Skip to content

v1.3.189

Choose a tag to compare

View all tags
@topcheer topcheer released this 02 Aug 07:33
· 1836 commits to main since this release
v1.3.189
b878385

v1.3.189

86 commits since v1.3.188 (2026-08-01)

Highlights

  • Write-time code quality guards: 9 new post-write integrity checks — Unicode character detection, commented-out code, unterminated strings/comments, Python indentation, duplicate declarations, context leak detection, config syntax validation, test gaming detection, and auto-format for 7 languages.
  • Streaming UX overhaul: wait_agent, wait_command, and run_command now stream live progress to the TUI. Running tool calls show animated rotating icon (◐◓◑◒).
  • Agent intelligence gates: Self-correction stability gate, convergence lock, change reconciliation, verification regression detection, task re-anchoring, and tool sequence validation.
  • Git workflow: Full git toolset (revert/reset/tag), smart commit partitioning, pre-commit build gate, protected branch warnings.
  • Cross-tool compatibility: Reads .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md.
  • MCP ecosystem: Dynamic tool refresh, elicitation support, sampling/createMessage.

New Features

Write-Time Code Quality Guards

Check Description
Unicode character detection Flags smart quotes, non-breaking spaces, zero-width chars, BOM, em-dashes introduced by LLM
Commented-out code detection Warns when large commented-out code blocks appear in edits
Unterminated string/comment detection Catches unclosed string literals, block comments, Python triple-quotes at EOF
Python indentation consistency Detects mixed tabs/spaces in .py indentation (would cause TabError)
Duplicate declaration detection Catches redeclared Go symbols in post-write check
Context leak detection Flags context.TODO()/context.Background() in functions that already receive ctx context.Context
Config syntax validation Post-edit validation for JSON/YAML/TOML/XML config files
Auto-format on write Runs gofmt/rustfmt/black/prettier/clang-format/dart format on file write
Test gaming detection Catches agents that delete tests, add t.Skip(), or remove assertions instead of fixing code

Agent Intelligence & Safety Gates

  • EIR-based self-correction stability gate: Prevents oscillating correction loops by tracking edit-to-verification ratio.
  • Post-verification convergence lock: Detects and warns about unnecessary edits after successful verification.
  • Pre-completion change reconciliation gate: Ensures all modified files are accounted for before task completion.
  • Verification regression detection: Detects when a correction introduces new failures.
  • Task re-anchoring: Prevents context collapse from losing track of the original task.
  • Tool call sequence validator: Detects cross-iteration anti-patterns (e.g., repeated failed edits to same file).
  • Undo-triggered correction feedback: Agent learns from undo operations.

Streaming UX

  • wait_agent live progress: Streams sub-agent status, tool count, current tool, and progress summary during polling.
  • wait_command live output: Scrolling log effect with [...]/[completed] status prefix.
  • run_command real-time output: Up to 5 lines of live command output during execution.
  • Animated tool call icon: Running tools show rotating quarter-circle (◐◓◑◒) instead of static circle.
  • start_command initial output: Shows first 3s of output (later reverted per user preference).

Git Workflow

Tool/Feature Description
git_revert Revert commits safely
git_reset Reset staging area / working tree
git_tag Create, list, delete tags
Smart commit partitioning Suggests logical commit grouping for large changes
Pre-commit build gate Verifies code compiles before allowing git_commit
Protected branch warning Warns when editing on main/master/develop
Destructive command detection Blocks git push --force, git reset --hard, etc.

MCP Ecosystem

  • Dynamic tool refresh: Server tool list changes are detected and applied without restart.
  • Elicitation (elicitation/create): Servers can request structured input from the user.
  • Sampling (sampling/createMessage): Servers can request LLM completions via the host agent.
  • Full capability detection: Complete Agent Card / tool / prompt / resource discovery.

Session & Context

  • Smart session title: Auto-generates descriptive titles from conversation content.
  • Cross-session search (/search): Full-text search across all session histories.
  • Session preview snippet: Resume picker shows a content preview for each session.
  • Recent git commits in system prompt: Injects last N commits for session continuity.
  • Per-session tool call budget: Progressive enforcement of tool call limits.

Cross-Tool Compatibility

  • Reads rules from .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md in addition to GGCODE.md/AGENTS.md/CLAUDE.md.

Provider & Model

  • OpenAI adaptive sampling: Per-turn temperature/top_p control based on task phase.
  • Gemini reasoning effort + tool choice: Full thinking_budget and tool_choice support.
  • Adaptive per-tool timeout: Based on tool category and observed latency baseline.

Developer Experience

  • open_editor tool: Launch files in user's external IDE with line/column jump.
  • undo_edit agent tool: Revert file edits made in the current session.
  • /undo-run command: Batch revert all file changes from a specific run.
  • /goal command: Persistent autopilot goal injection.
  • Auto-detect lint/format commands: Scans Makefile/package.json and injects into system prompt.
  • Auto-remove unused Go imports: On file write.
  • Post-edit LSP diagnostics: Cross-package diagnostic detection with baseline diffing.

Security

  • Network egress detection: Warns about outbound network calls (data exfiltration prevention).
  • Output redirection guard: Detects writes to sensitive system paths (/dev/, /etc/, etc.).
  • Blind write detection: Warns when overwriting files that haven't been read.
  • Critical infrastructure file warnings: Special warnings for editing lock files, CI configs, etc.
  • Pre-run dirty tree awareness: Captures working tree state before agent runs.

Fixes

  • Tool progress callbacks decoupled from lanchat: SetToolProgressCallback was inside SetLanChatHub() which was (a) never called if lanchat unconfigured, (b) called before r.program created. Extracted to standalone RegisterCallbacks().
  • Cron double-fire on restart: Three-layer defense (enqueue debounce, generation tracking, timer reset).
  • Cron debounce broke timer chain: Permanently stopped cron jobs after skip.
  • Cron system messages missing timestamps: Added timestamp display.
  • run_command/wait_command streaming not rendering: ToolProgressKey was not injected into tool execution context.
  • BashToolItem suppressed result body: suppressBody flag incorrectly hid command output.
  • Test flakiness: Hardened TestOpenEditor_LaunchAndVerify, TestPostEditVerifyHintWithImpactAndCoverage, TestSyncVerify against CI load.
  • Pre-commit diff scan: Now runs on git_commit all:true path.
  • Dead code removal: Cleaned up summarizeCompilerErrors (reviewer P2).

Internal

  • Per-tool description token budgeting for MCP schema bloat mitigation.
  • Build/test output intelligence — structured command result extraction.
  • Tool argument size guard for context efficiency.
  • Delayed input-needed notifications for blocked agent runs.
  • Companion file test coverage guard.
  • Press Up arrow to de-queue pending messages while agent is busy.
  • Persistent autopilot goal injection.

Compare

Assets 51
Loading