feat: FrameBuffer diff rendering, input box, cursor fix#205
Merged
Sunrisepeak merged 40 commits intomainfrom Mar 15, 2026
Merged
feat: FrameBuffer diff rendering, input box, cursor fix#205Sunrisepeak merged 40 commits intomainfrom
Sunrisepeak merged 40 commits intomainfrom
Conversation
- Add FrameBuffer off-screen line buffer for diff-based rendering: only rewrite changed lines instead of full clear+redraw each frame - Fix cursor positioning bug: track prev_cursor_line_ so cursor_up moves the correct distance (was overshooting by old_count - cursor_line) - Fix flush_to_scrollback and loop cleanup to use tracked cursor position - Add bordered input box component (╭╮╰╯│─) replacing flat separators - Input box uses amber border color during approval prompts - Hide cursor when no cursor position set (approval mode) - Add CancellationToken, utf8 module, platform RunCommand support - Add ContextManager 3-level cache, TokenTracker, OutputBuffer - Add ResourceTools (RunCommand/ViewOutput/SearchContent capabilities) - Expand agent loop with manage_tree, multi-provider support - Add download progress with speed/ETA in tree nodes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a turn completes and the tree is printed to scrollback, the root node title (user's message) was duplicated — already shown as the UserMsg block above. Now TurnTree prints only state icon + duration on a summary line, then children directly, skipping the root title. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ameters Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…mplate Extract llm_call_worker<Provider, ProviderConfig> template before run_one_turn to unify the Anthropic/OpenAI worker-thread blocks, which were identical except for their provider/config types. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…t capabilities Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Create a MemoryStore per session (loaded from AgentFS) and build the agent-specific capability registry with both memory and context tools, shadowing the outer non-agent registry that lacks these extensions. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Bump dependency to llmapi 0.2.3 which adds cacheCreationTokens and cacheReadTokens fields to Usage. Add cache_read_tokens/cache_write_tokens to TurnResult, accumulate them in run_one_turn, and pass them through to TokenTracker::record() so session cache stats are tracked correctly. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds MemorySummary struct to loop.cppm and extends build_system_prompt() to accept memory entries, appending a Remembered Context section when memories are present (capped at 20 with overflow notice). Updates the call site in cli.cppm to build summaries from MemoryStore.all_entries(). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…in tree Two bugs fixed: 1. complete_task() now calls mark_running_as() on the completed node's children, so Response/Thinking nodes under a finished task stop timing 2. Done Response nodes now show the first line of the model's reply as a "◆ summary" in the tree instead of being auto-hidden Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce LuaSandbox (lua_engine.cppm) that lets the LLM generate Lua code to orchestrate multiple capability calls in a single tool invocation. The sandbox exposes pkg/sys/ver modules via upvalue trampolines into the CapabilityRegistry, runs in a worker thread with debug.sethook for timeout/cancel protection, and returns structured ExecutionLog JSON. Wire execute_lua as a virtual tool in loop.cppm (same pattern as manage_tree), inject Lua API docs into system prompt, and create the sandbox instance in cli.cppm. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Remove leftover [tui] debug fprintf to stderr. Hide completed Response nodes in tree rendering (same as Thinking/Detail) since the full reply is already printed as AssistantText in scrollback — avoids duplicate display. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…anges When manage_tree creates subtasks, active_parent moves to a subtask node. The Response node created earlier at root level was never closed because the close logic only checked active_parent's children. Now also check the root node to close any lingering Running Response/Thinking nodes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update system prompt to emphasize that run_command and execute_lua should only be used when no other built-in tool can accomplish the task. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace FrameBuffer/diff rendering with print-stream + 2-line fixed area model to eliminate root duplication bug and reduce complexity. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add initialization step, multi-line text handling, completions/approval, download progress, flash messages, manage_tree sync, terminal resize. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace tinytui print-stream + 2-line fixed area with ftxui fullscreen layout: scrollable history area with recursive task tree, fixed bottom input box + status bar, mouse wheel scrolling, and 200ms timer for real-time elapsed time display. - tui.cppm: replace flat TaskEntry/TaskList with recursive TreeNode/TurnNode - ftxui_tui.cppm: new AgentScreen class with render_tree_node (Unicode connectors ├─└─│), render_turn, render_status_bar, mouse wheel support - loop.cppm: adapt handle_manage_tree for recursive tree with parent_id - cli.cppm: migrate from tinytui::Screen to AgentScreen, callbacks now operate on tree structure instead of print_line Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement render_input_box() that splits editor content at cursor position, renders the cursor character with inverted style, and wraps the input line between two separator lines for visual distinction. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…nage_tree Replace the LLM-driven task tree (manage_tree virtual tool) with a system-driven recursive model where LLM decides execute-or-decompose via the `decide` virtual tool, and the system handles DFS traversal, backtracking, and failure recovery. Key changes: - BehaviorNode: kind→type system (TypeRoot/Decompose/Execute/DirectExec/ ToolCall/Response), add result_summary/tool/tool_args fields - ABehaviorTree: replace plan management (add_plan/start_plan/complete_plan) with system-driven API (set_root/add_child/set_state/set_result) - loop.cppm: add run_task_tree entry point with process_node (recursive DFS), ask_decision, run_execute, run_tool_loop (extracted inner loop) - DirectExec: LLM specifies tool+args at decompose time, system calls directly with zero LLM overhead; tool name validated against registry - Failure recovery: children continue on failure, system adds LLM verification node to review results and decide final outcome - Cross-turn context: shared conversation injected at root level (depth 0) only, nested nodes use sibling_results + ancestor_path - ToolBridge: check exitCode in results, set isError when non-zero - Context budget: use per-node input tokens instead of stale session tracker - ftxui_tui: ⚙ icon for DirectExec nodes, type-based rendering - prompt_builder: add build_scoped() for per-node scoped prompts Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…lures - Verify node now uses a tool-less single LLM call (just summarize, no fix attempts). Prompt explicitly says "Do NOT attempt to fix". - Parent node state is Failed when any child failed, not always Done. - Previously verify ran run_execute with full tool access, causing it to try fixing failures instead of reviewing them, and always returned Done even when summary reported failures. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
DirectExec nodes now display the actual tool call like:
⚙ search_packages("nodejs") 41ms
✗ use_version("nodejs", "18") 0ms
instead of the LLM-generated task title. Makes it clear which
nodes are system-direct atomic operations vs LLM-driven tasks.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the 6-type node system with just 2 types: - TypeAtom(0): has tool+args, system direct execute, zero LLM - TypePlan(1): needs LLM, ask_decision decomposes into children Key changes: - Delete TypeRoot/TypeDecompose/TypeExecute/TypeToolCall/TypeResponse - Delete run_tool_loop, run_execute, ToolLoopConfig, NodeResult - process_node: Atom branch (direct tool call) + Plan branch (ask_decision → create children → DFS → re-plan if failures) - Re-plan: after child failures, ask_decision again with results context. LLM can accept (done) or add new subtasks (decompose). Max MAX_REPLAN=3 rounds. - decide protocol: "decompose" (split) or "done" (accept/finish) - Atom approval: inline policy check before bridge.execute - execute_lua: handled as special Atom in process_node - TUI: Atom shows ⚙ tool(args), Plan shows ○/⟳/✓/✗ title - Remove streaming/think_filter (no tool-use loop = no streaming) Tree-shaped streaming reasoning: each ask_decision sees scoped sibling_results, dynamic node addition = streaming reasoning in tree form. No capability loss vs flat tool-use loop. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Decide prompt now explicitly instructs LLM to group related operations under Plan subtasks instead of flat Atom lists. Includes example showing correct decomposition pattern. - Re-plan's new subtasks are wrapped under a visible "re-plan #N" Plan sub-node in the tree, making decision points visible in TUI. - Re-plan "done" decision adds a visible marker node showing the summary. Before: flat list of 14 Atom nodes at root level, re-plan invisible After: hierarchical Plan groups + visible re-plan decision nodes Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three fixes: 1. Available Tools section now shows inputSchema JSON so LLM knows exact parameter names (e.g. "target" not "name" for remove_package) 2. Decide prompt emphasizes checking sibling results before acting - if a sibling already shows a package is not installed, the Plan node should return "done" immediately instead of attempting removal 3. Re-plan prompt reframed: evaluate whether the PARENT TASK is complete, not whether to retry. "remove failed because package wasn't installed = task done" → prefer "done" over "decompose" Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…hecks 1. synthesize_children now includes Atom node results (tool: summary) instead of discarding them. This fixes sibling_results being useless "2/2 tasks completed" instead of actual tool output. 2. Decide prompt now says "Do NOT add check-if-installed steps before operations" and shows a direct Atom example with correct parameter names, encouraging the LLM to just act and let re-plan handle failures. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Comprehensive documentation covering: - Two-type node system (Atom + Plan) - process_node recursive DFS flow - Re-plan mechanism with visible markers - Context passing (conversation, sibling_results, ancestor_path) - ABehaviorTree thread-safe API - TUI rendering (icons, colors, tree structure) - Full data flow diagram Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Capture LLM reasoning text (before decide tool call) in Decision struct. Store it in Plan node's detail field so it's visible in the tree as context for what the LLM decided and why. 2. TUI: show Plan node detail as a dim secondary line when the node is terminal and has no children (leaf Plan with reasoning). 3. Final reply (◆): add a summary LLM call after tree completes to generate a human-readable response instead of mechanical "3/3 tasks completed" synthesis. Uses on_stream_chunk for real-time streaming to TUI. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. ask_decision now streams LLM output via on_stream_chunk, so the
TUI shows real-time "thinking..." text in the ◆ reply area while
the LLM reasons about how to decompose.
2. After decompose, a ◇ thinking node is inserted as the first child
of each Plan node, showing the LLM's decision reasoning (truncated
to ~80 chars, single line). Uses cyan color to distinguish from
Atom (⚙) and Plan (○) nodes.
3. on_stream_chunk restored in cli.cppm TreeConfig for streaming.
TUI display:
├─ ◇ 需要先搜索版本信息，然后安装第二新版本... ← cyan, thinking
├─ ⚙ search_packages("d2x") ← green, Atom
└─ ○ 安装第二新版本 ← amber, Plan
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
LLMs often don't output text before tool calls, so response.text() was empty and thinking nodes never appeared. Fix: add "thinking" as a required field in the decide tool schema. LLM must provide its reasoning inside the tool args, which is reliably parsed from JSON. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Critical context fix: execute_pending_children now inherits the parent's sibling_results (ancestor context) so nested Plan nodes can see results from upper levels. Previously each nesting level lost all parent context, causing the LLM to re-search for information that was already found at a higher level. Also: - Clear streaming text after each ask_decision to prevent accumulation - Increase thinking node display from 80 to 120 chars - Filter out __thinking__ nodes from sibling_results Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Code-level execution flow covering: - cli.cppm → run_task_tree → process_node call chain - Atom execution path (approval, execute_lua, bridge.execute) - Plan decision path (ask_decision, scoped prompt assembly) - Re-plan loop with visible marker nodes - Context data flow (conversation injection, sibling inheritance) - TUI update timing sequence (worker ↔ main thread) - Disconnected components inventory (ContextManager断路) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Proposes reducing 19 modules (4499 lines) to 11 modules (~3100 lines): - Delete 8 disconnected modules (context_manager, prompt_builder, resource_cache, resource_tools, package_tools, mcp_client, output_buffer, token_tracker) totaling 1370 lines - Replace ContextManager 3-level cache with simple conversation auto-compact in run_task_tree - Merge token tracking into TreeResult - Compress Atom result_summary from raw JSON Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove 10 obsolete agent modules (lua_engine, mcp_client, output_buffer, package_tools, prompt_builder, resource_cache, resource_tools, llm_config, ftxui_tui, old tui) — replaced by llm, prompt, runtime, tui/screen, tui/state - Add thinking nodes to behavior tree for LLM reasoning display - Unify quiet mode: replace separate exec_quiet/log::set_quiet with single platform::set_tui_mode() flag checked by exec(), log, and download renderer - Add agent architecture docs and plans Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Agent TUI: show real-time download progress on running tool nodes (↓ done/total pct%) - Fix download progress bar jumping 0→100% when some packages have unknown sizes (fall back to count-based progress when byte sizes unavailable) - Unify downloads: use tinyhttps streaming for all paths (remove curl subprocess) - Fix CLI progress bar leaking into agent TUI on Ctrl+C (cleanup order race) - Fix tuiThread writing ANSI cursor codes to stdout in TUI mode - Guard dispatch_data_event against TUI mode for download_progress events - Use tinyhttps download_to_file() for incremental progress callbacks Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Show each downloading file as a child node with aligned progress bar (█ filled cyan, ░ unfilled dark) under the install tool node - Install node shows total progress (↓ pct%) - Pass isCancelled through tinyhttps for responsive ESC abort during download - Clean up download_files on tool completion Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…vements - Download files are real BehaviorNode (TypeDownload) with progress bars that persist in the tree after completion - Fix node ID collision: use IdAllocator for download nodes - ESC interrupts downloads (pause → cancel downloader via isCancelled) - Show "interrupted by ESC" in red with hint text - Ctrl+C × 3 exit: ForceHandleCtrlC(false), counter hint in status bar, auto-clear after 1.5s timeout, only character input resets count - Tree connectors (├─ └─ │) styled with border_color - Paste debounce: suppress submit when Return arrives <5ms after chars Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add platform::Icon with ASCII fallback for Windows (Unicode on Linux/macOS) - Unify icon references: theme.cppm and screen.cppm use platform::Icon - Optimize tool descriptions: clarify parameter formats (plain name vs namespace:name), add fuzzy match hint to search_packages - Add download speed to agent TUI total progress (↓ 45% 3.2 MB/s) - Fix download nodes: set Done state on creation for already-finished files - Fix slash command completion: Enter auto-fills selected completion - Paste debounce: prevent multi-line paste from triggering submit Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
prev_cursor_line_socursor_upmoves correct distance (was overshooting byold_count - cursor_linelines, corrupting scrollback)╭╮╰╯│─) replacing flat separators, amber borders for approval prompts, cursor hidden in approval modeTest plan
rm -rf build && xmake build— compiles cleanxmake run xlings_tests— 209 tests pass>inside input box🤖 Generated with Claude Code