What shipped

All notable changes to G-Rump are documented here. The format follows Keep a Changelog and versions follow Semantic Versioning.

[Unreleased]

The reliability-and-quality release: the agent loop stops trusting "the model went quiet" as proof of "done", Claude Fable 5 is fully supported, adaptive thinking raises coding quality on every current Claude model, runs can go twice as long, and chat responses got a rendering overhaul — faster, more correct markdown, aligned tables, JSON highlighting, and reopenable thought processes.

Added

• Claude Fable 5 works end-to-end. Fable signs every reasoning block and expects them back: the app now captures native thinking blocks (including redacted ones) from the stream and replays them unchanged — first in the assistant turn — so tool-use continuations no longer risk rejection. When Fable's safety classifiers decline a request (HTTP 200, refusal stop reason), the chat says so instead of stopping silently.
• Adaptive thinking on by default. Anthropic requests on models that support it (Opus 4.6+, Sonnet 4.6+, Sonnet 5, Fable 5) now ask for thinking: adaptive — the recommended mode for coding and agentic work. Opus 4.8, the app default, previously ran with thinking off entirely.
• Mode switches flow mid-conversation. Changing Plan/Build/Spec with ⇧⇥ or the status-bar switcher now drops an explicit pivot note into the conversation, so the model carries context forward under the new mode instead of silently receiving different instructions.
• Longer autonomous runs. Max agent steps default doubled to 400 and the ceiling raised to 2,000 (Settings → Behavior).
• Response quality contract. The default system prompt now spells out how answers should read: lead with the outcome, calibrate structure to the question, complete sentences over fragment chains, no filler openers, file:line code references, and honest verified-vs-should-work reporting. A new Answering-vs-Acting rule keeps questions from triggering unasked edits — problems get diagnosed and a fix proposed, applied only on request.
• Completion gate. When a run that changed code (or has open plan steps) tries to finish, a fast outside check audits the original request first — open plan steps block completion deterministically; otherwise a light-model judge reviews changed files, build status, and the final message. Fails open, max 2 re-entries, off-switch in Settings → Behavior.
• Auto-verify after edits. At the completion point, if the run edited code and never proved a green build, the project builds automatically (ecosystem auto-detect or buildCommand in .grump/config.json) and failures go back to the agent for fixing — max 3 cycles, then an honest give-up note. Tests are strictly opt-in via testCommand.
• update_plan tool. The agent keeps a tracked checklist (pending / in_progress / done); the current plan rides along on every request and Build mode is instructed to plan first on multi-step tasks.
• Anthropic prompt caching. Three ephemeral breakpoints (system, tools, advancing message prefix) cut per-step cost and latency on long runs. Kill switch: AnthropicPromptCachingEnabled.
• Rolling context compaction. Long runs summarize their oldest turns (light model) instead of hard-dropping them; the original request is pinned and always survives truncation.
• Real eval battery. scripts/agent-eval.mjs runs 8 agentic coding tasks against the Anthropic API with the app's actual tool schemas (GRump --dump-tools), graded deterministically; completion-rate history lands in evals/history.jsonl.
• Thought process on past messages. Reasoning traces captured from thinking models were persisted but invisible — completed assistant messages now show a collapsed "Thought process" disclosure you can reopen any time (redacted blocks stay hidden).
• Markdown tables, done properly. Separator alignment (:---, :---:, ---:) is respected per column, ragged rows normalize to the header's column count (short rows pad, overflow folds into the last cell), and data rows zebra-stripe for scannability.
• JSON syntax highlighting. json / jsonc / json5 / jsonl / ndjson blocks highlight literals, strings, and numbers instead of rendering flat — the single most common block type in AI output.
• More inline markdown. _italic_ and __bold__ (word-boundary aware, so snake_case stays plain) and backslash escapes (\* renders a literal asterisk).

Fixed

• Shell tools were blind to stderr. run_command/run_build/friends read stderr but never returned it — the agent literally could not see compiler errors. stderr is now included and non-zero exits carry a deterministic [exit code: N] marker.
• Deep-run stream deaths. Transient stream errors now retry per turn (with backoff) instead of killing any run past step 3.
• edit_file ambiguity. Multi-location matches now error (with a replace_all opt-out) instead of silently replacing everything, and a whitespace-tolerant fallback handles indentation drift — applied only when unambiguous and flagged for re-verification.
• .grump/context.md was ignored for projects without a config.json.
• Truncation could orphan tool results (API 400s deep into long runs).
• Streaming could duplicate text fragments. The incremental markdown parser estimated block offsets to skip re-parsing, but the estimates undercounted blank lines — mid-stream re-parses could start inside already-rendered text and repeat it. Streaming now fully re-parses off the main thread (still debounced to ~60fps); correct by construction.
• Emphasis false positives. 2 * 3 * 4 no longer italicizes " 3 " — emphasis delimiters require non-whitespace flanking.
• Code blocks clipped at larger content sizes. Line heights were hardcoded for the default text scale; they now scale with the content-size preference, keeping the line-number gutter and code aligned.
• The streaming cursor vanished on long code blocks once the live view truncated to the last 8 lines (it compared against the untruncated list).

Changed

• Panel dock is now a single source of truth. The right-edge sidebar renders PanelTab.dockGroups, and tests assert every panel appears in the dock exactly once — a new panel can no longer be silently unreachable.
• Chat rendering is dramatically cheaper. Inline formatting batches plain text runs instead of appending one attributed character at a time (long paragraphs: thousands of appends → a handful); syntax-highlighted code renders one concatenated Text per line instead of one view per token; long messages no longer re-split their full content on every hover.

Removed

• Dead layout settings. panelAlignment, quickInputPosition, secondaryActivityBarVisible, and secondarySidebarVisible persisted to UserDefaults but were consumed nowhere; they are gone, along with a never-rendered "Quick Input Position" customizer section.
• The duplicate "Markdown" copy button on assistant messages — it was byte-identical to Copy.
• The Run button on unlabeled code blocks. Only explicitly tagged shell blocks (bash, zsh, sh, …) offer Run in Terminal; a block with no language tag no longer grows a button that silently executes arbitrary text.

[2.1.0] - 2026-07-14

2.1 turns G-Rump from a chat app with tools into a small IDE with an agent in it: a real build engine with run-to-simulator, a project navigator, an experimental self-learning loop, and local models via Ollama — no key needed. Honest note: the newest surfaces (build & run, learning loop) are lightly battle-tested. They work; they will also surprise you occasionally. The roadmap tracks what's solid and what's rough.

Added

• Build engine + run-to-simulator. Xcode-style toolbar above the chat: ⌘R builds via xcodebuild or SPM and, on a simulator destination, continues install → launch → live app logs streamed into the new Build console panel (Log and Issues tabs, Fix-with-G-Rump, reveal-in-navigator, xed --line). ⌘⇧. stops. The agent can drive the same loop through xcrun_simctl tools.
• Project navigator. ⌘0 toggles a left file-tree pane; build issues and agent file references reveal straight into it.
• Self-learning loop (experimental). Every run's outcome lands in a per-project ledger; a reflection pass distills short, confidence-scored lessons that ride along on future prompts; strong lesson clusters can become skill proposals you approve or reject as diffs in the new Learning panel (the 20th dock panel). Off-switch in Settings → Brain. Writes to SOUL.md, MIND.md, and skill directories always require approval.
• Local Ollama provider. Fifth provider, keyless: models are discovered live from localhost:11434, tool support is gated per model, and being offline is fine. Ollama was cut as dead code in 2.0.0; this is a fresh, working implementation. Never auto-routed.
• Welcome window. Xcode-style launch window (⇧⌘1): pinned recents, open, clone (streaming, cancellable), new project.
• Developer profile. Profile → You feeds a capped block (name, role, stack, conventions) into the system prompt; Profile → Your Agent embeds the SOUL editor.
• Settings as a real window. Native Settings{} scene on ⌘, — 21 tabs regrouped into 7 groups. All legacy entry points still work.
• Mode-select card. New conversations open with an inline Plan / Build / Spec card gating the first message; ⇧⇥ cycles modes from the input; the status bar shows the current mode with a switcher.
• Onboarding rework. Six typed steps with real gating; provider step validates keys inline (or probes Ollama reachability) and allows an explicit "add a key later" deferral. The skills-allowlist overwrite bug is dead.

Changed

• App icons regenerated on the proper 824pt Apple squircle grid.
• Docs (README, ARCHITECTURE, CLAUDE.md) rewritten against counted reality: 160 native tools, 67 MCP presets, 20 panels, 73 skills, ~67K LOC, 1,581 tests.

Removed

• Empty-state suggestion chips and the mode buttons row (replaced by the mode-select card); orphaned ChatAreaView, SpecQuestionsModal, RightPanelManager.

2.0.0 - 2026-07-09

G-Rump 2.0 is the multi-provider release. The app is no longer tied to a single AI vendor: bring your own key for Anthropic (default), OpenAI, Google, or OpenRouter, and G-Rump speaks each provider's native wire format with full streaming tool calls. The release also lands the cognitive-memory engine, the brain subsystems, and a large platform cleanup that removes the SaaS billing layer entirely — G-Rump is now a local-first app with your keys in the macOS Keychain and no accounts.

Added

• Multi-provider AI. Anthropic (default), OpenAI, Google, and OpenRouter. Anthropic and Gemini use native wire formats with streaming tool calls; OpenAI and OpenRouter ride a shared parameterized OpenAI-compatible transport.
• 2026 model catalog. Claude Opus 4.8 (default), Claude Fable 5 (premium, manual select only), Claude Sonnet 5, Claude Haiku 4.5, GPT-5.2, GPT-5.3-Codex, Gemini 3 Pro, Gemini 2.5 Flash, plus OpenRouter routes (Claude Sonnet 5, GPT-5.3-Codex, Gemini 3 Pro, Qwen3 Coder).
• API key validation on save. Each saved key gets a cheap authenticated probe with inline verified / rejected / could-not-verify feedback in both Settings and onboarding. Keys always save first; a failed probe warns and never blocks.
• Provider-aware model routing. Task routing picks sensible heavy/light chains within the active provider. Fable 5 is never auto-selected.
• Automatic settings migration. One-shot migration of Qwen-era configuration: provider and model IDs are remapped and stray keys are hoisted into the Keychain. No manual steps.
• Cognitive memory (Track 1: MemoryAgent). Cross-session memory with relevance x recency x salience ranking, recall within a fixed token budget, and deliberate forgetting of stale context.
• Brain subsystems. Vault, voice, eyes, mind, and daemon fused into the app.
• Richer chat rendering. Task lists, inline code pills, images, H4-H6 headings, improved streaming code blocks, expandable tool timeline with line counts and copy buttons, per-message word count, and copy-as-markdown.
• Keyboard shortcuts, conversation search, collapsible messages, and diff copy.
• Expanded syntax highlighting. SQL, YAML, HTML, CSS, and Dockerfile, plus decorator support.
• Code signing and notarization pipeline. make sign / make package / make notarize for signed, notarized distribution builds.
• Evaluation harness. scripts/judge-verify.mjs (chat, multi-turn tool calling, embeddings) and scripts/agent-eval.mjs (4-task agent battery on a real tool loop).
• CI hardening. SwiftLint strict configuration; green pipeline across tests, lint, backend tests, and release build. Suite currently at 1,449 tests.

Changed

• Default model is Claude Opus 4.8 (was Qwen Coder Plus).
• API keys live exclusively in the macOS Keychain, one account per provider. Keys are never written to UserDefaults or any config file.
• Agent modes consolidated from seven to three: Plan, Build, Spec.
• The Node.js backend is now an optional stateless proxy; the app calls providers directly by default.
• Anthropic requests pin anthropic-version: 2023-06-01 and omit temperature (Claude 4.7+/5 models reject it).
• Themes renamed to neutral names: Snow, Linen, Graphite, Amethyst, Ink.
• Relicensed under MIT.

Fixed

• Anthropic tool arguments now stream. input_json_delta chunks were dropped, so every Anthropic tool call arrived with empty arguments.
• Anthropic multi-turn tool loops no longer 400. System prompts ride the top-level system field, assistant turns emit tool_use blocks, and max-tokens is set per request.
• Gemini tool loops work past turn one. Tool results are now sent as functionResponse parts.
• Stop-reason normalization. All providers normalize to the same tool_calls / stop signals so the agent loop drives identically everywhere.
• Settings key foot-gun removed. The Account tab's "OpenRouter API key" field silently wrote to whichever provider was active; per-provider keys live in the Providers tab.
• Headless test crash and all pre-existing test failures; the suite runs green.
• 48 deployment gaps closed across security, legal, features, backend, and accessibility ahead of distribution.

Removed

• Billing / platform SaaS layer. JWT auth replaced by an optional APP_API_KEY gate on the proxy.
• OpenClaw chat-routing subsystem.
• GRumpServer SaaS target.
• Ollama, on-device CoreML, and model-mode dead code.
• "Local Only Mode" privacy toggle and "Fully Local" badge (they gated nothing).
• Legacy AIModel enum; the typed model catalog is the single source of truth.

Security

• Shell-injection vectors eliminated — subprocess calls use Process() argument arrays instead of interpolated shell strings.
• MCP hardening: path-traversal protection and authentication on tools/list.
• ProviderConfiguration excludes the API key from serialization, so keys cannot leak into UserDefaults backups.
• Key-validation probes send keys in headers only (Google via x-goog-api-key), never in URLs where they could reach logs.

1.0.0 - 2026-02-24

Initial release: a native macOS/iOS autonomous AI coding agent — chat with streaming responses, 100+ local tools (file, shell, git, Docker, browser, Apple-native), MCP client and server, skills and SOUL.md personality systems, approval-gated autonomous daemon, and dual-mode persistence (SwiftData/JSON).