mini-coder 0.5.4 → 0.5.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mini-coder",
-  "version": "0.5.4",
+  "version": "0.5.5",
   "description": "A small, fast CLI coding agent",
   "module": "src/index.ts",
   "type": "module",

package/src/index.ts

@@ -47,6 +47,7 @@ import {
 } from "./prompt.ts";
 import {
   appendMessage,
+  computeContextTokens,
   computeStats,
   createSession,
   filterModelMessages,
@@ -526,6 +527,8 @@ export interface AppState {
   messages: ReturnType<typeof loadMessages>;
   /** Cumulative session input/output/cost stats for the status bar. */
   stats: SessionStats;
+  /** Estimated model-visible context tokens for the next request. */
+  contextTokens: number;
   /** Discovered AGENTS.md files. */
   agentsMd: AgentsMdFile[];
   /** Discovered skills. */
@@ -610,6 +613,7 @@ export async function init(): Promise<AppState> {
   const effort = startup.effort;
   const messages: ReturnType<typeof loadMessages> = [];
   const stats = computeStats(messages);
+  const contextTokens = computeContextTokens(messages);
   const promptContext = await loadPromptContext(filterModelMessages(messages), {
     cwd,
   });
@@ -621,6 +625,7 @@ export async function init(): Promise<AppState> {
     effort,
     messages,
     stats,
+    contextTokens,
     agentsMd: promptContext.agentsMd,
     skills: promptContext.skills,
     plugins: promptContext.plugins,

package/src/session.ts

@@ -107,7 +107,7 @@ export interface UiMessage {
 }
 
 /** Any message persisted in session history. */
-type PersistedMessage = Message | UiMessage;
+export type PersistedMessage = Message | UiMessage;
 
 /** Options for creating a new session. */
 interface CreateSessionOpts {
@@ -1071,3 +1071,153 @@ export function computeStats(
 
   return stats;
 }
+
+// ---------------------------------------------------------------------------
+// Context estimation
+// ---------------------------------------------------------------------------
+
+/** Conservative fixed estimate for an image block's token footprint. */
+const ESTIMATED_IMAGE_TOKENS = 1_200;
+
+/** Calculate context tokens from assistant usage, falling back when `totalTokens` is zero. */
+function calculateUsageTokens(usage: AssistantMessage["usage"]): number {
+  return (
+    usage.totalTokens ||
+    usage.input + usage.output + usage.cacheRead + usage.cacheWrite
+  );
+}
+
+/** Estimate token usage from a character count using a conservative chars/4 heuristic. */
+function estimateCharacterTokens(charCount: number): number {
+  return Math.ceil(charCount / 4);
+}
+
+type UserMultipartContent = Exclude<
+  Extract<Message, { role: "user" }>["content"],
+  string
+>;
+type TextOrImageContentBlock =
+  | UserMultipartContent[number]
+  | Extract<Message, { role: "toolResult" }>["content"][number];
+
+function estimateTextOrImageContentTokens(
+  content: readonly TextOrImageContentBlock[],
+): number {
+  let chars = 0;
+  let imageTokens = 0;
+
+  for (const block of content) {
+    if (block.type === "text") {
+      chars += block.text.length;
+      continue;
+    }
+    if (block.type === "image") {
+      imageTokens += ESTIMATED_IMAGE_TOKENS;
+    }
+  }
+
+  return estimateCharacterTokens(chars) + imageTokens;
+}
+
+function estimateUserMessageTokens(
+  message: Extract<Message, { role: "user" }>,
+): number {
+  if (typeof message.content === "string") {
+    return estimateCharacterTokens(message.content.length);
+  }
+  return estimateTextOrImageContentTokens(message.content);
+}
+
+function estimateAssistantBlockCharacters(
+  block: Extract<Message, { role: "assistant" }>["content"][number],
+): number {
+  if (block.type === "text") {
+    return block.text.length;
+  }
+  if (block.type === "thinking") {
+    return block.thinking.length;
+  }
+  return block.name.length + JSON.stringify(block.arguments).length;
+}
+
+function estimateAssistantMessageTokens(
+  message: Extract<Message, { role: "assistant" }>,
+): number {
+  const chars = message.content.reduce((total, block) => {
+    return total + estimateAssistantBlockCharacters(block);
+  }, 0);
+  return estimateCharacterTokens(chars);
+}
+
+function estimateToolResultMessageTokens(
+  message: Extract<Message, { role: "toolResult" }>,
+): number {
+  return estimateTextOrImageContentTokens(message.content);
+}
+
+/** Estimate token usage for a model-visible message. */
+function estimateMessageTokens(message: Message): number {
+  switch (message.role) {
+    case "user":
+      return estimateUserMessageTokens(message);
+    case "assistant":
+      return estimateAssistantMessageTokens(message);
+    case "toolResult":
+      return estimateToolResultMessageTokens(message);
+  }
+}
+
+/**
+ * Fold one persisted message into the running context estimate for the next request.
+ *
+ * Assistant messages with valid usage anchor the full model-visible context for
+ * that point in the transcript, so they replace the running estimate. All other
+ * model-visible messages are added incrementally using the same conservative
+ * estimation logic used before the first valid assistant usage appears.
+ *
+ * @param contextTokens - Running estimate before this message.
+ * @param message - Persisted message to fold into the estimate.
+ * @returns Updated context-token estimate.
+ */
+export function addMessageToContextTokens(
+  contextTokens: number,
+  message: PersistedMessage,
+): number {
+  if (message.role === "ui") {
+    return contextTokens;
+  }
+
+  const usage = getAssistantUsage(message);
+  if (
+    message.role === "assistant" &&
+    usage &&
+    message.stopReason !== "aborted" &&
+    message.stopReason !== "error"
+  ) {
+    return calculateUsageTokens(usage);
+  }
+
+  return contextTokens + estimateMessageTokens(message);
+}
+
+/**
+ * Estimate the current model-visible context size for the next request.
+ *
+ * Recomputed on session-load boundaries and maintained incrementally during an
+ * active turn so render-time status-bar updates do not need to rescan the full
+ * message history.
+ *
+ * @param messages - Full persisted session history.
+ * @returns Estimated context tokens visible to the next model request.
+ */
+export function computeContextTokens(
+  messages: readonly PersistedMessage[],
+): number {
+  let contextTokens = 0;
+
+  for (const message of messages) {
+    contextTokens = addMessageToContextTokens(contextTokens, message);
+  }
+
+  return contextTokens;
+}

package/src/submit.ts

@@ -19,6 +19,7 @@ import {
 } from "./index.ts";
 import { parseInput } from "./input.ts";
 import {
+  addMessageToContextTokens,
   addMessageToStats,
   appendMessage,
   appendPromptHistory,
@@ -212,9 +213,17 @@ function handleAgentEvent(event: AgentEvent, state: AppState): void {
     case "assistant_message":
       state.messages.push(event.message);
       state.stats = addMessageToStats(state.stats, event.message);
+      state.contextTokens = addMessageToContextTokens(
+        state.contextTokens,
+        event.message,
+      );
       break;
     case "tool_result":
       state.messages.push(event.message);
+      state.contextTokens = addMessageToContextTokens(
+        state.contextTokens,
+        event.message,
+      );
       break;
     case "text_delta":
     case "thinking_delta":
@@ -277,6 +286,10 @@ export async function submitResolvedInput(
 
   const turn = appendMessage(state.db, session.id, userMessage);
   state.messages.push(userMessage);
+  state.contextTokens = addMessageToContextTokens(
+    state.contextTokens,
+    userMessage,
+  );
   hooks?.onUserMessage?.(state);
 
   state.git = await loadGitState(state.cwd);

package/src/ui/commands.test.ts

@@ -8,6 +8,7 @@ import type { AppState } from "../index.ts";
 import { COMMANDS } from "../input.ts";
 import {
   appendPromptHistory,
+  computeContextTokens,
   createSession,
   openDatabase,
 } from "../session.ts";
@@ -75,6 +76,7 @@ function createTestState(): AppState {
     effort: "medium",
     messages: [],
     stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
+    contextTokens: 0,
     agentsMd: [],
     skills: [],
     plugins: [],
@@ -377,6 +379,7 @@ describe("ui/commands", () => {
         totalOutput: 0,
         totalCost: 0,
       });
+      expect(state.contextTokens).toBe(computeContextTokens(state.messages));
     } finally {
       state.db.close();
     }
@@ -436,6 +439,7 @@ describe("ui/commands", () => {
       { role: "ui", kind: "info", content: "old", timestamp: 1 },
     ];
     state.stats = { totalInput: 10, totalOutput: 20, totalCost: 0.5 };
+    state.contextTokens = 123;
 
     try {
       expect(controller.handleCommand("new", state)).toBe(true);
@@ -449,6 +453,7 @@ describe("ui/commands", () => {
         totalOutput: 0,
         totalCost: 0,
       });
+      expect(state.contextTokens).toBe(0);
       expect(state.agentsMd).toEqual([
         { path: "/tmp/reloaded/AGENTS.md", content: "Reloaded context" },
       ]);

package/src/ui/commands.ts

@@ -19,6 +19,7 @@ import type { AppState } from "../index.ts";
 import { getAvailableModels, saveOAuthCredentials } from "../index.ts";
 import { COMMANDS } from "../input.ts";
 import {
+  computeContextTokens,
   computeStats,
   forkSession,
   listPromptHistory,
@@ -387,6 +388,7 @@ export function createCommandController(
             state.session = picked;
             state.messages = loadMessages(state.db, picked.id);
             state.stats = computeStats(state.messages);
+            state.contextTokens = computeContextTokens(state.messages);
             runtime.scrollConversationToBottom();
           }
         }
@@ -402,6 +404,7 @@ export function createCommandController(
     state.session = null;
     state.messages = [];
     state.stats = { totalInput: 0, totalOutput: 0, totalCost: 0 };
+    state.contextTokens = 0;
     await runtime.reloadPromptContext(state);
     runtime.scrollConversationToBottom();
     runtime.render();
@@ -415,6 +418,7 @@ export function createCommandController(
     state.session = forked;
     state.messages = loadMessages(state.db, forked.id);
     state.stats = computeStats(state.messages);
+    state.contextTokens = computeContextTokens(state.messages);
     runtime.appendInfoMessage("Forked session.", state);
   };
 
@@ -434,6 +438,7 @@ export function createCommandController(
     if (removed) {
       state.messages = loadMessages(state.db, state.session.id);
       state.stats = computeStats(state.messages);
+      state.contextTokens = computeContextTokens(state.messages);
       runtime.scrollConversationToBottom();
       runtime.render();
     }

package/src/ui/status.test.ts

@@ -4,7 +4,11 @@ import {
   registerFauxProvider,
 } from "@mariozechner/pi-ai";
 import type { AppState } from "../index.ts";
-import { createUiMessage, openDatabase } from "../session.ts";
+import {
+  computeContextTokens,
+  createUiMessage,
+  openDatabase,
+} from "../session.ts";
 import { DEFAULT_SHOW_REASONING, DEFAULT_VERBOSE } from "../settings.ts";
 import { DEFAULT_THEME } from "../theme.ts";
 import { renderStatusBar } from "./status.ts";
@@ -87,6 +91,7 @@ function createTestState(): AppState {
     effort: "medium",
     messages: [],
     stats: { totalInput: 0, totalOutput: 0, totalCost: 0 },
+    contextTokens: 0,
     agentsMd: [],
     skills: [],
     plugins: [],
@@ -109,6 +114,11 @@ function createTestState(): AppState {
   };
 }
 
+function setMessages(state: AppState, messages: AppState["messages"]): void {
+  state.messages = messages;
+  state.contextTokens = computeContextTokens(messages);
+}
+
 function makeAssistantWithUsage(
   text: string,
   usage: Partial<ReturnType<typeof fauxAssistantMessage>["usage"]>,
@@ -217,11 +227,11 @@ describe("ui/status", () => {
       ] as const;
 
       for (const testCase of cases) {
-        state.messages = [
+        setMessages(state, [
           makeAssistantWithUsage("context anchor", {
             totalTokens: testCase.totalTokens,
           }),
-        ];
+        ]);
         expect(getUsagePill(state)).toMatchObject({
           text: `in:0 out:0 · ${testCase.totalTokens.toFixed(1)}%/100 · $0.00`,
           bgColor: testCase.bgColor,
@@ -234,6 +244,24 @@ describe("ui/status", () => {
     }
   });
 
+  test("renderStatusBar uses cached contextTokens instead of rescanning messages", () => {
+    const faux = registerFauxProvider();
+    const state = createTestState();
+    state.model = {
+      ...faux.getModel(),
+      contextWindow: 100,
+    };
+    state.messages = [{ role: "user", content: "x".repeat(400), timestamp: 1 }];
+    state.contextTokens = 25;
+
+    try {
+      expect(getUsageSummary(state)).toBe("in:0 out:0 · 25.0%/100 · $0.00");
+    } finally {
+      faux.unregister();
+      state.db.close();
+    }
+  });
+
   test("renderStatusBar colors the model and usage pills independently", () => {
     const faux = registerFauxProvider();
     const state = createTestState();
@@ -244,9 +272,9 @@ describe("ui/status", () => {
 
     try {
       state.effort = "xhigh";
-      state.messages = [
+      setMessages(state, [
         makeAssistantWithUsage("cold context", { totalTokens: 10 }),
-      ];
+      ]);
       expect(getModelPill(state)).toMatchObject({
         bgColor: "color09",
         fgColor: "color00",
@@ -257,9 +285,9 @@ describe("ui/status", () => {
       });
 
       state.effort = "low";
-      state.messages = [
+      setMessages(state, [
         makeAssistantWithUsage("hot context", { totalTokens: 95 }),
-      ];
+      ]);
       expect(getModelPill(state)).toMatchObject({
         bgColor: "color02",
         fgColor: "color00",
@@ -288,26 +316,29 @@ describe("ui/status", () => {
     };
 
     try {
-      state.messages = [
+      setMessages(state, [
         { role: "user", content: "first", timestamp: 1 },
         makeAssistantWithUsage("Second.", {
           input: 200,
           output: 50,
           totalTokens: 250,
         }),
-      ];
+      ]);
       const anchoredPercent = getUsagePercent(state);
 
-      state.messages = [...state.messages, createUiMessage("x".repeat(5_000))];
+      setMessages(state, [
+        ...state.messages,
+        createUiMessage("x".repeat(5_000)),
+      ]);
       const withUiOnlyPercent = getUsagePercent(state);
 
-      state.messages = [
+      setMessages(state, [
         ...state.messages,
         { role: "user", content: "12345678", timestamp: 2 },
-      ];
+      ]);
       const withTrailingUserPercent = getUsagePercent(state);
 
-      state.messages = [
+      setMessages(state, [
         ...state.messages,
         {
           role: "toolResult",
@@ -317,7 +348,7 @@ describe("ui/status", () => {
           isError: false,
           timestamp: 3,
         },
-      ];
+      ]);
       const withToolResultPercent = getUsagePercent(state);
 
       expect(getUsageSummary(state)).toBe(
@@ -341,16 +372,19 @@ describe("ui/status", () => {
     };
 
     try {
-      state.messages = [{ role: "user", content: "12345678", timestamp: 1 }];
+      setMessages(state, [{ role: "user", content: "12345678", timestamp: 1 }]);
       const initialPercent = getUsagePercent(state);
 
-      state.messages = [createUiMessage("x".repeat(5_000)), ...state.messages];
+      setMessages(state, [
+        createUiMessage("x".repeat(5_000)),
+        ...state.messages,
+      ]);
       const withUiOnlyPercent = getUsagePercent(state);
 
-      state.messages = [
+      setMessages(state, [
         ...state.messages,
         { role: "user", content: "more visible text", timestamp: 2 },
-      ];
+      ]);
       const withSecondUserPercent = getUsagePercent(state);
 
       expect(initialPercent).toBeGreaterThan(0);
@@ -383,7 +417,7 @@ describe("ui/status", () => {
         ...faux.getModel(),
         contextWindow: 1_000,
       };
-      state.messages = [
+      setMessages(state, [
         { role: "user", content: "first", timestamp: 1 },
         makeAssistantWithUsage("valid", {
           input: 150,
@@ -392,7 +426,7 @@ describe("ui/status", () => {
         }),
         aborted,
         { role: "user", content: "1234", timestamp: 2 },
-      ];
+      ]);
       return state;
     };
 
@@ -422,7 +456,7 @@ describe("ui/status", () => {
       ...faux.getModel(),
       contextWindow: 1_000,
     };
-    fromComponents.messages = [
+    setMessages(fromComponents, [
       makeAssistantWithUsage("fallback", {
         input: 120,
         output: 30,
@@ -430,8 +464,8 @@ describe("ui/status", () => {
         cacheWrite: 25,
         totalTokens: 0,
       }),
-    ];
-    fromTotalTokens.messages = [
+    ]);
+    setMessages(fromTotalTokens, [
       makeAssistantWithUsage("fallback", {
         input: 120,
         output: 30,
@@ -439,7 +473,7 @@ describe("ui/status", () => {
         cacheWrite: 25,
         totalTokens: 200,
       }),
-    ];
+    ]);
 
     try {
       expect(getUsagePercent(fromComponents)).toBe(

package/src/ui/status.ts

@@ -2,8 +2,7 @@
  * Status-bar formatting and rendering for the terminal UI.
  *
  * Computes the cumulative usage pill, abbreviates the working directory, and
- * estimates current model-visible context usage from persisted conversation
- * history.
+ * formats the cached current-context estimate maintained on application state.
  *
  * @module
  */
@@ -12,14 +11,9 @@ import { homedir } from "node:os";
 import { Spacer } from "@cel-tui/components";
 import { HStack, Text, visibleWidth } from "@cel-tui/core";
 import type { Node } from "@cel-tui/types";
-import type { AssistantMessage, Message } from "@mariozechner/pi-ai";
 import type { AppState } from "../index.ts";
-import { filterModelMessages, getAssistantUsage } from "../session.ts";
 import type { StatusTone, Theme } from "../theme.ts";
 
-/** Conservative fixed estimate for an image block's token footprint. */
-const ESTIMATED_IMAGE_TOKENS = 1_200;
-
 /**
  * Abbreviate a path with `~` for the home directory.
  *
@@ -84,143 +78,12 @@ function formatModelInfo(state: AppState): string {
   return `${state.model.provider}/${state.model.id} · ${formatEffort(state.effort)}`;
 }
 
-/** Calculate context tokens from assistant usage, falling back when `totalTokens` is zero. */
-function calculateUsageTokens(usage: AssistantMessage["usage"]): number {
-  return (
-    usage.totalTokens ||
-    usage.input + usage.output + usage.cacheRead + usage.cacheWrite
-  );
-}
-
-/** Estimate token usage from a character count using a conservative chars/4 heuristic. */
-function estimateCharacterTokens(charCount: number): number {
-  return Math.ceil(charCount / 4);
-}
-
-type UserMultipartContent = Exclude<
-  Extract<Message, { role: "user" }>["content"],
-  string
->;
-type TextOrImageContentBlock =
-  | UserMultipartContent[number]
-  | Extract<Message, { role: "toolResult" }>["content"][number];
-
-function estimateTextOrImageContentTokens(
-  content: readonly TextOrImageContentBlock[],
-): number {
-  let chars = 0;
-  let imageTokens = 0;
-
-  for (const block of content) {
-    if (block.type === "text") {
-      chars += block.text.length;
-      continue;
-    }
-    if (block.type === "image") {
-      imageTokens += ESTIMATED_IMAGE_TOKENS;
-    }
-  }
-
-  return estimateCharacterTokens(chars) + imageTokens;
-}
-
-function estimateUserMessageTokens(
-  message: Extract<Message, { role: "user" }>,
-): number {
-  if (typeof message.content === "string") {
-    return estimateCharacterTokens(message.content.length);
-  }
-  return estimateTextOrImageContentTokens(message.content);
-}
-
-function estimateAssistantBlockCharacters(
-  block: Extract<Message, { role: "assistant" }>["content"][number],
-): number {
-  if (block.type === "text") {
-    return block.text.length;
-  }
-  if (block.type === "thinking") {
-    return block.thinking.length;
-  }
-  return block.name.length + JSON.stringify(block.arguments).length;
-}
-
-function estimateAssistantMessageTokens(
-  message: Extract<Message, { role: "assistant" }>,
-): number {
-  const chars = message.content.reduce((total, block) => {
-    return total + estimateAssistantBlockCharacters(block);
-  }, 0);
-  return estimateCharacterTokens(chars);
-}
-
-function estimateToolResultMessageTokens(
-  message: Extract<Message, { role: "toolResult" }>,
-): number {
-  return estimateTextOrImageContentTokens(message.content);
-}
-
-/** Estimate token usage for a model-visible message. */
-function estimateMessageTokens(message: Message): number {
-  switch (message.role) {
-    case "user":
-      return estimateUserMessageTokens(message);
-    case "assistant":
-      return estimateAssistantMessageTokens(message);
-    case "toolResult":
-      return estimateToolResultMessageTokens(message);
-  }
-}
-
-/** Find the latest assistant usage that can anchor context estimation. */
-function getLatestValidAssistantUsage(
-  messages: readonly Message[],
-): { index: number; tokens: number } | null {
-  for (let i = messages.length - 1; i >= 0; i--) {
-    const message = messages[i];
-    if (
-      message?.role === "assistant" &&
-      message.stopReason !== "aborted" &&
-      message.stopReason !== "error"
-    ) {
-      const usage = getAssistantUsage(message);
-      if (!usage) {
-        continue;
-      }
-      return {
-        index: i,
-        tokens: calculateUsageTokens(usage),
-      };
-    }
-  }
-  return null;
-}
-
-/** Estimate the current model-visible context size for the next request. */
-function estimateCurrentContextTokens(state: AppState): number {
-  const messages = filterModelMessages(state.messages);
-  const latestUsage = getLatestValidAssistantUsage(messages);
-
-  if (!latestUsage) {
-    return messages.reduce((total, message) => {
-      return total + estimateMessageTokens(message);
-    }, 0);
-  }
-
-  let total = latestUsage.tokens;
-  for (let i = latestUsage.index + 1; i < messages.length; i++) {
-    total += estimateMessageTokens(messages[i]!);
-  }
-  return total;
-}
-
 /** Estimate current context usage as a percentage of the active model window. */
 function getContextPercentage(state: AppState): number {
   if (!state.model || state.model.contextWindow <= 0) {
     return 0;
   }
-  const contextTokens = estimateCurrentContextTokens(state);
-  return (contextTokens / state.model.contextWindow) * 100;
+  return (state.contextTokens / state.model.contextWindow) * 100;
 }
 
 /** Format cumulative session totals plus estimated current context usage for the status bar. */