {
+  const manifest = await shardDiff(runDir)
+  await logLine(runDir, {
+    stage: 'shard',
+    status: 'done',
+    shards: manifest.shards.length,
+    budget: manifest.budget,
+    files: manifest.totalFiles,
+    lines: manifest.totalLines,
+  })
+}
+
 async function runTestsCheck(runDir: string): Promise<void> {
   const diff = await Bun.file(join(runDir, 'diff.patch')).text()
   const findings = detectMissingTests(diff)

package/skills/magpie/scripts/shard.ts

@@ -0,0 +1,171 @@
+import { mkdir, readdir, rm, writeFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { type FileChunk, splitFileChunks } from './diff-chunks.ts'
+
+export const DEFAULT_SHARD_BUDGET = 6000
+export const DEFAULT_SHARD_MAX_FILES = 80
+
+export type ShardEntry = { id: number; path: string; files: string[]; lines: number }
+export type ShardManifest = {
+  budget: number
+  maxFiles: number
+  totalFiles: number
+  totalLines: number
+  shards: ShardEntry[]
+}
+
+/**
+ * The directory a file is grouped under, capped at two segments. Cohesion is
+ * not cosmetic: the architecture and code-smells focuses look for duplication,
+ * boundary violations, and shotgun surgery, and splitting a module across
+ * shards hides exactly those.
+ */
+export function groupKey(path: string): string {
+  const parts = path.split('/')
+  if (parts.length <= 1) return '.'
+  return parts.slice(0, Math.min(2, parts.length - 1)).join('/')
+}
+
+export function planShards(chunks: FileChunk[], budget: number, maxFiles: number): FileChunk[][] {
+  const groups = new Map<string, FileChunk[]>()
+  for (const c of chunks) {
+    const key = groupKey(c.path)
+    const existing = groups.get(key)
+    if (existing) existing.push(c)
+    else groups.set(key, [c])
+  }
+
+  const shards: FileChunk[][] = []
+  let current: FileChunk[] = []
+  let currentLines = 0
+  const flush = () => {
+    if (current.length > 0) {
+      shards.push(current)
+      current = []
+      currentLines = 0
+    }
+  }
+
+  for (const members of groups.values()) {
+    const groupLines = members.reduce((n, c) => n + c.lines, 0)
+    // Start a fresh shard rather than straddle a group across a boundary, when
+    // the whole group could fit on its own.
+    if (
+      current.length > 0 &&
+      (currentLines + groupLines > budget || current.length + members.length > maxFiles)
+    ) {
+      flush()
+    }
+    for (const c of members) {
+      if (
+        current.length > 0 &&
+        (currentLines + c.lines > budget || current.length + 1 > maxFiles)
+      ) {
+        flush()
+      }
+      current.push(c)
+      currentLines += c.lines
+    }
+  }
+  flush()
+  return shards
+}
+
+/** Read `$RUN_DIR/diff.patch`, tolerating only its absence. A missing
+ *  diff.patch means "nothing to shard"; anything else (permission error,
+ *  disk error, EISDIR, ...) is a genuine failure and must propagate rather
+ *  than silently turn into an empty manifest. */
+async function readDiffPatch(runDir: string): Promise<string> {
+  try {
+    return await Bun.file(join(runDir, 'diff.patch')).text()
+  } catch (err) {
+    const code = err && typeof err === 'object' && 'code' in err ? err.code : undefined
+    if (code === 'ENOENT') return ''
+    throw err
+  }
+}
+
+/** Delete `shard-<n>.patch` files left by an earlier split. Re-splitting 7 shards
+ *  into 4 would otherwise leave `shard-5.patch` .. `shard-7.patch` on disk: the
+ *  manifest is authoritative for dispatch, but a stale patch makes the directory
+ *  claim coverage the manifest does not have, and a resume that reads shard ids
+ *  off disk would review a file set that no longer matches its id. Only the shard
+ *  patches are touched: `manifest.json` stays in place until it is overwritten at
+ *  the end (so there is never a window with no manifest), and `diff.patch` lives
+ *  outside this directory and is what a single-shard manifest points at. */
+async function clearStaleShardPatches(shardsDir: string): Promise<void> {
+  let entries: string[]
+  try {
+    entries = await readdir(shardsDir)
+  } catch {
+    return
+  }
+  for (const name of entries) {
+    if (/^shard-\d+\.patch$/.test(name)) await rm(join(shardsDir, name), { force: true })
+  }
+}
+
+/**
+ * Split `$RUN_DIR/diff.patch` into budgeted shards. `diff.patch` is never
+ * modified: shards are views over it, which is what keeps the critic's diff
+ * excerpts, dedupe's evidence check, and the report's rendering working
+ * unchanged.
+ *
+ * Re-splitting an already-sharded run invalidates any
+ * `findings/<focus>.shard-<n>.json` from the previous split: the same shard id
+ * now names a different file set. Stage 4's resume rule keys on those files, so
+ * the caller must delete them before re-dispatching.
+ */
+export async function shardDiff(
+  runDir: string,
+  opts: { budget?: number; maxFiles?: number } = {},
+): Promise<ShardManifest> {
+  const budget = opts.budget ?? DEFAULT_SHARD_BUDGET
+  const maxFiles = opts.maxFiles ?? DEFAULT_SHARD_MAX_FILES
+  const diff = await readDiffPatch(runDir)
+  const chunks = splitFileChunks(diff)
+  const totalLines = chunks.reduce((n, c) => n + c.lines, 0)
+  const shardsDir = join(runDir, 'shards')
+  await clearStaleShardPatches(shardsDir)
+  await mkdir(shardsDir, { recursive: true })
+
+  const planned = planShards(chunks, budget, maxFiles)
+  let shards: ShardEntry[]
+  if (planned.length === 0) {
+    shards = []
+  } else if (planned.length === 1) {
+    // Single-shard passthrough: stage 4 takes the same path it took on 0.9.0.
+    const only = planned[0] ?? []
+    shards = [
+      {
+        id: 1,
+        path: 'diff.patch',
+        files: only.map((c) => c.path),
+        lines: only.reduce((n, c) => n + c.lines, 0),
+      },
+    ]
+  } else {
+    shards = []
+    for (const [i, members] of planned.entries()) {
+      const id = i + 1
+      const rel = `shards/shard-${id}.patch`
+      await writeFile(join(runDir, rel), members.map((c) => c.text).join(''))
+      shards.push({
+        id,
+        path: rel,
+        files: members.map((c) => c.path),
+        lines: members.reduce((n, c) => n + c.lines, 0),
+      })
+    }
+  }
+
+  const manifest: ShardManifest = {
+    budget,
+    maxFiles,
+    totalFiles: chunks.length,
+    totalLines,
+    shards,
+  }
+  await writeFile(join(shardsDir, 'manifest.json'), `${JSON.stringify(manifest, null, 2)}\n`)
+  return manifest
+}

package/skills/magpie/scripts/status-cmd.ts

@@ -1,7 +1,10 @@
 import { readFile } from 'node:fs/promises'
 import { join } from 'node:path'
 
-const ORDER = [
+/** The resume ladder. Exported so `skill-lint` can assert that the diagnostic
+ *  stage names SKILL.md logs (`filter`, `tests-check`, `shard-coverage`, ...) stay
+ *  out of it: a diagnostic entry that lands on this list would move `next`. */
+export const ORDER = [
   'setup',
   'context',
   'specialists',