@et0and/ovid 0.0.2 → 0.0.4

package/package.json

@@ -1,11 +1,8 @@
 {
   "name": "@et0and/ovid",
-  "version": "0.0.2",
+  "version": "0.0.4",
   "description": "Browse a repository's files by semantic meaning",
   "type": "module",
-  "bin": {
-    "ovid": "./bin/semantic-navigator.js"
-  },
   "files": [
     "bin",
     "src",

package/src/auth.ts

@@ -74,9 +74,9 @@ async function fetchCopilotToken(githubToken: string): Promise<{ token: string;
     method: "GET",
     headers: {
       Authorization: `token ${githubToken}`,
-      "Editor-Version": "semantic-navigator/1.0.0",
-      "Editor-Plugin-Version": "semantic-navigator/1.0.0",
-      "User-Agent": "semantic-navigator",
+      "Editor-Version": "vscode/1.95.0",
+      "Editor-Plugin-Version": "copilot/1.246.0",
+      "User-Agent": "GitHubCopilotChat/0.22.4",
     },
   })
 

package/src/cache.ts

@@ -0,0 +1,119 @@
+/**
+ * Two-layer result cache for semantic-navigator.
+ *
+ * Layer 1 — Embedding cache (~/.cache/semantic-navigator/embeddings.json)
+ *   Content-addressed: maps sha256(text)[0:16] → number[]
+ *   Per-entry granularity: only re-embed chunks whose text changed.
+ *
+ * Layer 2 — Tree cache (~/.cache/semantic-navigator/trees/<fingerprint>.json)
+ *   Keyed by sha256(model + sorted(path:contentHash pairs)).
+ *   A single changed file invalidates the whole tree, forcing a fresh
+ *   cluster + label run, but embeddings are still reused from layer 1.
+ */
+
+import { createHash } from "node:crypto"
+import { mkdirSync, existsSync, readFileSync, writeFileSync } from "node:fs"
+import { join } from "node:path"
+import { homedir } from "node:os"
+import type { Tree } from "./tree.ts"
+
+// ---------------------------------------------------------------------------
+// Paths
+// ---------------------------------------------------------------------------
+
+const CACHE_DIR = join(homedir(), ".cache", "semantic-navigator")
+const EMBED_CACHE_PATH = join(CACHE_DIR, "embeddings.json")
+const TREES_DIR = join(CACHE_DIR, "trees")
+
+function ensureDirs(): void {
+  mkdirSync(CACHE_DIR, { recursive: true })
+  mkdirSync(TREES_DIR, { recursive: true })
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+/** Short hex digest: sha256(text).slice(0, 16) */
+export function textHash(text: string): string {
+  return createHash("sha256").update(text).digest("hex").slice(0, 16)
+}
+
+/**
+ * Fingerprint for the tree cache: sha256(model + sorted path:contentHash pairs).
+ * `fileHashes` is a map from relative path to sha256(file content).
+ */
+export function treeFingerprint(model: string, fileHashes: Map<string, string>): string {
+  const entries = Array.from(fileHashes.entries())
+    .map(([p, h]) => `${p}:${h}`)
+    .sort()
+    .join("\n")
+  return createHash("sha256").update(model + "\n" + entries).digest("hex")
+}
+
+// ---------------------------------------------------------------------------
+// Layer 1: Embedding cache
+// ---------------------------------------------------------------------------
+
+type EmbedCacheMap = Record<string, number[]>
+
+let _embedCache: EmbedCacheMap | null = null
+
+function loadEmbedCache(): EmbedCacheMap {
+  if (_embedCache !== null) return _embedCache
+  if (existsSync(EMBED_CACHE_PATH)) {
+    try {
+      _embedCache = JSON.parse(readFileSync(EMBED_CACHE_PATH, "utf-8")) as EmbedCacheMap
+    } catch {
+      _embedCache = {}
+    }
+  } else {
+    _embedCache = {}
+  }
+  return _embedCache
+}
+
+/** Look up a cached embedding by text content. Returns null on miss. */
+export function getCachedEmbedding(text: string): Float32Array | null {
+  const cache = loadEmbedCache()
+  const key = textHash(text)
+  const vec = cache[key]
+  if (vec === undefined) return null
+  return Float32Array.from(vec)
+}
+
+/** Store an embedding in the in-memory cache (call flushEmbedCache to persist). */
+export function setCachedEmbedding(text: string, embedding: Float32Array): void {
+  const cache = loadEmbedCache()
+  const key = textHash(text)
+  cache[key] = Array.from(embedding)
+}
+
+/** Persist the in-memory embedding cache to disk. */
+export function flushEmbedCache(): void {
+  if (_embedCache === null) return
+  ensureDirs()
+  writeFileSync(EMBED_CACHE_PATH, JSON.stringify(_embedCache), "utf-8")
+}
+
+// ---------------------------------------------------------------------------
+// Layer 2: Tree cache
+// ---------------------------------------------------------------------------
+
+/** Look up a cached Tree by fingerprint. Returns null on miss. */
+export function getCachedTree(fingerprint: string): Tree | null {
+  const treePath = join(TREES_DIR, `${fingerprint}.json`)
+  if (!existsSync(treePath)) return null
+  try {
+    return JSON.parse(readFileSync(treePath, "utf-8")) as Tree
+  } catch {
+    return null
+  }
+}
+
+/** Persist a Tree to the tree cache. */
+export function setCachedTree(fingerprint: string, tree: Tree): void {
+  ensureDirs()
+  const treePath = join(TREES_DIR, `${fingerprint}.json`)
+  writeFileSync(treePath, JSON.stringify(tree), "utf-8")
+}

package/src/cluster.ts

@@ -95,19 +95,24 @@ function connectedComponents(indices: Int32Array[], N: number): number {
 }
 
 /**
- * Build the (dense) normalised Laplacian from the affinity matrix (stored as
- * a list of sparse {row,col,val} triples) and return it as a dense matrix
- * plus the degree diagonal `dd`.
+ * Build the degree diagonal `dd` from sparse affinity triplets, then return
+ * a sparse matvec closure for the *negated* normalised Laplacian (-L_norm).
+ *
+ * L_norm = I - D^{-1/2} A D^{-1/2}
+ * -L_norm = D^{-1/2} A D^{-1/2} - I
+ *
+ * The matvec avoids allocating an N×N dense matrix — at N=2000, k_sparse≈7,
+ * this reduces the cost per multiply from O(N²)=4M to O(N·k)=14k ops.
  */
-function buildNormalisedLaplacian(
+function buildNormLaplacianSparseMatvec(
   sparseAffinity: Array<{ i: number; j: number; v: number }>,
   N: number
-): { L: Matrix; dd: Float64Array } {
+): { matvec: (v: Float64Array) => Float64Array; dd: Float64Array } {
   // Accumulate row sums (degree) for normalisation
   const degree = new Float64Array(N)
   for (const { i, j, v } of sparseAffinity) {
-    degree[i] = (degree[i] ?? 0) + v
-    if (i !== j) degree[j] = (degree[j] ?? 0) + v
+    degree[i]! += v
+    if (i !== j) degree[j]! += v
   }
 
   const dd = new Float64Array(N)
@@ -115,31 +120,29 @@ function buildNormalisedLaplacian(
     dd[i] = degree[i]! > 1e-12 ? 1 / Math.sqrt(degree[i]!) : 0
   }
 
-  // L_norm = I - D^{-1/2} A D^{-1/2}
-  // We start from identity
-  const L: Matrix = Array.from({ length: N }, (_, i) => {
-    const row = new Float64Array(N)
-    row[i] = 1
-    return row
-  })
-
-  // Subtract normalised affinity contributions
-  for (const { i, j, v } of sparseAffinity) {
-    const w = v * dd[i]! * dd[j]!
-    const rowI = L[i]!
-    rowI[j] = (rowI[j] ?? 0) - w
-    if (i !== j) {
-      const rowJ = L[j]!
-      rowJ[i] = (rowJ[i] ?? 0) - w
+  // Pre-compute normalised weights once so the closure stays cheap.
+  // w_ij = v * dd[i] * dd[j]  (the off-diagonal contribution to A_norm)
+  const normAffinity = sparseAffinity.map(({ i, j, v }) => ({
+    i,
+    j,
+    w: v * dd[i]! * dd[j]!,
+  }))
+
+  // Matvec for -L_norm = A_norm - I
+  // result[i] = -v[i] + sum_j w_ij * v[j]   (using symmetry)
+  const matvecFn = (vec: Float64Array): Float64Array => {
+    const out = new Float64Array(N)
+    // Start from -I · vec
+    for (let i = 0; i < N; i++) out[i] = -vec[i]!
+    // Add symmetric A_norm contributions
+    for (const { i, j, w } of normAffinity) {
+      out[i]! += w * vec[j]!
+      if (i !== j) out[j]! += w * vec[i]!
     }
+    return out
   }
 
-  // Clamp diagonal to 1 (matches scipy behaviour after set_diag)
-  for (let i = 0; i < N; i++) {
-    L[i]![i] = 1
-  }
-
-  return { L, dd }
+  return { matvec: matvecFn, dd }
 }
 
 /**
@@ -171,35 +174,24 @@ function normaliseVec(v: Float64Array): number {
   return n
 }
 
-/** Multiply matrix M by vector v */
-function matvec(M: Matrix, v: Float64Array): Float64Array<ArrayBuffer> {
-  const N = M.length
-  const out = new Float64Array(N) as Float64Array<ArrayBuffer>
-  for (let i = 0; i < N; i++) {
-    out[i] = dot(M[i]!, v)
-  }
-  return out
-}
-
 /**
  * Randomised power-iteration with deflation to extract the `k` eigenpairs
- * corresponding to the *smallest* eigenvalues of a symmetric matrix M.
- *
- * M is the **negated** Laplacian (M = -L), so its *largest* eigenvalues
- * correspond to L's smallest — matching the Python code which does `laplacian *= -1`.
+ * corresponding to the *smallest* eigenvalues of a symmetric matrix.
  *
- * We use shifted inverse iteration: to find small eigenvalues of L we find
- * large eigenvalues of (-L + shift*I) where shift ≈ 1 (the diagonal was set
- * to 1 above).  We iterate on M = -L and take the top-k eigenvectors, then
- * negate the eigenvalues back.
+ * Instead of a dense matrix, accepts a sparse `matvecFn` closure so that the
+ * per-iteration cost is O(N·k_sparse) rather than O(N²).  The closure should
+ * implement multiplication by the *negated* Laplacian (-L_norm), whose top
+ * eigenvalues correspond to L_norm's bottom ones (matching the Python code
+ * which does `laplacian *= -1`).
  */
 function topKEigenpairs(
-  negL: Matrix,
+  matvecFn: (v: Float64Array) => Float64Array,
+  n: number,
   k: number,
   maxIter = 300,
   tol = 1e-6
 ): { values: Float64Array<ArrayBuffer>; vectors: Float64Array<ArrayBuffer>[] } {
-  const N = negL.length
+  const N = n
   const rng = seededRng(42)
 
   const vectors: Float64Array<ArrayBuffer>[] = []
@@ -207,7 +199,7 @@ function topKEigenpairs(
 
   for (let idx = 0; idx < k; idx++) {
     // Random start
-    let v = Float64Array.from({ length: N }, () => rng() * 2 - 1)
+    let v = Float64Array.from({ length: N }, () => rng() * 2 - 1) as Float64Array<ArrayBuffer>
     normaliseVec(v)
 
     // Deflate against already-found vectors
@@ -216,7 +208,7 @@ function topKEigenpairs(
 
     let lambda = 0
     for (let iter = 0; iter < maxIter; iter++) {
-      const Mv = matvec(negL, v)
+      const Mv = matvecFn(v) as Float64Array<ArrayBuffer>
 
       // Deflate
       for (const u of vectors) subtractProjection(Mv, u)
@@ -364,6 +356,29 @@ export function splitCluster(input: Cluster): Cluster[] {
 
   const normalized = normaliseRows(matFromEmbeds(input.entries))
 
+  // --- Precompute all pairwise distances once (O(N²)) ---
+  // Each row is sorted ascending so we can slice any k cheaply.
+  const allDistances: Array<Array<[number, number]>> = Array.from({ length: N }, () => [])
+  for (let i = 0; i < N; i++) {
+    for (let j = 0; j < N; j++) {
+      if (j === i) continue
+      allDistances[i]!.push([cosDist(normalized[i]!, normalized[j]!), j])
+    }
+    allDistances[i]!.sort((a, b) => a[0] - b[0])
+  }
+
+  /** Slice sorted rows to get k-NN result for any k in O(N·k). */
+  function knnFromPrecomputed(k: number): { distances: Float64Array[]; indices: Int32Array[] } {
+    const distances: Float64Array[] = []
+    const indices: Int32Array[] = []
+    for (let i = 0; i < N; i++) {
+      const row = allDistances[i]!.slice(0, k)
+      distances.push(Float64Array.from(row.map((x) => x[0])))
+      indices.push(Int32Array.from(row.map((x) => x[1])))
+    }
+    return { distances, indices }
+  }
+
   // --- Adaptive k-NN: find smallest k that gives 1 connected component ---
   const candidateKs: number[] = []
   for (let n = 0; ; n++) {
@@ -377,7 +392,7 @@ export function splitCluster(input: Cluster): Cluster[] {
   let chosenKnnResult: { distances: Float64Array[]; indices: Int32Array[] } | null = null
 
   for (const k of candidateKs) {
-    const knnResult = knn(normalized, k)
+    const knnResult = knnFromPrecomputed(k)
     const nComponents = connectedComponents(knnResult.indices, N)
     if (nComponents === 1) {
       chosenK = k
@@ -387,8 +402,7 @@ export function splitCluster(input: Cluster): Cluster[] {
   }
 
   if (chosenKnnResult === null) {
-    // Fallback: compute for the last candidate (floor(N/2))
-    chosenKnnResult = knn(normalized, chosenK)
+    chosenKnnResult = knnFromPrecomputed(chosenK)
   }
 
   const { distances, indices } = chosenKnnResult
@@ -411,19 +425,12 @@ export function splitCluster(input: Cluster): Cluster[] {
     }
   }
 
-  // --- Normalised Laplacian ---
-  const { L, dd } = buildNormalisedLaplacian(sparseAffinity, N)
-
-  // Negate L (as Python does `laplacian *= -1`) so power iteration finds
-  // eigenvectors of -L, whose top eigenvalues correspond to L's bottom ones.
-  const negL: Matrix = L.map((row) => {
-    const r = Float64Array.from(row)
-    for (let i = 0; i < r.length; i++) r[i]! *= -1
-    return r
-  })
+  // --- Sparse normalised Laplacian matvec ---
+  // Avoids building an N×N dense matrix; matvec cost is O(N·k_sparse) vs O(N²).
+  const { matvec: negLMatvec, dd } = buildNormLaplacianSparseMatvec(sparseAffinity, N)
 
   const k = Math.min(MAX_CLUSTERS + 1, N)
-  const { values: rawValues, vectors } = topKEigenpairs(negL, k)
+  const { values: rawValues, vectors } = topKEigenpairs(negLMatvec, N, k)
 
   // Eigenvalues were of -L; flip sign back to get L eigenvalues
   const eigenvalues = Float64Array.from(rawValues, (v) => -v)

package/src/embed.ts

@@ -1,5 +1,6 @@
 import { pipeline, type FeatureExtractionPipeline } from "@huggingface/transformers"
 import type { Chunk } from "./tokenize.ts"
+import { getCachedEmbedding, setCachedEmbedding, flushEmbedCache } from "./cache.ts"
 
 export const DEFAULT_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2"
 
@@ -19,6 +20,8 @@ export interface EmbedOptions {
   model: string
   batchSize: number
   concurrency: number
+  /** When true, skip reading from cache (but still write to it). */
+  noCache?: boolean
 }
 
 let _pipe: FeatureExtractionPipeline | null = null
@@ -81,6 +84,7 @@ async function embedBatch(
 
 /**
  * Embed all chunks using the local model, with batching + concurrency limits.
+ * Chunks whose text is already in the embedding cache are skipped.
  * Calls `onProgress(done, total)` after each batch completes.
  */
 export async function embedChunks(
@@ -90,44 +94,60 @@ export async function embedChunks(
 ): Promise<EmbedEntry[]> {
   if (chunks.length === 0) return []
 
-  const pipe = await getEmbedPipeline(opts.model)
-
-  const batches: Chunk[][] = []
-  for (let i = 0; i < chunks.length; i += opts.batchSize) {
-    batches.push(chunks.slice(i, i + opts.batchSize))
+  // --- Cache pass: resolve hits immediately, collect misses for the model ---
+  const entries: EmbedEntry[] = new Array(chunks.length)
+  const missIndices: number[] = []
+
+  if (!opts.noCache) {
+    for (let i = 0; i < chunks.length; i++) {
+      const cached = getCachedEmbedding(chunks[i]!.text)
+      if (cached !== null) {
+        entries[i] = { path: chunks[i]!.path, text: chunks[i]!.text, embedding: cached }
+      } else {
+        missIndices.push(i)
+      }
+    }
+  } else {
+    for (let i = 0; i < chunks.length; i++) missIndices.push(i)
   }
 
-  const entries: EmbedEntry[] = new Array(chunks.length)
-  let chunkIndex = 0
-  let done = 0
+  let done = chunks.length - missIndices.length
+  onProgress?.(done, chunks.length)
 
-  for (let i = 0; i < batches.length; i += opts.concurrency) {
-    const concurrentBatches = batches.slice(i, i + opts.concurrency)
-    const startIndex = chunkIndex
+  if (missIndices.length > 0) {
+    const pipe = await getEmbedPipeline(opts.model)
+    const missChunks = missIndices.map((i) => chunks[i]!)
 
-    const batchResults = await Promise.all(
-      concurrentBatches.map((batch) =>
-        embedBatch(pipe, batch.map((c) => c.text))
+    const batches: number[][] = []
+    for (let i = 0; i < missChunks.length; i += opts.batchSize) {
+      batches.push(missIndices.slice(i, i + opts.batchSize))
+    }
+
+    for (let i = 0; i < batches.length; i += opts.concurrency) {
+      const concurrentBatches = batches.slice(i, i + opts.concurrency)
+
+      const batchResults = await Promise.all(
+        concurrentBatches.map((idxBatch) =>
+          embedBatch(pipe, idxBatch.map((ci) => chunks[ci]!.text))
+        )
       )
-    )
-
-    let offset = startIndex
-    for (let b = 0; b < concurrentBatches.length; b++) {
-      const batch = concurrentBatches[b]!
-      const embeddings = batchResults[b]!
-      for (let j = 0; j < batch.length; j++) {
-        const chunk = batch[j]!
-        entries[offset] = {
-          path: chunk.path,
-          text: chunk.text,
-          embedding: embeddings[j]!,
+
+      for (let b = 0; b < concurrentBatches.length; b++) {
+        const idxBatch = concurrentBatches[b]!
+        const embeddings = batchResults[b]!
+        for (let j = 0; j < idxBatch.length; j++) {
+          const ci = idxBatch[j]!
+          const chunk = chunks[ci]!
+          const embedding = embeddings[j]!
+          entries[ci] = { path: chunk.path, text: chunk.text, embedding }
+          setCachedEmbedding(chunk.text, embedding)
         }
-        offset++
+        done += idxBatch.length
+        onProgress?.(done, chunks.length)
       }
-      chunkIndex += batch.length
-      done += batch.length
-      onProgress?.(done, chunks.length)
     }
+
+    flushEmbedCache()
   }
 
   return entries

package/src/labels.ts

@@ -106,7 +106,8 @@ async function chatComplete(
     })
 
     if (!resp.ok) {
-      lastError = new Error(`Copilot API error: HTTP ${resp.status} ${resp.statusText}`)
+      const errBody = await resp.text().catch(() => "")
+      lastError = new Error(`Copilot API error: HTTP ${resp.status} ${resp.statusText} — ${errBody}`)
       // Retry on 5xx
       if (resp.status >= 500) {
         await Bun.sleep(1000 * (attempt + 1))

package/src/main.ts

@@ -15,6 +15,7 @@
 
 import { Command } from "commander"
 import path from "node:path"
+import { createHash } from "node:crypto"
 
 import { discoverFiles, DEFAULT_EXCLUDES, type FsOptions } from "./fs.ts"
 import { chunkFile } from "./tokenize.ts"
@@ -24,22 +25,24 @@ import { buildTree } from "./tree.ts"
 import { clearAuthCache, getCopilotToken } from "./auth.ts"
 import { SemanticNavigatorUI, type ProgressState } from "./ui.ts"
 import type { CopilotConfig } from "./labels.ts"
+import { treeFingerprint, getCachedTree, setCachedTree } from "./cache.ts"
 
 // ---------------------------------------------------------------------------
 // CLI definition
 // ---------------------------------------------------------------------------
 
 const program = new Command()
-  .name("semantic-navigator")
+  .name("ovid")
   .description("Browse a repository's files by semantic meaning")
   .argument("[directory]", "Directory to analyse (default: current working directory)", ".")
-  .option("--completion-model <model>", "Copilot model to use for labelling", "gpt-4o-mini")
+  .option("--completion-model <model>", "Copilot model to use for labelling", "gpt-5-mini")
   .option("--max-files <n>", "Maximum number of files to index", (v) => parseInt(v, 10), 2000)
   .option("--max-file-bytes <n>", "Skip files larger than this many bytes", (v) => parseInt(v, 10), 1_000_000)
   .option("--exclude-glob <pattern...>", "Glob patterns to exclude (repeatable)", DEFAULT_EXCLUDES)
   .option("--read-concurrency <n>", "Concurrent file reads", (v) => parseInt(v, 10), 64)
   .option("--embed-batch-size <n>", "Chunks per embedding batch", (v) => parseInt(v, 10), 32)
   .option("--embed-concurrency <n>", "Concurrent embedding batches", (v) => parseInt(v, 10), 2)
+  .option("--no-cache", "Skip reading from cache; force re-embed and re-label")
   .option("--logout", "Clear cached GitHub / Copilot credentials and exit")
   .helpOption("-h, --help", "Show help")
 
@@ -60,6 +63,7 @@ async function main(): Promise<void> {
     readConcurrency: number
     embedBatchSize: number
     embedConcurrency: number
+    cache: boolean
     logout: boolean | undefined
   }>()
 
@@ -167,6 +171,7 @@ async function main(): Promise<void> {
     model: DEFAULT_EMBEDDING_MODEL,
     batchSize: opts.embedBatchSize,
     concurrency: opts.embedConcurrency,
+    noCache: !opts.cache,
   }
 
   let embedEntriesRaw: EmbedEntry[] | undefined
@@ -184,7 +189,32 @@ async function main(): Promise<void> {
   }
 
   // ---------------------------------------------------------------------------
-  // Step 5: Spectral clustering (CPU-bound, synchronous)
+  // Step 5: Compute tree fingerprint and check tree cache
+  // ---------------------------------------------------------------------------
+
+  // Build a map of relativePath → sha256(content) for all discovered files.
+  const fileHashes = new Map<string, string>()
+  for (const f of resolvedFiles) {
+    fileHashes.set(
+      f.relativePath,
+      createHash("sha256").update(f.content).digest("hex").slice(0, 16)
+    )
+  }
+
+  const fingerprint = treeFingerprint(opts.completionModel, fileHashes)
+  const noCache = !opts.cache
+
+  if (!noCache) {
+    const cached = getCachedTree(fingerprint)
+    if (cached !== null) {
+      ui.setTree(cached)
+      // The UI event loop keeps the process alive until the user presses q/Esc.
+      return
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Step 6: Spectral clustering (CPU-bound, synchronous)
   // ---------------------------------------------------------------------------
 
   ui.updateProgress({
@@ -201,7 +231,7 @@ async function main(): Promise<void> {
   await Bun.sleep(0)
 
   // ---------------------------------------------------------------------------
-  // Step 6: Build labelled tree
+  // Step 7: Build labelled tree
   // ---------------------------------------------------------------------------
 
   ui.updateProgress({
@@ -223,8 +253,11 @@ async function main(): Promise<void> {
   }
   const tree = treeRaw!
 
+  // Persist to tree cache for future runs
+  setCachedTree(fingerprint, tree)
+
   // ---------------------------------------------------------------------------
-  // Step 7: Hand the tree to the UI
+  // Step 8: Hand the tree to the UI
   // ---------------------------------------------------------------------------
 
   ui.setTree(tree)

package/src/tree.ts

@@ -93,25 +93,30 @@ export async function labelNodes(
   const children = splitCluster(cluster)
 
   if (children.length === 1) {
-    // Leaf cluster: label each file individually
+    // Leaf cluster: label each file individually, in batches of 5 to stay
+    // within the model's prompt token limit.
     const entries = cluster.entries
-    let labels = await labelFiles(config, entries)
-
-    // Guard: align label count with entry count (Copilot may return fewer)
-    if (labels.length < entries.length) {
-      const missing = entries.length - labels.length
-      labels = [
-        ...labels,
-        ...Array.from({ length: missing }, () => ({
-          overarchingTheme: "",
-          distinguishingFeature: "",
-          label: "unlabelled",
-        })),
-      ]
+    const BATCH = 5
+    const allLabels: import("./labels.ts").Label[] = []
+    for (let i = 0; i < entries.length; i += BATCH) {
+      const batch = entries.slice(i, i + BATCH)
+      let batchLabels = await labelFiles(config, batch)
+      if (batchLabels.length < batch.length) {
+        const missing = batch.length - batchLabels.length
+        batchLabels = [
+          ...batchLabels,
+          ...Array.from({ length: missing }, () => ({
+            overarchingTheme: "",
+            distinguishingFeature: "",
+            label: "unlabelled",
+          })),
+        ]
+      }
+      allLabels.push(...batchLabels)
     }
 
     return entries.map((entry, i) => ({
-      label: `${entry.path}: ${labels[i]!.label}`,
+      label: `${entry.path}: ${allLabels[i]!.label}`,
       files: [entry.path],
       children: [],
     }))

package/src/ui.ts

@@ -42,6 +42,8 @@ const THEME = {
   statusColor:  "#bb9af7",
 }
 
+const SPINNER_FRAMES = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"]
+
 // ---------------------------------------------------------------------------
 // Flat node model for scroll-based rendering
 // ---------------------------------------------------------------------------
@@ -79,7 +81,7 @@ function nodeKey(tree: Tree, depth: number): string {
 // ---------------------------------------------------------------------------
 
 export interface ProgressState {
-  phase: "reading" | "embedding" | "clustering" | "labelling" | "done"
+  phase: "reading" | "embedding" | "clustering" | "labelling" | "working" | "done"
   done: number
   total: number
   message?: string
@@ -104,6 +106,8 @@ export class SemanticNavigatorUI {
 
   // Progress state (shown before tree is ready)
   private progress: ProgressState = { phase: "reading", done: 0, total: 0 }
+  private spinnerTimer: ReturnType<typeof setInterval> | null = null
+  private spinnerIndex = 0
 
   // Row renderables (reused by rebuildRows)
   private rowRenderables: TextRenderable[] = []
@@ -144,7 +148,7 @@ export class SemanticNavigatorUI {
     })
     this.headerText = new TextRenderable(this.renderer, {
       id: "header-text",
-      content: " semantic-navigator ",
+      content: " ovid ",
       fg: THEME.statusColor,
     })
     headerBox.add(this.headerText)
@@ -277,6 +281,11 @@ export class SemanticNavigatorUI {
 
   updateProgress(state: ProgressState): void {
     this.progress = state
+    if (state.phase === "working") {
+      this.startSpinner()
+    } else {
+      this.stopSpinner()
+    }
     this.renderProgress()
   }
 
@@ -290,6 +299,7 @@ export class SemanticNavigatorUI {
       case "embedding":  phaseLabel = "Embedding"; break
       case "clustering": phaseLabel = "Clustering"; break
       case "labelling":  phaseLabel = "Labelling"; break
+      case "working":    phaseLabel = "Working"; break
       default:           phaseLabel = "Done"; break
     }
 
@@ -297,14 +307,17 @@ export class SemanticNavigatorUI {
       ? `${phaseLabel}: ${done}/${total} (${pct}%)`
       : message ?? phaseLabel
 
-    this.headerText.content = ` semantic-navigator  ${statusLine}`
+    this.headerText.content = ` ovid ${statusLine}`
 
     // Clear rows and show a single status line
     this.clearRows()
     const id = "status-line"
+    const spinner = phase === "working"
+      ? `${SPINNER_FRAMES[this.spinnerIndex % SPINNER_FRAMES.length] ?? "•"} `
+      : ""
     const statusText = new TextRenderable(this.renderer, {
       id,
-      content: `  ${statusLine}…`,
+      content: `  ${spinner}${statusLine}…`,
       fg: THEME.statusColor,
     })
     this.scrollBox.add(statusText)
@@ -312,11 +325,28 @@ export class SemanticNavigatorUI {
     this.rowIds = [id]
   }
 
+  private startSpinner(): void {
+    if (this.spinnerTimer !== null) return
+    this.spinnerTimer = setInterval(() => {
+      if (this.progress.phase !== "working") return
+      this.spinnerIndex = (this.spinnerIndex + 1) % SPINNER_FRAMES.length
+      this.renderProgress()
+    }, 80)
+  }
+
+  private stopSpinner(): void {
+    if (this.spinnerTimer === null) return
+    clearInterval(this.spinnerTimer)
+    this.spinnerTimer = null
+    this.spinnerIndex = 0
+  }
+
   // ---------------------------------------------------------------------------
   // Tree display (after tree is ready)
   // ---------------------------------------------------------------------------
 
   setTree(tree: Tree): void {
+    this.stopSpinner()
     this.tree = tree
 
     // Auto-expand root and its direct children
@@ -326,7 +356,7 @@ export class SemanticNavigatorUI {
     this.renderRows()
 
     this.headerText.content =
-      ` semantic-navigator  ${tree.label} (${tree.files.length} files)`
+      ` ovid ${tree.label} (${tree.files.length} files)`
   }
 
   private rebuildFlatList(): void {
@@ -455,6 +485,7 @@ export class SemanticNavigatorUI {
   // ---------------------------------------------------------------------------
 
   destroy(): void {
+    this.stopSpinner()
     this.renderer.destroy()
   }
 }