Skip to content

v1.3.183

Latest

Choose a tag to compare

@topcheer topcheer released this 29 Jul 10:25

ggcode v1.3.183

Highlights

  • Prompt injection guard — tool outputs from external/untrusted sources (web, files, commands, MCP servers) are scanned for common injection patterns and wrapped with an explicit "untrusted data" boundary warning before entering the context.
  • Post-write secret detection — file writing/editing tools scan written content for 18 secret patterns (AWS, GitHub, OpenAI incl. sk-proj-, GCP, Slack, JWT, private keys...) and warn immediately after the write.
  • Windows destructive command gaterd/rmdir /s, del /s, format, reg delete detection with block/ask tiers; system-drive wipes hard-blocked, quote paths and all drive letters covered, Unix path false positives eliminated.
  • Persistent session cost/token display — composer hint and status bar show live session token usage and estimated USD cost with per-model pricing fallback chain.
  • /commit slash command — generates an AI commit message from the current staged/unstaged diff; supports /commit <message> and /commit --amend.

Fixes and improvements

  • agent: restore "do NOT wrap up" constraint in 95%/99% context-budget warnings (long autopilot tasks must continue through compaction)
  • context: preserve image blocks/metadata through context compaction instead of dropping them
  • truncation: all context-boundary truncation is now UTF-8 safe (subagent results, tool output guard, run_command, mobile, IM, reflection) — no more split multi-byte characters
  • subagent: result truncation keeps head+tail so the conclusion (usually at the bottom) survives
  • windows: PowerShell no longer treats native-command stderr as errors ($ErrorActionPreference='Continue' + $PSNativeCommandUseErrorActionPreference=$false for pwsh 7.4+)
  • webui: 1MB HTTP body limit via MaxBytesHandler (WebSocket payloads unaffected)
  • cron: skipped firings are now logged for observability
  • git_branch: output capped at 100 branches with guidance
  • tui: spawn_agent labels show the effective model (desc [model]) even when inheriting the parent's runtime model
  • agent: injected prompt text (context warnings, budget guard, checkpoints) slimmed down to save context tokens

Upgrade notes

No breaking changes. No configuration migration required.

Compare

v1.3.182...v1.3.183