Release Agent! 1.1.87.287 · AgentiLoop/Agent

Agent! 1.1.87 — Release

New /auto <goal> mode: goal-driven, unattended cycles with a time budget (8db7ce86)
Runs with no cycle or iteration cap, works on LLM tabs, keeps a goal history (/auto last, /auto #N) (7f76c106)
Sessions survive app restarts and resume on the same tab (0a30275e)
Esc stops only the current cycle. Stop All (or /auto stop all) ends the session (bcdb4ef9, e6c923de)

Fluxion AI, with OpenAI and Anthropic protocol options (10934543, c28ac4e6)
Muse Code, which reuses your muse login subscription (f684a61e)
Requesty, A2Agent, OrcaRouter and Qwen Code (Coding Plan) (dd11db62, 067665af, e136cb5a, 8b314d08)
fm serve (experimental): Apple Foundation Models over a local Chat Completions API (5d2f1adc)
Vision detection now comes from provider catalog metadata, so Force Vision is no longer needed (7ea881e8)
APIProvider is now the single provider registry; per-provider settings and state were collapsed into one code path (b019fd46 … 09f5afc2)

The critic review is enforced: an unchanged diff is refused, a changed diff is re-reviewed, and issues can't be dismissed as "out of scope" (2f2bfb78)
The critic now supports Codex and Apple Intelligence, and logs the issues it found and whether the code changed afterward (53324247, 55f83547)
New Jev (TypeSafe System One) decision layer advises the tool loop. Settings are in the new LLM Common Settings (337c94a2, 42503723)

Thresholds are sized from the model actually in use (tab or fallback). Fetched Ollama context windows persist, which fixes a 16K threshold (c5db2b4b, 162590d6)
The kept tail and microcompact are bounded by token budget, and oversized blocks are capped before compaction (b7a80ffc, 289075e1, ad395cb4)
Context-overflow and max_tokens detection work across providers (2e67eddc)
The context bar shows the last request's input tokens. Tab token counters reset per task (34c92da5)

macOS 14.6+ supported on Apple Silicon and Intel. FoundationModels is gated behind macOS 26 (ea5ce624, 3a993205)
Localized into Spanish, French, German, Chinese (Simplified), Russian, Korean and Japanese (7e923fec)
Fixed activity log, Steps and HUD output bleeding across tabs. The main tab keeps its own persisted provider (78dfbaa3, 8c5e2974, 5582df0b, 0b7079fe)
Copy-to-clipboard for API keys, clear/restore buttons for recent folders, and a separate Agent!-debug.app / .debug bundle ID for Debug builds (957ecf95, cc797e2e, 98088c3b)
New accessibility actions: wait_until_actionable, select_text_range and observe_start/poll/stop/list (02992255)

Recursive deletion of the current project folder is blocked (fa0ab6f0)
App-wide bug hunt: fixed PlanMode step matching, an Ollama < streaming stall and a ShellSafety & read-only bypass. Also fixed CodingService range crashes (8c9d63a6, 7dce8fe7)
task_complete is refused when the summary points to output that was never written (1f9410f0, e8848437)
Fixed duplicate tool_call_ids on parallel calls. Local LLMs that run out of memory now stop immediately with a clear reason (cb6daac8, c2573a8a)

Install with Homebrew: brew update && brew install --cask agentiloop-agent (2cb41b91)
Relicensed to PolyForm Noncommercial 1.0.0. Added a CLA, Code of Conduct, issue/PR templates and a sponsorship program (fc72aaf5, 9c42e19c, 98e8fafc, f459372c)
Bundle IDs and the team are driven by Project.xcconfig (daaccbb5)
READMEs are now in 8 languages, with the AgentiLoop CLI (Rust + Go), the website and a star history chart (92bb9839, 61a69ce6, c140587e)
Swift 6.4 / Xcode 27 CI. All SPM packages raised to their latest tags (1ec241f8, fb7d9cdc)

Full Changelog: v1.1.33.233...v1.1.87.287