@shuyhere/bb-agent 0.0.11 → 0.0.14

package/CHANGELOG.md

@@ -5,18 +5,75 @@ All notable changes to BB-Agent will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.0.14] - 2026-04-12
+
+### Added
+
+- fullscreen now supports extension-driven workflows and structured slash-command outcomes, including menus, hidden dispatches, and richer command result handling
+- `/settings` in fullscreen now exposes compaction controls for `Auto-compact`, `Reserve tokens`, and `Keep recent tokens`
+- skills can now be listed, disabled, and re-enabled from the CLI without deleting their installed files
+- startup model selection now prefers configured provider/model defaults more consistently, with better OpenAI startup fallback behavior
+- added a parity test script against installed pi compaction logic to keep BB token accounting aligned with upstream behavior
+
+### Fixed
+
+- session resume now restores the prior model and thinking level instead of starting with mismatched runtime defaults
+- fullscreen/TUI terminal rendering now sanitizes terminal control text more reliably and avoids ANSI leakage into the UI
+- auto-compaction token estimation now matches pi more closely by using the last successful assistant usage plus trailing estimates, using ceil-based token heuristics, computing `tokens_before` from rebuilt context instead of raw payload size, and ignoring assistant usage from before the latest compaction boundary
+- fullscreen compaction behavior and status reporting are more consistent after auto-compaction and manual compaction events, and local fullscreen actions now show an animated elapsed-time status while they run
+
+### Changed
+
+- fullscreen extension workflows and session compaction support are now merged into the main interaction path on `master`
+
+## [0.0.13] - 2026-04-09
+
+### Added
+
+- fullscreen screenshot and image clipboard paste now works on the normal paste path, with macOS clipboard fallbacks and Codex image preservation so pasted images reach image-capable models correctly
+- model registry metadata now tracks image input capability, making `/models` truthful about image support and allowing runtime warnings when users attach images to text-only models
+
+### Fixed
+
+- fullscreen clipboard image attach no longer leaks helper `true` / `false` output or stray follow-up paste text into the input block
+- attached image chips can now be removed with `Backspace`, image-only prompts can be submitted, and optimistic user messages keep attachment chip previews in the transcript
+- rebuilt fullscreen session transcripts now preserve user image attachment markers instead of silently dropping image blocks
+- managed `bb-clipboard-*.png` temp files are now cleaned up after removal or ingestion instead of lingering in `/tmp`
+- the fullscreen input block now hides raw `@file` tokens when the corresponding attachment chip is already shown, preventing duplicated `@file` text in the editor
+- fullscreen tool-header regression tests now match the intended live bash-header rendering and running-dot animation behavior
+
+## [0.0.12] - 2026-04-06
+
+### Fixed
+
+- direct `@image` references in print mode and fullscreen now attach real image inputs instead of falling back to UTF-8 read warnings
+- `@path with spaces` parsing now correctly keeps the full file path before trailing prompt text, including whole-message forms
+- image tool results are now preserved through provider conversion so models can actually see images returned by tools instead of responding as if no image was provided
+- fullscreen `@` folder navigation now keeps the completion menu open when you select a directory and immediately shows the next level, including directories with spaces
+- the fullscreen input block now shows attached files as `[name, sizeKB]`, keeps those chips visible, and places the cursor below them so typing starts after the attachments
+
+### Changed
+
+- binary office/document inputs (`pdf`, `docx`, `pptx`, `xlsx`) now degrade to format-aware metadata notes instead of misleading invalid-UTF-8 errors
+
 ## [0.0.11] - 2026-04-07
 
 ### Added
 
 - startup update notices in the fullscreen transcript are now highlighted so available updates stand out clearly during startup
 - read-tool line ranges in fullscreen tool activity now highlight the requested span, so values like `2148-2267/5006` stand out while the model is using tools
+- fullscreen footer and `/session` info now show the active execution posture so safety vs yolo is visible during a run
 
 ### Improved
 
 - npm install now caches verified native binaries by version/platform and reuses them on reinstall instead of re-downloading every time
 - npm install now shows more frequent download progress with transfer rate information to make slow installs easier to understand
 - npm install now avoids unnecessary re-verification on cache hits, making repeat installs faster
+- safety mode now restricts built-in `write` and `edit` to the active workspace, while `yolo` keeps unrestricted file mutation behavior
+
+### Migration
+
+- `execution_mode` now defaults to `safety`; set `"execution_mode": "yolo"` if your workflow intentionally edits files outside the current workspace
 
 ## [0.0.10] - 2026-04-07
 
@@ -71,6 +128,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - latest published package includes the post-0.0.7 startup, auth, model-default, and update-notice improvements
 
+[0.0.14]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.14
+[0.0.13]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.13
+[0.0.12]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.12
 [0.0.11]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.11
 [0.0.10]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.10
 [0.0.9]: https://github.com/shuyhere/bb-agent/releases/tag/v0.0.9

package/README.md

@@ -101,6 +101,7 @@ bb --list-models                      # List all available models
 - **Fullscreen TUI** — rich terminal interface with streaming output, markdown rendering, syntax highlighting
 - **Multi-provider** — Anthropic (Claude), OpenAI, Google (Gemini), Groq, xAI, OpenRouter, and custom OpenAI-compatible endpoints
 - **Built-in tools** — `read`, `write`, `edit`, `bash`, `find`, `grep`, `ls`, `web_search`, `web_fetch`, `browser_fetch`
+- **Safety and yolo execution modes** — default safety posture restricts built-in `write` and `edit` to the active workspace; yolo removes that guard
 - **Session persistence** — SQLite-backed sessions with branching, forking, and tree navigation
 - **Extensions** — JS/TS plugin system for custom tools, commands, and hooks
 - **Skills** — markdown-based instruction files that auto-load contextual knowledge
@@ -144,9 +145,11 @@ BB-Agent uses layered configuration:
 
 ```json
 {
+  "execution_mode": "safety",
   "default_model": "claude-sonnet-4-20250514",
   "default_provider": "anthropic",
   "default_thinking": "medium",
+  "execution_mode": "safety",
   "models": [
     {
       "id": "my-local-model",
@@ -160,6 +163,21 @@ BB-Agent uses layered configuration:
 }
 ```
 
+### Execution Modes
+
+BB-Agent exposes the active permission posture in fullscreen and `/session`.
+
+- `safety` is the default. Built-in `write` and `edit` stay inside the current workspace, and bash commands use the safer approval/sandboxed posture.
+- `yolo` is the opt-in less-restrictive mode.
+
+Example:
+
+```json
+{
+  "execution_mode": "yolo"
+}
+```
+
 ## Keyboard Shortcuts
 
 | Key | Action |
@@ -228,7 +246,6 @@ Or set this in `~/.bb-agent/settings.json`:
 - [Development Guide](docs/development.md) — build from source, dev workflow, project structure, debugging
 - [Contributing](CONTRIBUTING.md) — code style, PR process
 - [Changelog](CHANGELOG.md) — release history
-- [Security](SECURITY.md) — vulnerability reporting, security model
 
 ## Development
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@shuyhere/bb-agent",
-  "version": "0.0.11",
+  "version": "0.0.14",
   "description": "BB-Agent — a Rust-native AI coding agent for the terminal",
   "license": "MIT",
   "repository": {

package/scripts/test-pi-token-parity.mjs

@@ -0,0 +1,141 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import * as pi from '/home/shuyhere/projects/tako/node_modules/@mariozechner/pi-coding-agent/dist/core/compaction/compaction.js';
+
+const settings = { enabled: true, reserveTokens: 16384, keepRecentTokens: 20000 };
+
+function bbEstimateContextTokens(messages) {
+  function calc(usage) {
+    return usage.total_tokens > 0
+      ? usage.total_tokens
+      : usage.input + usage.output + usage.cache_read + usage.cache_write;
+  }
+
+  function estimateText(text) {
+    return Math.ceil(text.length / 4);
+  }
+
+  function estimateMessage(message) {
+    switch (message.role) {
+      case 'user':
+        return message.content.reduce((sum, block) => sum + (block.type === 'text' ? estimateText(block.text) : 1200), 0);
+      case 'assistant':
+        return message.content.reduce((sum, block) => {
+          if (block.type === 'text') return sum + estimateText(block.text);
+          if (block.type === 'thinking') return sum + estimateText(block.thinking);
+          if (block.type === 'toolCall') return sum + estimateText(block.name) + estimateText(JSON.stringify(block.arguments ?? {}));
+          return sum;
+        }, 0);
+      case 'toolResult':
+      case 'custom':
+        return message.content.reduce((sum, block) => sum + (block.type === 'text' ? estimateText(block.text) : 1200), 0);
+      case 'bashExecution':
+        return estimateText(message.command) + estimateText(message.output);
+      case 'branchSummary':
+      case 'compactionSummary':
+        return estimateText(message.summary);
+      default:
+        return 0;
+    }
+  }
+
+  const lastUsageIndex = [...messages.keys()].reverse().find((i) => {
+    const m = messages[i];
+    return m.role === 'assistant' && m.stopReason !== 'aborted' && m.stopReason !== 'error' && calc(m.usage) > 0;
+  });
+
+  if (lastUsageIndex == null) {
+    const trailingTokens = messages.reduce((sum, m) => sum + estimateMessage(m), 0);
+    return { tokens: trailingTokens, usageTokens: 0, trailingTokens, lastUsageIndex: null };
+  }
+
+  const usageTokens = calc(messages[lastUsageIndex].usage);
+  const trailingTokens = messages.slice(lastUsageIndex + 1).reduce((sum, m) => sum + estimateMessage(m), 0);
+  return { tokens: usageTokens + trailingTokens, usageTokens, trailingTokens, lastUsageIndex };
+}
+
+test('bb estimator matches installed pi on shared fixtures', () => {
+  const fixtures = [
+    [
+      {
+        role: 'assistant',
+        content: [{ type: 'text', text: 'done' }],
+        provider: 'test',
+        model: 'test',
+        usage: { input: 100, output: 20, cacheRead: 10, cacheWrite: 5, totalTokens: 0 },
+        stopReason: 'stop',
+        timestamp: Date.now(),
+      },
+      { role: 'user', content: [{ type: 'text', text: '12345678' }], timestamp: Date.now() },
+    ],
+    [
+      {
+        role: 'assistant',
+        content: [{ type: 'text', text: 'aborted' }],
+        provider: 'test',
+        model: 'test',
+        usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 500 },
+        stopReason: 'aborted',
+        timestamp: Date.now(),
+      },
+      { role: 'user', content: [{ type: 'text', text: '12345678' }], timestamp: Date.now() },
+    ],
+    [
+      {
+        role: 'assistant',
+        content: [
+          { type: 'text', text: 'hello' },
+          { type: 'thinking', thinking: 'reasoning text' },
+          { type: 'toolCall', id: '1', name: 'read', arguments: { path: 'src/main.rs' } },
+        ],
+        provider: 'test',
+        model: 'test',
+        usage: { input: 300, output: 40, cacheRead: 50, cacheWrite: 0, totalTokens: 0 },
+        stopReason: 'stop',
+        timestamp: Date.now(),
+      },
+      {
+        role: 'toolResult',
+        toolCallId: '1',
+        toolName: 'read',
+        content: [{ type: 'text', text: 'fn main() {}' }],
+        isError: false,
+        timestamp: Date.now(),
+      },
+      {
+        role: 'bashExecution',
+        command: 'ls -la',
+        output: 'file1\nfile2\n',
+        cancelled: false,
+        truncated: false,
+        timestamp: Date.now(),
+      },
+    ],
+  ];
+
+  for (const messages of fixtures) {
+    const piEstimate = pi.estimateContextTokens(messages);
+    const bbEstimate = bbEstimateContextTokens(
+      JSON.parse(
+        JSON.stringify(messages)
+          .replace(/cacheRead/g, 'cache_read')
+          .replace(/cacheWrite/g, 'cache_write')
+          .replace(/totalTokens/g, 'total_tokens')
+          .replace(/stopReason/g, 'stopReason')
+      )
+    );
+
+    assert.deepEqual(bbEstimate, {
+      tokens: piEstimate.tokens,
+      usageTokens: piEstimate.usageTokens,
+      trailingTokens: piEstimate.trailingTokens,
+      lastUsageIndex: piEstimate.lastUsageIndex,
+    });
+
+    const contextWindow = 128000;
+    assert.equal(
+      bbEstimate.tokens > contextWindow - settings.reserveTokens,
+      pi.shouldCompact(piEstimate.tokens, contextWindow, settings)
+    );
+  }
+});