v1.3.101

v1.3.101

Critical fixes

Harness loop now learns from previous failures

The harness auto-run loop was repeating identical errors across retry attempts. Each retry started a fresh subprocess with zero conversation history, so the agent had no way to know what went wrong in the previous attempt.

Fix: BuildRunPrompt now injects the previous attempt's error, verification status, and log file path into the prompt when task.Attempt > 1. The agent sees exactly what failed and is explicitly instructed not to repeat the same approach.

Context compaction: rolling single-pass design

Redesigned the compaction system from a complex multi-pass dual-threshold approach to a clean rolling design:

• Single-pass summarize: All pre-compaction messages are compressed into one summary in a single LLM call. No more multi-pass loops with adaptive retention budgets.
• Fixed post-compaction target: ~30K tokens (or 25% of context window for smaller models), instead of a proportional ratio.
• Summary output cap: 5% of context window, capped at 12K tokens absolute (prevents 50K summaries on 1M context models).
• Threshold lowered to 0.99: Compaction triggers at 99% of usable budget (was 100%), leaving a small margin for the last turn.

Precompact anti-collision

Background precompact (async compression) was racing with the agent's regular LLM turn, causing API rate-limit collisions and snapshot invalidation:

• 6-second stagger delay: The compression LLM call now fires 6 seconds after the agent's turn starts, avoiding simultaneous API requests.
• Snapshot guard: ensurePromptSendable and tryReactiveCompact now skip local compaction (Microcompact/TruncateOldest) while a background precompact is running, preventing fingerprint changes that would invalidate the snapshot.
• Lossy apply: ApplyCompactResult now applies the compaction result even if messages changed within the snapshot range (lossy summary is always better than no compaction).

Session-scoped todos

todo_write now stores todos per-session at ~/.ggcode/todos/<sessionID>.json instead of per-workspace. This prevents todo bleed between concurrent sessions in the same workspace. The session ID is propagated from TUI/daemon/pipe to the agent.

Cancel no longer shows stuck "Cancelling..." state

Double Ctrl+C during an agent run was leaving the TUI in a stuck "Cancelling..." state with the spinner still running. The context.Canceled error path now performs full cleanup: stops spinner, finishes running tools, clears status, restores pending input.

Pending image queue fix

Images pasted with a submission while the agent was busy were lost when the submission was consumed from the pending queue. The pending queue now carries Images alongside text through enqueueWithImages → consumeDetailed.

Features

• debug_log tool: New tool that lets the LLM query the in-memory debug log ring buffer (500 entries). Useful for diagnosing provider errors, agent loop behavior, or harness failures. Category filterable (e.g. agent, provider, harness).

• Debug log noise cleanup: Removed ~2500 lines of noise per session from debug logs:

  • estimateTokensForMessages polling (1400 lines/session)
  • OpenAI tools JSON dump (60+ large lines)
  • TUI CATCHALL cursor blink + keypress logging
  • Agent permission decision trio (check + decision + execute)
  • QQ heartbeat spam (3 lines × 42 cycles)
  • Permission check full-input JSON dump
  • Stream frame ticks, byte counts, encoder selection
  • Emitter tool_result spam

• First-token timeout: Reduced from 5 minutes to 120 seconds. If the provider hasn't sent a response header in 2 minutes, it's effectively down.

• Slash command descriptions: Added missing descriptions for /mode, /allow, /export, /bug, /config, /reflect, /rules in both English and Chinese.

• Lanchat history compact format: lanchat tool result history now shows one line per message with a 50-char preview, instead of multi-line body dumps.

History rendering alignment (TUI + GUI)

Historical tool rendering in both TUI session reload and GUI GetSessionHistory now mirrors the real-time streaming path exactly:

• todo_write → TodoToolItem (was generic ToolItem)
• spawn_agent → uses DisplayName (was empty, showing raw JSON)
• exit_plan_mode → rendered as assistant text message
• Skipped tools (read_command_output, wait_command, stop_command, write_command_input, list_commands, enter_plan_mode, lsp_*) → omitted entirely
• tool_result formatting: suppressToolResult now resolves rawArgs from the matching tool_use block for proper formatting

Files changed

45 files changed since v1.3.100.