kimaki 0.30.0 → 0.30.1

package/src/session-handler/event-stream-state.ts

@@ -879,7 +879,8 @@ const MIN_PROMPT_CACHE_READ_TO_TRACK = 1024
 const PROMPT_CACHE_DROP_RATIO = 0.5
 
 export type PromptCacheClear = {
-  previousCacheRead: number
+  // Tokens that should have been read from cache: the smaller of the previous cached prefix and the current prompt.
+  expectedCacheRead: number
   currentCacheRead: number
   previousMessageId: string
   currentMessageId: string
@@ -915,13 +916,6 @@ function hasPruneBetween({
   return false
 }
 
-function promptInputTokens(tokens: {
-  input: number
-  cache: { read: number; write: number }
-}): number {
-  return tokens.input + tokens.cache.read + tokens.cache.write
-}
-
 function getCompletedAssistantAt({
   events,
   sessionId,
@@ -963,7 +957,8 @@ function isComparableCacheAssistant(message: AssistantMessage): boolean {
   return true
 }
 
-// Same-model cache drop vs the previous completed assistant. Errors and compaction summaries block the pair.
+// Same-model cache drop vs the previous completed assistant. Aborted/errored replies are skipped,
+// compaction summaries and pruned tool output block the pair since they rewrite the prompt.
 export function getPromptCacheClear({
   events,
   sessionId,
@@ -1003,13 +998,13 @@ export function getPromptCacheClear({
       continue
     }
     seen.add(info.id)
-    if (!isComparableCacheAssistant(info) || !info.tokens) {
+    if (!isUserFacingAssistantMessage(info)) {
       return undefined
     }
-    if (info.modelID !== current.modelID || info.providerID !== current.providerID) {
-      return undefined
+    if (!isComparableCacheAssistant(info) || !info.tokens) {
+      continue
     }
-    if (info.tokens.cache.read < MIN_PROMPT_CACHE_READ_TO_TRACK) {
+    if (info.modelID !== current.modelID || info.providerID !== current.providerID) {
       return undefined
     }
     if (hasPruneBetween({
@@ -1020,17 +1015,19 @@ export function getPromptCacheClear({
     })) {
       return undefined
     }
-    if (current.tokens.cache.read > info.tokens.cache.read * PROMPT_CACHE_DROP_RATIO) {
+    // Anthropic reports a fresh cache as write only, so read alone misses the turn after a miss.
+    const previousCached = info.tokens.cache.read + info.tokens.cache.write
+    const currentPrompt = current.tokens.input + current.tokens.cache.read + current.tokens.cache.write
+    // A reverted (shorter) prompt can only reuse its own length from cache.
+    const expectedCacheRead = Math.min(previousCached, currentPrompt)
+    if (expectedCacheRead < MIN_PROMPT_CACHE_READ_TO_TRACK) {
       return undefined
     }
-    const previousPrompt = promptInputTokens(info.tokens)
-    const currentPrompt = promptInputTokens(current.tokens)
-    // Pruning blanks old tool output and shrinks the next prompt. That is not a cache clear.
-    if (currentPrompt < previousPrompt) {
+    if (current.tokens.cache.read > expectedCacheRead * PROMPT_CACHE_DROP_RATIO) {
       return undefined
     }
     return {
-      previousCacheRead: info.tokens.cache.read,
+      expectedCacheRead,
       currentCacheRead: current.tokens.cache.read,
       previousMessageId: info.id,
       currentMessageId: current.id,
@@ -1052,7 +1049,7 @@ export function formatPromptCacheClearMessage(
   clear: PromptCacheClear,
   systemDiff?: { additions: number; deletions: number },
 ): string {
-  const tokens = `prompt cache missed (${formatCompactTokenCount(clear.previousCacheRead)} → ${formatCompactTokenCount(clear.currentCacheRead)})`
+  const tokens = `prompt cache missed (${formatCompactTokenCount(clear.expectedCacheRead)} → ${formatCompactTokenCount(clear.currentCacheRead)})`
   if (!systemDiff || (systemDiff.additions === 0 && systemDiff.deletions === 0)) {
     return tokens
   }

package/src/session-handler/thread-session-runtime.ts

@@ -50,7 +50,7 @@ import {
 import type { DiscordFileAttachment, SessionPartKind } from '../message-formatting.js'
 import {
   asDiscordQuote,
-  asSystemLine,
+  asSubtext,
   formatPart,
   formatTaskToolTitle,
   planAssistantTurnFlush,
@@ -2349,16 +2349,15 @@ export class ThreadSessionRuntime {
       upToIndex: this.eventBuffer.length - 2,
     })
     const completedAt = msg.time.completed
-    if (
-      !wasAlreadyCompleted
-      && typeof completedAt === 'number'
-      && isAssistantMessageNaturalCompletion({ message: msg })
-    ) {
-      await this.handleNaturalAssistantCompletion({
-        completedMessageId: msg.id,
-        completedAt,
-      })
-      return
+    if (!wasAlreadyCompleted && typeof completedAt === 'number') {
+      if (isAssistantMessageNaturalCompletion({ message: msg })) {
+        await this.handleNaturalAssistantCompletion({
+          completedMessageId: msg.id,
+          completedAt,
+        })
+        return
+      }
+      await this.maybeNotifyPromptCacheClear({ sessionId, messageId: msg.id })
     }
 
     // Context usage notice.
@@ -2398,7 +2397,7 @@ export class ThreadSessionRuntime {
       return
     }
     this.lastDisplayedContextPercentage = thresholdCrossed
-    const chunk = `${STATUS_PREFIX}context usage ${currentPercentage}%`
+    const chunk = asSubtext(`context usage ${currentPercentage}%`)
     const sendResult = await this.thread.send({ content: chunk, flags: SILENT_MESSAGE_FLAGS })
       .catch((e) => new DiscordOperationError({ operation: 'sendMessage', cause: e }))
     if (sendResult instanceof Error) {
@@ -2615,7 +2614,7 @@ export class ThreadSessionRuntime {
             }
             return ` (${pct.toFixed(1)}%)`
           })()
-          const chunk = `${STATUS_PREFIX}${part.tool} returned ${formattedTokens} tokens${percentageSuffix}`
+          const chunk = asSubtext(`${STATUS_PREFIX}${part.tool} returned ${formattedTokens} tokens${percentageSuffix}`)
           const largeOutputResult = await this.thread.send({
             content: chunk,
             flags: SILENT_MESSAGE_FLAGS,
@@ -3285,7 +3284,7 @@ export class ThreadSessionRuntime {
       return secs > 0 ? `${mins}m ${secs}s` : `${mins}m`
     })()
 
-    const chunk = `${STATUS_PREFIX}${message} - retrying in ${duration} (attempt #${attempt})`
+    const chunk = asSubtext(`${message} - retrying in ${duration} (attempt #${attempt})`)
     const retryResult = await this.thread.send({ content: chunk, flags: SILENT_MESSAGE_FLAGS })
       .catch((e) => new DiscordOperationError({ operation: 'sendMessage', cause: e }))
     if (retryResult instanceof Error) {
@@ -3371,7 +3370,7 @@ export class ThreadSessionRuntime {
     const titlePrefix = properties.title
       ? `${properties.title.trim()}: `
       : ''
-    const chunk = `${STATUS_PREFIX}${properties.variant}: ${titlePrefix}${toastMessage}`
+    const chunk = asSubtext(`${properties.variant}: ${titlePrefix}${toastMessage}`)
     const toastResult = await this.thread.send({ content: chunk, flags: SILENT_MESSAGE_FLAGS })
       .catch((e) => new DiscordOperationError({ operation: 'sendMessage', cause: e }))
     if (toastResult instanceof Error) {
@@ -4051,17 +4050,20 @@ export class ThreadSessionRuntime {
     }
   }
 
-  abortActiveRun(reason: string): void {
+  /** Explicit user abort: stops the run and drops all queued messages. Returns the removed items. */
+  async abortActiveRun(reason: string): Promise<threadState.QueuedMessage[]> {
     const outcome = this.abortActiveRunInternal({
       reason,
     })
     if (outcome.apiAbortPromise) {
       void outcome.apiAbortPromise
     }
-    // Drain local queued messages after explicit abort.
-    void this.dispatchAction(() => {
-      return this.tryDrainQueue({ showIndicator: true })
+    // Enqueued synchronously, so it runs before the abort's session.idle can drain the queue.
+    let cleared: threadState.QueuedMessage[] = []
+    await this.dispatchAction(async () => {
+      cleared = await this.clearQueueNow()
     })
+    return cleared
   }
 
   async abortActiveRunAndWait({
@@ -4112,20 +4114,25 @@ export class ThreadSessionRuntime {
   async clearQueue(): Promise<threadState.QueuedMessage[]> {
     let cleared: threadState.QueuedMessage[] = []
     await this.dispatchAction(async () => {
-      const persistResult = await deleteThreadQueueItems(this.threadId).catch((error) => {
-        return new Error('Failed to clear persisted queue', { cause: error })
-      })
-      if (persistResult instanceof Error) {
-        logger.error(
-          `[QUEUE] Failed to clear persisted queue for thread ${this.threadId}: ${persistResult.message}`,
-        )
-        return
-      }
-      cleared = threadState.clearQueueItems(this.threadId)
+      cleared = await this.clearQueueNow()
     })
     return cleared
   }
 
+  // Must run inside dispatchAction.
+  private async clearQueueNow(): Promise<threadState.QueuedMessage[]> {
+    const persistResult = await deleteThreadQueueItems(this.threadId).catch((error) => {
+      return new Error('Failed to clear persisted queue', { cause: error })
+    })
+    if (persistResult instanceof Error) {
+      logger.error(
+        `[QUEUE] Failed to clear persisted queue for thread ${this.threadId}: ${persistResult.message}`,
+      )
+      return []
+    }
+    return threadState.clearQueueItems(this.threadId)
+  }
+
   /** Remove a queued message by its 1-based position. */
   async removeQueuePosition(position: number): Promise<threadState.QueuedMessage | undefined> {
     let removed: threadState.QueuedMessage | undefined
@@ -5174,7 +5181,7 @@ export class ThreadSessionRuntime {
       : ''
     const result = await sendThreadMessage(
       this.thread,
-      asSystemLine(`*using ${modelLabel}${agentLabel}*`),
+      asSubtext(`*using ${modelLabel}${agentLabel}*`),
       { flags: SILENT_MESSAGE_FLAGS },
     ).catch((e) => new DiscordOperationError({ operation: 'sendMessage', cause: e }))
     if (result instanceof Error) {
@@ -5327,7 +5334,7 @@ export class ThreadSessionRuntime {
         })
       : undefined
     const mention = mentionUserId ? ` <@${mentionUserId}>` : ''
-    const footerText = asSystemLine(
+    const footerText = asSubtext(
       `*${projectInfo}${sessionDuration}${contextInfo}${modelInfo}${agentInfo}*${mention}`,
     )
     this.stopTyping()
@@ -5356,6 +5363,11 @@ export class ThreadSessionRuntime {
     sessionId: string
     messageId: string
   }): Promise<void> {
+    // Only the first reply after a user prompt can show a cold cache. Later steps re-read the turn's own writes.
+    const [firstAssistantId] = this.getAssistantMessageIdsForCurrentTurn({ sessionId })
+    if (firstAssistantId !== messageId) {
+      return
+    }
     const cacheClear = getPromptCacheClear({
       events: this.eventBuffer,
       sessionId,
@@ -5369,7 +5381,7 @@ export class ThreadSessionRuntime {
       previousMessageId: cacheClear.previousMessageId,
       currentMessageId: cacheClear.currentMessageId,
     })
-    const chunk = `${STATUS_PREFIX}${formatPromptCacheClearMessage(cacheClear, systemDiff)}`
+    const chunk = asSubtext(formatPromptCacheClearMessage(cacheClear, systemDiff))
     const sendResult = await this.thread.send({ content: chunk, flags: SILENT_MESSAGE_FLAGS })
       .catch((e) => new DiscordOperationError({ operation: 'sendMessage', cause: e }))
     if (sendResult instanceof Error) {

package/src/sleep.e2e.test.ts

@@ -347,7 +347,7 @@ describe('kimaki_sleep', () => {
         -# *using deterministic-provider/deterministic-v2*
         > waiting for the deploy
 
-        ┣ kimaki_sleep until 2030-01-01T09:00:00Z _waiting for the deploy_
+        -# ┣ kimaki_sleep until 2030-01-01T09:00:00Z _waiting for the deploy_
 
         > sleep-started
         -# *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
@@ -466,7 +466,7 @@ describe('kimaki_sleep', () => {
         SLEEP_CANCEL_MARKER wait for something
         --- from: assistant (TestBot)
         -# *using deterministic-provider/deterministic-v2*
-        ┣ kimaki_sleep for 2h _cancel me_
+        -# ┣ kimaki_sleep for 2h _cancel me_
 
         > cancel-sleep-started
         -# *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*

package/src/task-runner.ts

@@ -2,7 +2,7 @@
 
 import { DiscordAPIError, type REST, Routes } from 'discord.js'
 import { createDiscordRest } from './discord-urls.js'
-import { ensureThreadMember } from './discord-utils.js'
+import { buildThreadStartEmbeds, ensureThreadMember } from './discord-utils.js'
 import YAML from 'yaml'
 import {
   claimScheduledTaskRunning,
@@ -97,7 +97,7 @@ async function executeThreadScheduledTask({
       : {}),
     ...(payload.parentSessionId ? { parentSessionId: payload.parentSessionId } : {}),
   }
-  const embed = [{ color: 0x2b2d31, footer: { text: YAML.stringify(marker) } }]
+  const embed = await buildThreadStartEmbeds(marker)
   // Newline between prefix and prompt so leading /command detection can
   // find the command on its own line.
   const prefixedPrompt = `${QUEUE_PREFIX}**kimaki-cli:**\n${prompt}`
@@ -296,7 +296,7 @@ async function executeChannelScheduledTask({
         ...(payload.parentSessionId ? { parentSessionId: payload.parentSessionId } : {}),
       }
   const embeds = marker
-    ? [{ color: 0x2b2d31, footer: { text: YAML.stringify(marker) } }]
+    ? await buildThreadStartEmbeds(marker)
     : undefined
 
   const starterResult = await rest

package/src/task-schedule.ts

@@ -2,7 +2,7 @@
 
 import { CronExpressionParser } from 'cron-parser'
 import * as errore from 'errore'
-import { STATUS_PREFIX } from './message-formatting.js'
+import { asSubtext } from './message-formatting.js'
 
 export type ScheduledTaskPayload = {
   preRunCommand: string | null
@@ -194,7 +194,7 @@ export function formatSessionSleepWakePrompt({
 }): string {
   const until = formatSessionSleepWakeAt(wakeAt)
   const reasonLine = reason?.trim() ? `\nReason: ${reason.trim()}` : ''
-  return `${STATUS_PREFIX}Woke after sleeping until ${until}${reasonLine}\nContinue the work you were waiting for.`
+  return `${asSubtext(`Woke after sleeping until ${until}`)}${reasonLine}\nContinue the work you were waiting for.`
 }
 
 // Tool result after persist. Keep "Sleeping until" first so e2e matchers still hit.

package/src/thread-message-queue.e2e.test.ts

@@ -714,7 +714,7 @@ e2eTest('thread message queue ordering', () => {
         > running create file
         > creating marker
 
-        ┣ bash _Create marker file for e2e test_
+        -# ┣ bash _Create marker file for e2e test_
 
         > file created
         -# *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"

package/src/undo-redo.e2e.test.ts

@@ -202,7 +202,7 @@ e2eTest('/undo sets revert state and cleans up on next prompt', () => {
         -# *using deterministic-provider/deterministic-v2*
         > creating undo file
 
-        ┣ bash _Create undo marker file_
+        -# ┣ bash _Create undo marker file_
 
         > undo file created
         -# *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*