Capture token-savings as core mission; automate usage visibility

AGENTS.md

@@ -8,6 +8,20 @@ This file is loaded into the agent's system prompt when it runs in this repo.
 - Add unit tests for new tools, providers, and loop behavior (Vitest).
 - Run `npm run lint && npm run typecheck && npm test` before considering a change done.
 
+## Token minimalism
+Keeping token counts low is a core design concern, not an afterthought. Savings
+should be automatic — the user should not need custom configuration to avoid
+runaway costs.
+
+- Keep the system prompt short. Tool descriptions are generated from Zod schemas;
+  don't add redundant prose.
+- Tool output must be bounded. Always cap result sets (grep matches, glob hits,
+  file lines). Prefer targeted reads over full-file slurps.
+- Surface usage automatically. Token counts appear after every turn and as a
+  session total on exit; no opt-in required.
+- Prefer features that reduce tokens structurally (output caps, compaction) over
+  features that merely expose knobs for users to tune manually.
+
 ## Boundaries
 - No business logic. This is a general-purpose tool.
 - Don't add a second state paradigm or heavy dependencies without a clear reason.

README.md

@@ -1,9 +1,11 @@
 # tiny-code
 
-A small, extensible CLI coding agent. Interactive terminal REPL, interchangeable
-**Anthropic** and **Gemini** models, and just the core features you actually use:
-read/write/edit files, run shell commands, search code, and a custom
-commands/skills system. No business logic baked in.
+A small, extensible CLI coding agent built around one constraint: **keep token
+usage low**. As coding-agent costs climb, tiny-code automates the savings so
+you don't have to. Interactive terminal REPL, interchangeable **Anthropic** and
+**Gemini** models, and just the core features you actually use: read/write/edit
+files, run shell commands, search code, and a custom commands/skills system.
+No business logic baked in.
 
 > Status: early (v0.x). Published as `@therr/tiny-code`; the binary is
 > `tiny-code`. Names may change before the first npm publish.
@@ -96,6 +98,35 @@ CLI flags.
 `allow` pre-approves mutating actions so they skip the confirmation prompt:
 `bash` matches command prefixes, `write` matches path globs for write/edit.
 
+## Token efficiency
+
+Minimizing token usage is a first-class goal — coding-agent bills grow fast,
+and you shouldn't need a complex setup to control them. tiny-code automates
+the savings where it can:
+
+- **Usage visible by default.** Every assistant turn prints `↑ in  ↓ out tokens`
+  with no configuration. On exit you get a session total.
+- **Bounded tool output.** grep caps at 200 matches, glob at 500 files, and
+  `read_file` supports `offset`/`limit` to pull only the lines you need —
+  preventing runaway context growth automatically.
+- **Minimal system prompt.** The built-in persona is kept short. Tool schemas
+  are generated from Zod (no duplicate prose). Project context is opt-in.
+- **Concise agent instructions.** The agent is explicitly told to avoid
+  restating the task or narrating completed steps.
+- **Effort control.** For Anthropic models, `effort` tunes the adaptive
+  thinking budget. Drop it from the default `"high"` to `"medium"` or `"low"`
+  for simpler, cheaper tasks:
+
+  ```json
+  { "effort": "medium" }
+  ```
+
+  Or set it per-session with `TINY_CODE_EFFORT=medium`.
+
+**Coming:** automatic conversation compaction once histories grow long
+(see `TODO.md`), which will keep input-token counts from compounding across
+many turns without any user action.
+
 ## Development
 
 ```bash

TODO.md

@@ -3,6 +3,20 @@
 Deferred features, roughly in priority order. Each entry notes the rationale and
 a rough approach so it can be picked up later.
 
+Token efficiency is a first-class goal. Features that reduce token usage
+automatically (without user configuration) are prioritized over features that
+add capability at higher cost.
+
+## Conversation compaction
+Input tokens compound quickly in long sessions because the full message history
+is resent every turn. Compaction trims the history automatically once it grows
+past a threshold, keeping costs from ballooning without any user action.
+**Approach:** track cumulative `inputTokens` from `AgentLoop.getUsage()`; when
+it crosses a configurable threshold (e.g. 50k), summarize earlier messages into
+a single condensed block. For Anthropic use the compaction beta; for Gemini
+summarize via a lightweight call to a cheap model. Pair with conversation
+persistence so compacted sessions can be resumed.
+
 ## Sub-agents
 Spawn isolated agent runs for parallel exploration/research (like a lightweight
 Explore/Plan agent). **Approach:** a `spawn_agent` tool whose `execute` constructs
@@ -34,8 +48,8 @@ exits; permission gate falls back to allowlist-only (no prompts).
 
 ## Conversation persistence / resume
 Save/restore `AgentLoop.getMessages()` to disk; `--resume` to continue a session.
-Pair with token-budget-aware compaction once histories get long (Anthropic
-compaction beta; manual summarization for Gemini).
+Pair with the compaction feature above so resumed sessions don't carry a bloated
+history.
 
 ## ripgrep-backed grep
 The `grep` tool currently walks the tree in JS. **Approach:** detect `rg` on

src/agent/loop.ts

@@ -4,6 +4,8 @@ import type { PermissionGate } from '../permissions/gate.js';
 import type { ToolResult } from '../tools/types.js';
 import type { Message, ToolResultBlock, ToolUseBlock } from './types.js';
 
+export type { Usage };
+
 /** Sink for everything the loop wants to surface. The REPL provides the real one. */
 export interface AgentUI {
   onText(delta: string): void;
@@ -39,6 +41,7 @@ export class AgentLoop {
   private readonly cwd: string;
   private readonly maxIterations: number;
   private readonly messages: Message[] = [];
+  private sessionUsage: Usage = { inputTokens: 0, outputTokens: 0 };
 
   constructor(opts: AgentLoopOptions) {
     this.provider = opts.provider;
@@ -55,6 +58,11 @@ export class AgentLoop {
     return this.messages;
   }
 
+  /** Cumulative token usage across all turns in this session. */
+  getUsage(): Usage {
+    return { ...this.sessionUsage };
+  }
+
   /** Run one user turn to completion (through any number of tool round-trips). */
   async run(userInput: string): Promise<void> {
     this.messages.push({ role: 'user', content: [{ type: 'text', text: userInput }] });
@@ -75,6 +83,8 @@ export class AgentLoop {
         } else if (event.type === 'tool_call') {
           toolCalls.push({ type: 'tool_use', id: event.id, name: event.name, input: event.input });
         } else {
+          this.sessionUsage.inputTokens += event.usage.inputTokens;
+          this.sessionUsage.outputTokens += event.usage.outputTokens;
           this.ui.onUsage(event.usage);
         }
       }

src/agent/systemPrompt.ts

@@ -13,7 +13,9 @@ Guidelines:
 - Read files before editing them. Prefer small, targeted edits.
 - When you run commands, explain briefly what you are doing only if it is non-obvious.
 - Match the conventions of the surrounding code.
-- When the task is complete, give a short summary of what changed. Do not narrate routine steps.`;
+- When the task is complete, give a short summary of what changed. Do not narrate routine steps.
+- Be concise. Avoid restating the task or narrating completed steps. Every output token has a cost.
+- Prefer targeted reads (offset/limit) and filtered searches (glob patterns) over reading entire files or scanning broadly.`;
 
 /** Compose the system prompt from persona, environment, tools, and project context. */
 export function buildSystemPrompt(params: SystemPromptParams): string {

src/repl.ts

@@ -111,7 +111,14 @@ export async function startRepl(overrides: CliOverrides): Promise<void> {
   };
 
   rl.on('close', () => {
-    console.log(pc.dim('\nBye.'));
+    const usage = agent.getUsage();
+    if (usage.inputTokens > 0 || usage.outputTokens > 0) {
+      const fmtN = (n: number) => n.toLocaleString('en-US');
+      console.log(
+        pc.dim(`\nSession: ↑ ${fmtN(usage.inputTokens)}  ↓ ${fmtN(usage.outputTokens)} tokens total`),
+      );
+    }
+    console.log(pc.dim('Bye.'));
     process.exit(0);
   });
 

src/ui/render.ts

@@ -10,6 +10,10 @@ function preview(name: string, input: unknown): string {
   return JSON.stringify(obj);
 }
 
+function fmtN(n: number): string {
+  return n.toLocaleString('en-US');
+}
+
 function truncate(s: string, n: number): string {
   const oneLine = s.replace(/\s*\n\s*/g, ' ').trim();
   return oneLine.length > n ? `${oneLine.slice(0, n)}…` : oneLine;
@@ -46,8 +50,8 @@ export function createTerminalUI(): AgentUI {
       ensureNewline();
       write(pc.yellow(`  ⊘ ${name} denied\n`));
     },
-    onUsage() {
-      // Token usage is available here; kept silent to reduce noise in the MVP.
+    onUsage(usage) {
+      write(pc.dim(`  ↑ ${fmtN(usage.inputTokens)}  ↓ ${fmtN(usage.outputTokens)} tokens\n`));
     },
     onAssistantEnd() {
       ensureNewline();

tests/agent/loop.test.ts

@@ -165,6 +165,31 @@ describe('AgentLoop', () => {
     }
   });
 
+  it('accumulates token usage across a single turn', async () => {
+    const provider = new ScriptedProvider([
+      [
+        { type: 'text', delta: 'hi' },
+        { type: 'done', usage: { inputTokens: 10, outputTokens: 5 }, stopReason: 'end_turn' },
+      ],
+    ]);
+    const { ui } = recordingUI();
+    const loop = makeLoop(provider, ui, gateWith('yes'));
+    await loop.run('hello');
+    expect(loop.getUsage()).toEqual({ inputTokens: 10, outputTokens: 5 });
+  });
+
+  it('accumulates token usage across multiple run() calls', async () => {
+    const provider = new ScriptedProvider([
+      [{ type: 'done', usage: { inputTokens: 10, outputTokens: 5 }, stopReason: 'end_turn' }],
+      [{ type: 'done', usage: { inputTokens: 20, outputTokens: 8 }, stopReason: 'end_turn' }],
+    ]);
+    const { ui } = recordingUI();
+    const loop = makeLoop(provider, ui, gateWith('yes'));
+    await loop.run('first');
+    await loop.run('second');
+    expect(loop.getUsage()).toEqual({ inputTokens: 30, outputTokens: 13 });
+  });
+
   it('stops at the iteration guard when tools never stop', async () => {
     const looping: ProviderEvent[][] = [];
     for (let i = 0; i < 10; i += 1) {

tests/ui/render.test.ts

@@ -39,6 +39,15 @@ describe('createTerminalUI', () => {
     expect(out).toContain('max iterations');
   });
 
+  it('displays token usage per turn', () => {
+    const out = capture(() => {
+      const ui = createTerminalUI();
+      ui.onUsage({ inputTokens: 1234, outputTokens: 567 });
+    });
+    expect(out).toContain('1,234');
+    expect(out).toContain('567');
+  });
+
   it('previews path- and pattern-based tools', () => {
     const out = capture(() => {
       const ui = createTerminalUI();