Repository navigation
Releases
Releases · RedGreenOrg/RedGreen
v0.1.10
- OpenRouter support: new provider (OPENROUTER_API_KEY, openrouter/auto default, or any model id). Picks it in the init wizard or via
px redgreen init -p openrouter. - Any OpenAI-compatible endpoint (Groq, Together, LM Studio, llama.cpp, ...) via "provider": "openai" + "baseUrl" - no new provider needed. Custom endpoints use the chat-completions wire protocol; OpenAI proper keeps its native API.
- README provider table + note; wizard gains the OpenRouter row.
- New tests: openrouter env-key chat, labeled missing-key error, custom baseUrl, provider/env wiring pin; model-defaults regression pin now covers openrouter/auto.
v0.1.9
- Fixed the attack-round-3 bug: REFACTOR_STEP shadowed the final attack step, so rounds stopped at 2/3 in both headless and interactive modes. The refactor step now sits after the third attack round, and the pipeline runs the full three rounds.
- Regression test: a headless REDGREEN_ATTACK run asserts all three rounds survive (events + summary "attacks 3/3").
- Interactive refactor tests updated for the extra executes per round.
v0.1.8
- Headless attack auto-advance: REDGREEN_ATTACK=1 runs through all three attack rounds unattended (released in 0.1.7).
- CI on every push/PR: typecheck, tests, and build run in GitHub Actions.
- New test coverage: 7 tests for the LLM client (env-key/auth errors, model-default regression pins) and 7 for TUI timeline windowing.
- Refactor: the seven TUI overlays (phase guide, help, hints, solution, test review, refactor, themes) and layout helpers moved out of PhaseTui.tsx (1627 -> 955 lines) into src/tui/overlays.tsx.
v0.1.7
- REDGREEN_ATTACK=1: headless runs now auto-advance through all three attack rounds (mirrors REDGREEN_REFACTOR=1), so unattended runs match the README's "survive three rounds" framing.
- CI on every push/PR: typecheck, tests, and build now run in GitHub Actions.
- Removed the stale and misleading specsheet.md planning doc (Supabase/Vercel stack, unimplemented runners, nonexistent login command).
v0.1.6
Model defaults now point at current, non-retired models per provider:
- OpenAI: gpt-4o -> gpt-5.2 (gpt-4o base sunsets Oct 23, 2026)
- Anthropic: claude-3-5-sonnet -> claude-sonnet-5 (bare name never valid)
- Gemini: gemini-1.5-pro -> gemini-3.5-flash (1.5-pro retired, 2.5-pro retires Oct 20)
The model is editable in the init wizard or via the model key in the config file (now documented in the README).
v0.1.5
Doc-only release: README now covers the tutorial coach overlay (single-press phase advance) and the corrected keybindings (s = skip, S = solution, r = reject/retry, j/k/g scrolling, / commands, t = themes).
v0.1.4
Adds the guided
edgreen tutorial\ command (RED -> GREEN -> attacks -> refactor with a per-phase coach overlay), line-based scrolling TUI, and session hardening: absolute cwd resolution, GREEN watcher no longer aborts on approve, and no repeated auto-apply of consumed refactor suggestions.