codex-session-insights 0.2.2 → 0.2.3

package/lib/llm-insights.js

@@ -598,6 +598,7 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
   const facetModel = options.facetModel || DEFAULT_FACET_MODEL
   const fastSectionModel = options.fastSectionModel || DEFAULT_FAST_SECTION_MODEL
   const insightModel = options.insightModel || DEFAULT_INSIGHT_MODEL
+  const provider = options.provider || DEFAULT_PROVIDER
   const cacheRoot = resolveCacheRoot(options.cacheDir)
   const facetCacheDir = path.join(cacheRoot, 'facets')
   await fs.mkdir(facetCacheDir, { recursive: true })
@@ -619,8 +620,10 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
   let combineSummaryCalls = 0
   let estimatedFacetInputTokens = 0
   let estimatedFacetOutputTokens = 0
-  let estimatedSummaryInputTokens = 0
-  let estimatedSummaryOutputTokens = 0
+  let estimatedChunkSummaryInputTokens = 0
+  let estimatedChunkSummaryOutputTokens = 0
+  let estimatedCombineSummaryInputTokens = 0
+  let estimatedCombineSummaryOutputTokens = 0
   const facetSystemPrompt = buildFacetSystemPrompt(options.lang)
 
   for (const job of uncachedFacetJobs) {
@@ -628,7 +631,7 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
     const transcriptChars = transcript.length
     if (!transcriptChars) {
       estimatedFacetInputTokens += estimateModelInputTokens({
-        provider: options.provider || DEFAULT_PROVIDER,
+        provider,
         systemPrompt: facetSystemPrompt,
         userPrompt: buildFacetExtractionPrompt(job.thread, `${job.thread.title || '(untitled)'}\n${job.thread.firstUserMessage || ''}`.trim(), options.lang),
         schema: FACET_SCHEMA,
@@ -640,7 +643,7 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
 
     if (transcriptChars <= LONG_TRANSCRIPT_THRESHOLD) {
       estimatedFacetInputTokens += estimateModelInputTokens({
-        provider: options.provider || DEFAULT_PROVIDER,
+        provider,
         systemPrompt: facetSystemPrompt,
         userPrompt: buildFacetExtractionPrompt(job.thread, transcript, options.lang),
         schema: FACET_SCHEMA,
@@ -653,28 +656,28 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
     const chunks = chunkText(transcript, TRANSCRIPT_CHUNK_SIZE)
     chunkSummaryCalls += chunks.length
     for (const chunk of chunks) {
-      estimatedSummaryInputTokens += estimateModelInputTokens({
-        provider: options.provider || DEFAULT_PROVIDER,
+      estimatedChunkSummaryInputTokens += estimateModelInputTokens({
+        provider,
         systemPrompt: `${FACET_TRANSCRIPT_SUMMARY_DIRECTIVE}\n\nPreserve user goal, outcome, friction, command/tool issues, and what the assistant actually achieved.`,
         userPrompt: `Chunk 1 of ${chunks.length}\n\n${chunk}`,
         structured: false,
       })
-      estimatedSummaryOutputTokens += 260
+      estimatedChunkSummaryOutputTokens += 260
     }
     const combinedSummaryChars = chunks.length * 1100
     if (combinedSummaryChars > LONG_TRANSCRIPT_THRESHOLD) {
       combineSummaryCalls += 1
-      estimatedSummaryInputTokens += estimateModelInputTokens({
-        provider: options.provider || DEFAULT_PROVIDER,
+      estimatedCombineSummaryInputTokens += estimateModelInputTokens({
+        provider,
         systemPrompt:
           'Combine these coding-session chunk summaries into one compact transcript summary. Keep only material signal for later facet extraction. Do not carry boilerplate, stack traces, or command details.',
         userPrompt: makePlaceholderText(combinedSummaryChars, 'Chunk summaries'),
         structured: false,
       })
-      estimatedSummaryOutputTokens += 320
+      estimatedCombineSummaryOutputTokens += 320
     }
     estimatedFacetInputTokens += estimateModelInputTokens({
-      provider: options.provider || DEFAULT_PROVIDER,
+      provider,
       systemPrompt: facetSystemPrompt,
       userPrompt: buildFacetExtractionPrompt(
         job.thread,
@@ -693,14 +696,25 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
   const estimatedSectionInputs = estimateSectionInputs(candidateThreads, facetJobs, options)
   const fastSectionCalls = SECTION_DEFS.filter(section => section.modelTier === 'fast').length
   const fullSectionCalls = SECTION_DEFS.filter(section => section.modelTier !== 'fast').length
-  const estimatedFastSectionInputTokens = SECTION_DEFS.filter(section => section.modelTier === 'fast')
+  let estimatedFastSectionInputTokens = SECTION_DEFS.filter(section => section.modelTier === 'fast')
     .reduce((sum, section) => sum + estimatedSectionInputs[section.contextKind] + 500, 0)
   const estimatedFastSectionOutputTokens = fastSectionCalls * 500
-  const estimatedFullSectionInputTokens = SECTION_DEFS.filter(section => section.modelTier !== 'fast')
+  let estimatedFullSectionInputTokens = SECTION_DEFS.filter(section => section.modelTier !== 'fast')
     .reduce((sum, section) => sum + estimatedSectionInputs[section.contextKind] + 650, 0)
   const estimatedFullSectionOutputTokens = fullSectionCalls * 700
-  const estimatedAtAGlanceInputTokens = estimatedSectionInputs.at_a_glance + 2200
+  let estimatedAtAGlanceInputTokens = estimatedSectionInputs.at_a_glance + 2200
   const estimatedAtAGlanceOutputTokens = 260
+  const estimatedSummaryInputTokens =
+    estimatedChunkSummaryInputTokens + estimatedCombineSummaryInputTokens
+  const estimatedSummaryOutputTokens =
+    estimatedChunkSummaryOutputTokens + estimatedCombineSummaryOutputTokens
+
+  estimatedFacetInputTokens += estimateCodexCliFreshOverhead(provider, facetModel, uncachedFacetJobs.length)
+  estimatedChunkSummaryInputTokens += estimateCodexCliFreshOverhead(provider, facetModel, chunkSummaryCalls)
+  estimatedCombineSummaryInputTokens += estimateCodexCliFreshOverhead(provider, facetModel, combineSummaryCalls)
+  estimatedFastSectionInputTokens += estimateCodexCliFreshOverhead(provider, fastSectionModel, fastSectionCalls)
+  estimatedFullSectionInputTokens += estimateCodexCliFreshOverhead(provider, insightModel, fullSectionCalls)
+  estimatedAtAGlanceInputTokens += estimateCodexCliFreshOverhead(provider, insightModel, 1)
 
   const byStage = [
     buildEstimateBucket(
@@ -713,16 +727,16 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
     buildEstimateBucket(
       'transcript_summary:chunk',
       chunkSummaryCalls,
-      estimatedSummaryInputTokens - (combineSummaryCalls ? estimateTokensFromChars(combineSummaryCalls * 1100) + combineSummaryCalls * 180 : 0),
+      estimatedChunkSummaryInputTokens,
       0,
-      estimatedSummaryOutputTokens - combineSummaryCalls * 320,
+      estimatedChunkSummaryOutputTokens,
     ),
     buildEstimateBucket(
       'transcript_summary:combine',
       combineSummaryCalls,
-      combineSummaryCalls ? estimateTokensFromChars(combineSummaryCalls * 1100) + combineSummaryCalls * 180 : 0,
+      estimatedCombineSummaryInputTokens,
       0,
-      combineSummaryCalls * 320,
+      estimatedCombineSummaryOutputTokens,
     ),
     buildEstimateBucket(
       'section:fast',
@@ -822,6 +836,14 @@ export async function estimateLlmAnalysisCost({ threadSummaries, options = {} })
   }
 }
 
+function estimateCodexCliFreshOverhead(provider, model, calls) {
+  if (provider !== 'codex-cli' || !calls) return 0
+  const normalized = String(model || '').trim()
+  if (normalized === 'gpt-5.4') return calls * 25_000
+  if (normalized === 'gpt-5.4-mini' || normalized === 'gpt-5.3-codex-spark') return calls * 4_500
+  return calls * 8_000
+}
+
 async function planFacetJobs(threadSummaries, { cacheDir, model, uncachedLimit }) {
   const jobs = []
   let uncachedCount = 0

package/package.json

@@ -1,10 +1,10 @@
 {
   "name": "codex-session-insights",
-  "version": "0.2.2",
+  "version": "0.2.3",
   "description": "Generate a report analyzing your Codex sessions.",
   "type": "module",
   "bin": {
-    "codex-session-insights": "./bin/codex-insights.js"
+    "codex-session-insights": "bin/codex-insights.js"
   },
   "files": [
     "bin",