Dev

.github/workflows/release.yml

@@ -52,6 +52,19 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: gh release edit ${{ github.ref_name }} --draft=false
 
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@06116385d9baf250c9f4dcb4858b16962ea869c3 # v4
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@d7f5e7f509e45cec5c76c4d5afdd7de93d0b3df5 # v4
+
+      - name: Log in to GHCR
+        uses: docker/login-action@650006c6eb7dba73a995cc03b0b2d7f5ca915bee # v4
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
       - name: Publish to npm and registries
         env:
           TINYCODE_VERSION: ${{ github.ref_name }}

CHANGELOG.md

@@ -2,19 +2,51 @@
 
 ## [Unreleased]
 
+## [1.17.1] — 2026-08-10
+
 ### Added
+- OpenRouter as full-fledged provider with auto-discovery — set `OPENROUTER_API_KEY` to discover 300+ models with capability detection, pricing, and cost tracking
+- OpenRouter in `/connect` TUI dialog for API key entry
 - Tool-call failure tracking with warning toast after 3+ consecutive failures (#69)
 - Auto-compact on model switch to smaller context window (#74)
 - Model size field in config for custom/finetuned models (#75)
 - Context warning for models with <8K context window (#77)
 
 ### Changed
+- omt tools and system prompt injection moved from plugin to native registration
 - Skip tool injection for models with `capabilities.toolcall=false` to save context tokens (#69)
 - Sanitize markdown fences and trailing commas from malformed tool-call JSON (#69)
 
 ### Fixed
+- oh-my-tiny tool schemas: nested `enum` arrays broke OpenAI strict JSON Schema validation
 - Surface friendly error on ECONNREFUSED instead of stack trace (#76)
 
+## [1.17.0] — 2026-08-07
+
+### Added
+- Unified command palette (`Ctrl+P`) searches across commands, agents, sessions, and skills with frecency ranking (#59)
+- Subagent output rendered inline in parent session as collapsible preview (#87)
+- Leader-key session navigation using vim-style hjkl (`<leader>h/j/k/l`) — replaces bare arrow keys (#88)
+- OpenAPI spec and Scalar docs UI served at `/api/docs`
+- First 5 Minutes quick-start section in user guide
+- Comprehensive user manual and deployment verification guide
+
+### Changed
+- Session list moved from `<leader>l` to `<leader>o` (freed for sibling navigation)
+- Code block concealment moved from `<leader>h` to `<leader>;` (freed for sibling navigation)
+- System prompts: agent prompts compose with base instead of replacing
+- System prompts: tool infrastructure separated from personality
+
+### Fixed
+- Command palette crash causing unrecoverable stuck/dimmed TUI state (#86)
+- Non-interactive run mode (`bun dev run`) — SSE events now fully delivered before exit (#84)
+- Docker multi-platform build — `TARGETARCH` moved to global scope in Dockerfile (#83)
+- `modelSizeB` regex handles decimal model sizes like "3.5b" (#80)
+- ErrorComponent double-fault from invalid `new URL("")` constructor
+- Dialog-scoped ErrorBoundary for graceful dialog error recovery
+- Release workflow: added buildx setup for multi-platform Docker builds
+- File path whitespace trimming in read/edit/write tools
+
 ## [0.2.1] — 2026-07-06
 
 ### Added

CLAUDE.md

@@ -142,7 +142,8 @@ Auto-generated from the OpenAPI spec. Regenerate with `./packages/sdk/js/script/
 - **Dual runtime**: `src/storage/db.ts` and `src/pty/` use conditional imports (`#db`, `#pty`) to swap Bun vs Node implementations.
 - **`bun dev` = `tinycode`**: During development, `bun dev` from the repo root is equivalent to the `tinycode` CLI.
 - **Local LLMs first**: The primary use case is local LLM inference via Ollama, ramalama, or any OpenAI-compatible endpoint. Configure via `~/.config/tinycode/config.json`. ramalama support is enabled via `TINYCODE_RAMALAMA_HOST` env var (e.g., `http://localhost:8080`).
-- **Provider filtering**: `enabled_providers` and `disabled_providers` in config apply to **all** providers — both custom API-configured providers and locally-discovered ones (Ollama, vLLM, ramalama, MaaS). Filters apply during discovery, so disabled providers are completely hidden from the provider list.
+- **OpenRouter provider**: Full-fledged cloud provider with auto-discovery. Set `OPENROUTER_API_KEY` and tinycode discovers available models via the OpenRouter API, filtering to tool-capable models with pricing/capability mapping. Discovery runs in `src/provider/local-discovery.ts` alongside local provider polling. Uses `@openrouter/ai-sdk-provider` SDK. Cost tracking via OpenRouter's generation cost API (`src/session/llm.ts`).
+- **Provider filtering**: `enabled_providers` and `disabled_providers` in config apply to **all** providers — both custom API-configured providers and locally-discovered ones (Ollama, vLLM, ramalama, MaaS, OpenRouter). Filters apply during discovery, so disabled providers are completely hidden from the provider list.
 - **oh-my-tiny**: Built-in plugin at `src/omt/` providing extended orchestration tools (notepad, wiki, state management, AST grep).
 - **ACP mode**: Run with `bun dev acp` or `tinycode acp` to enable IDE integration via the Agent Client Protocol. Communicates via stdio with editor extensions. See `docs/acp-integration.md` for developer guide.
 - **Session tree sidebar** (`src/cli/cmd/tui/component/session-tree.tsx`): Toggleable ASCII tree showing session hierarchy. Press `<leader>b` in the TUI to toggle visibility. Sessions are organized by parent-child relationships.
@@ -159,5 +160,6 @@ Auto-generated from the OpenAPI spec. Regenerate with `./packages/sdk/js/script/
 - **Skill parameter substitution**: Skills can declare `params:` in SKILL.md frontmatter (e.g., `params: [name, language]`). Parameters map to `$1`, `$2` placeholders in the skill content, substituted by `SessionPrompt.command()`.
 - **Plugin marketplace**: Curated plugin registry at `src/plugin/registry.json` with search via `tinycode plugin-search [query]`. Plugin install resolves registry names to npm packages automatically.
 - **Graceful shutdown**: `tinycode serve` handles SIGTERM/SIGINT — disposes all instances, emits global disposed event for SSE clients, then drains HTTP connections with 25s timeout.
-- **Frecency ranking**: Command palette sorts by usage frequency via namespaced frecency store (`src/cli/cmd/tui/component/prompt/frecency.tsx`). File autocomplete also uses frecency.
+- **Frecency ranking**: Command palette sorts by usage frequency via namespaced frecency store (`src/cli/cmd/tui/component/prompt/frecency.tsx`). File autocomplete also uses frecency. The unified palette (`ctrl+p`) searches across commands, agents, sessions, and skills — all categories use frecency.
+- **Subagent output inline**: The `Task` TUI component renders subagent results as a collapsible block below the task header. Extracted from `<task_result>` XML in `props.output` using the `BlockTool`/`collapseToolOutput` pattern.
 - **tc-doctor skill** (`src/skill/defaults/tc-doctor/`): Self-diagnostic skill with 14+ checks covering Ollama install/health, model availability, tool-call capability probe, ramalama provider, Mac-specific checks (Metal, Rosetta, RAM pressure), tinycode↔Ollama integration, vLLM provider health, and disk space. All checks use pure bash — no python3 dependency. Run with `/tc-doctor`.

docs/UI-UX-REVIEW.md

@@ -66,15 +66,9 @@ The help dialog dumps raw markdown syntax — headers show as `## text`, tables
 
 **Files:** `keybind.ts:43`, `tui-schema.ts:31`
 
-### 6. Subagent navigation paradigm shift
+### 6. Subagent navigation paradigm shift — RESOLVED
 
-Parent → child: `ctrl+x j` (leader key chord)
-Inside child: `up` (parent), `left/right` (prev/next siblings)
-Arrow keys work because child view has no prompt, but this is non-obvious and undiscoverable without the SubagentFooter hint text.
-
-**Fix:** Add `<leader>k` for parent (mirror of `<leader>j`) alongside arrow keys.
-
-**Files:** `keybind.ts:98-101`, `subagent-footer.tsx:96-128`, `index.tsx:444-449`
+Vim-style hjkl navigation: `<leader>h` (prev sibling), `<leader>j` (child), `<leader>k` (parent), `<leader>l` (next sibling). Bare arrow keys removed to avoid conflicts with OS window management (#88).
 
 ### 7. "tinycode" branding remnants (user-visible)
 

docs/architecture.md

@@ -108,11 +108,13 @@ Plugin registry at `src/plugin/registry.json` provides a curated catalog of know
 
 Auto-discovery polls every 30 seconds with a 2-second probe timeout (`src/provider/local-discovery.ts`).
 
+OpenRouter is also auto-discovered when `OPENROUTER_API_KEY` is set. The probe fetches `https://openrouter.ai/api/v1/models` (5-second timeout), filters to tool-capable non-free models, and maps pricing/capabilities/limits. Uses `@openrouter/ai-sdk-provider` SDK with generation cost tracking via OpenRouter's billing API.
+
 On startup, tinycode sends a warmup probe to the configured Ollama model (`src/provider/warmup.ts`). The probe sends a tool-call request to `/api/chat` with `keep_alive: "10m"`, pre-loading the model into GPU memory and verifying tool-calling capability. Results are shown as a toast (TUI) or footer message (direct mode).
 
 ### Bundled Cloud Providers (via Vercel AI SDK)
 
-anthropic, openai, google, amazon-bedrock, azure, google-vertex, xai, mistral, groq, deepinfra, cerebras, cohere, openrouter, togetherai, perplexity, vercel, alibaba, gateway, gitlab, venice, and any generic openai-compatible endpoint.
+anthropic, openai, google, amazon-bedrock, azure, google-vertex, xai, mistral, groq, deepinfra, cerebras, cohere, openrouter, togetherai, perplexity, vercel, alibaba, gateway, gitlab, venice, and any generic openai-compatible endpoint. OpenRouter is the only cloud provider with auto-discovery — others require manual model selection from the bundled catalog.
 
 ### Plugin Providers
 

docs/cheatsheet.md

@@ -10,11 +10,11 @@ Quick reference for the most common keyboard shortcuts, agents, and commands.
 |-----|--------|
 | `Ctrl+C` or `Ctrl+D` or `<leader>q` | Exit tinycode |
 | `F1` | Open help dialog |
-| `Ctrl+P` | Show command palette |
-| `<leader>l` | List all sessions |
+| `Ctrl+P` | Command palette (commands, agents, sessions, skills) |
+| `<leader>o` | List all sessions |
 | `<leader>n` | Create a new session |
-| `<leader>j` / `<leader>k` | Go to first/parent child session |
-| `left` / `right` | Cycle to previous/next child session |
+| `<leader>h` / `<leader>l` | Previous/next sibling session |
+| `<leader>j` / `<leader>k` | Go to first child / parent session |
 | `<leader>m` | List available models |
 | `F2` / `Shift+F2` | Cycle to next/previous recent model |
 | `<leader>a` | List available agents |
@@ -28,7 +28,7 @@ Quick reference for the most common keyboard shortcuts, agents, and commands.
 | `<leader>x` | Export session transcript |
 | `<leader>y` | Copy message |
 | `<leader>u` / `<leader>r` | Undo/redo message |
-| `<leader>h` | Toggle code block concealment |
+| `<leader>;` | Toggle code block concealment |
 | `Ctrl+R` | Rename session |
 | `Ctrl+D` | Delete session |
 | `Escape` | Interrupt current session |
@@ -133,11 +133,12 @@ Follow the suggestions
 Write guide for new developers on session architecture
 ```
 
-### Session navigation
+### Session navigation (vim-style hjkl)
 ```
+<leader>h          # Go to previous sibling session
 <leader>j          # Go to first child session
-<leader>k / up     # Go to parent session
-left / right       # Cycle through siblings
+<leader>k          # Go to parent session
+<leader>l          # Go to next sibling session
 ```
 
 ## Configuration
@@ -146,7 +147,7 @@ Edit `~/.config/tinycode/config.json`:
 
 ```json
 {
-  "model": "ollama/qwen3.5:9b",
+  "model": "ollama/qwen2.5:latest",
   "lsp": true
 }
 ```
@@ -175,7 +176,7 @@ After listing models (`<leader>m`):
 - Use `<leader>c` to compact a session before exporting
 - Export to HTML with `tinycode export --format html <session-id>`
 - Type `@filename` to reference a file in your prompt
-- Use `<leader>h` to collapse code blocks and focus on analysis
+- Use `<leader>;` to collapse code blocks and focus on analysis
 - Session tree shows hierarchy with `<leader>b`
 - **Model warmup:** On startup, tinycode probes the Ollama model to verify tool-call support and pre-load it into GPU memory. Look for "qwen3.5:9b ready — tool calling supported" in the toast/footer
 - **Tool-call warnings:** If you see "Multiple tool call failures detected," switch to a larger model via `<leader>m` — tinycode auto-repairs common JSON issues, but very small models may not support tool calling at all