mistake

.github/workflows/ci.yml

@@ -30,3 +30,46 @@ jobs:
 
       - name: Run ${{ matrix.task }}
         run: ${{ matrix.command }}
+
+  # KR-eval regression run (replay mode — recorded tool fixtures, no DART/KRX/Naver network;
+  # LLM keys for agent + judge come from repo secrets). This is OPT-IN: it runs only when a
+  # PR carries the `run-eval` label, NOT on every main push. Rationale: agent + judge are
+  # live LLM calls (token cost per run), and KR_EVAL_REQUIRE_RUN makes a keyless run FAIL —
+  # so gating main pushes on it would turn main red until OPENAI_API_KEY (etc.) is set as a
+  # repo secret. Keeping it label-only keeps main green with zero token cost by default.
+  # To re-enable an always-on gate once secrets are configured: add `github.event_name ==
+  # 'push'` (and restore the weekly `schedule` cron in `on:`) to the condition below.
+  #
+  # MAINTENANCE: fixtures are verbatim recorded tool outputs — whenever a tool's OUTPUT
+  # SHAPE changes (new fields the skills/prompts instruct the agent to read), re-record
+  # with `bun run kr-eval:record` in a keyed environment, or replay answers will hedge on
+  # missing fields and judge scores will drift.
+  kr-eval-replay:
+    if: >-
+      github.event_name == 'pull_request' && github.event.action == 'labeled' && github.event.label.name == 'run-eval'
+    runs-on: ubuntu-latest
+    # Rapid successive merges: only the newest main run carries signal — cancel superseded ones.
+    concurrency:
+      group: kr-eval-replay-${{ github.ref }}
+      cancel-in-progress: true
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - name: Install dependencies
+        run: bun install --frozen-lockfile --ignore-scripts
+
+      - name: KR eval (replay)
+        run: bun run kr-eval:replay
+        env:
+          KR_EVAL_REQUIRE_RUN: '1'
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
+          # Optional repo variables to pin models to whatever key is configured
+          # (defaults: agent + judge = gpt-5.5, which needs OPENAI_API_KEY).
+          KR_RUNNER_MODEL: ${{ vars.KR_RUNNER_MODEL }}
+          KR_EVAL_JUDGE_MODEL: ${{ vars.KR_EVAL_JUDGE_MODEL }}

.github/workflows/deploy-pages.yml

@@ -0,0 +1,33 @@
+name: Deploy landing page
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'landing/**'
+      - '.github/workflows/deploy-pages.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+concurrency:
+  group: pages
+  cancel-in-progress: true
+
+jobs:
+  deploy:
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/configure-pages@v5
+      - uses: actions/upload-pages-artifact@v3
+        with:
+          path: landing
+      - id: deployment
+        uses: actions/deploy-pages@v4

.github/workflows/release.yml

@@ -0,0 +1,67 @@
+name: Release Desktop
+
+# Build the desktop app on each OS's own runner so the bundled bun binary matches
+# the target platform (prepare-core.mjs stages the host's bun). Pushing a tag like
+# `v1.0.0` publishes a DRAFT GitHub Release with the dmg/exe; review then publish.
+on:
+  push:
+    tags: ['v*']
+  workflow_dispatch: # manual run → uploads installers as workflow artifacts (no publish)
+
+# electron-builder needs write access to create/upload the draft GitHub Release.
+permissions:
+  contents: write
+
+jobs:
+  build:
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: macos-14 # Apple Silicon (arm64)
+          - os: windows-latest # x64
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: latest
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      # Core deps for the staged sidecar. --ignore-scripts skips the playwright/
+      # better-sqlite3 build (the sidecar uses bun:sqlite, no chromium).
+      - name: Install core dependencies
+        run: bun install --frozen-lockfile --ignore-scripts
+
+      # Desktop deps. postinstall runs electron-rebuild for better-sqlite3 against
+      # this runner's Electron ABI — must happen on the target OS.
+      - name: Install desktop dependencies
+        working-directory: desktop
+        run: npm ci
+
+      - name: Stage core, build, package${{ startsWith(github.ref, 'refs/tags/') && ' & publish' || '' }}
+        working-directory: desktop
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          # Unsigned for now. To sign: add CSC_LINK/CSC_KEY_PASSWORD (mac) and
+          # WIN_CSC_LINK/WIN_CSC_KEY_PASSWORD (win) as repo secrets and drop this.
+          CSC_IDENTITY_AUTO_DISCOVERY: 'false'
+        run: |
+          npm run stage
+          npx electron-vite build
+          npx electron-builder --publish ${{ startsWith(github.ref, 'refs/tags/') && 'always' || 'never' }}
+
+      - name: Upload installers
+        uses: actions/upload-artifact@v4
+        with:
+          name: dexter-${{ matrix.os }}
+          path: |
+            desktop/release/*.dmg
+            desktop/release/*.exe
+            desktop/release/*.blockmap
+            desktop/release/latest*.yml
+          if-no-files-found: warn

.gitignore

@@ -60,4 +60,7 @@ logs/
 coverage.json
 
 # Scratchpad files
-scratchpads/*
+scratchpads/*
+
+# Claude Code runtime locks
+.claude/scheduled_tasks.lock

AGENTS.md

@@ -1,24 +1,24 @@
 # Repository Guidelines
 
 - Repo: https://github.com/virattt/dexter
-- Dexter is a CLI-based AI agent for deep financial research, built with TypeScript, Ink (React for CLI), and LangChain.
+- Dexter is a CLI-based AI agent for deep financial research, built with TypeScript, a terminal UI (`@mariozechner/pi-tui`), and LangChain.
 
 ## Project Structure
 
 - Source code: `src/`
   - Agent core: `src/agent/` (agent loop, prompts, scratchpad, token counting, types)
-  - CLI interface: `src/cli.tsx` (Ink/React), entry point: `src/index.tsx`
-  - Components: `src/components/` (Ink UI components)
-  - Hooks: `src/hooks/` (React hooks for agent runner, model selection, input history)
+  - CLI interface: `src/cli.ts` (pi-tui; no JSX), entry point: `src/index.tsx` (the `.tsx` extension is historical)
+  - Components: `src/components/` (pi-tui UI components, plain `.ts`)
   - Model/LLM: `src/model/llm.ts` (multi-provider LLM abstraction)
-  - Tools: `src/tools/` (financial search, web search, browser, skill tool)
+  - Tools: `src/tools/` (financial data, web search, browser, skill tool)
   - Tool descriptions: `src/tools/descriptions/` (rich descriptions injected into system prompt)
   - Finance tools: `src/tools/finance/` (prices, fundamentals, filings, insider trades, etc.)
-  - Search tools: `src/tools/search/` (Exa preferred, Tavily fallback)
+  - Korean tools: `src/tools/finance-kr/` (DART/KRX/Naver — financials, filings, foreign ownership, short balance, NPS, market data)
+  - Search tools: `src/tools/search/` (Exa → Perplexity → Tavily → LangSearch fallback chain)
   - Browser: `src/tools/browser/` (Playwright-based web scraping)
   - Skills: `src/skills/` (SKILL.md-based extensible workflows, e.g. DCF valuation)
   - Utils: `src/utils/` (env, config, caching, token estimation, markdown tables)
-  - Evals: `src/evals/` (LangSmith evaluation runner with Ink UI)
+  - Evals: `src/evals/` (LangSmith evaluation runner). KR eval harness: `scripts/kr-eval/` (record/replay + LLM-judge)
 - Config: `.dexter/settings.json` (persisted model/provider selection)
 - Environment: `.env` (API keys; see `env.example`)
 - Scripts: `scripts/release.sh`
@@ -27,16 +27,17 @@
 
 - Runtime: Bun (primary). Use `bun` for all commands.
 - Install deps: `bun install`
-- Run: `bun run start` or `bun run src/index.tsx`
+- Run: `bun start`
 - Dev (watch mode): `bun run dev`
 - Type-check: `bun run typecheck`
 - Tests: `bun test`
 - Evals: `bun run src/evals/run.ts` (full) or `bun run src/evals/run.ts --sample 10` (sampled)
+- KR evals: `bun run kr-eval` (live) / `kr-eval:record` / `kr-eval:replay` (deterministic, no keys)
 - CI runs `bun run typecheck` and `bun test` on push/PR.
 
 ## Coding Style & Conventions
 
-- Language: TypeScript (ESM, strict mode). JSX via React (Ink for CLI rendering).
+- Language: TypeScript (ESM, strict mode). The CLI renders via `@mariozechner/pi-tui` (no JSX in `src/cli.ts`).
 - Prefer strict typing; avoid `any`.
 - Keep files concise; extract helpers rather than duplicating code.
 - Add brief comments for tricky or non-obvious logic.
@@ -45,26 +46,28 @@
 
 ## LLM Providers
 
-- Supported: OpenAI (default), Anthropic, Google, xAI (Grok), OpenRouter, Ollama (local).
+- Supported: OpenAI (default), Anthropic, Google, xAI (Grok), OpenRouter, Moonshot, DeepSeek, Ollama (local).
 - Default model: `gpt-5.5`. Provider detection is prefix-based (`claude-` -> Anthropic, `gemini-` -> Google, etc.).
 - Fast models for lightweight tasks: see `FAST_MODELS` map in `src/model/llm.ts`.
 - Anthropic uses explicit `cache_control` on system prompt for prompt caching cost savings.
 - Users switch providers/models via `/model` command in the CLI.
 
 ## Tools
 
-- `financial_search`: primary tool for all financial data queries (prices, metrics, filings). Delegates to multiple sub-tools internally.
-- `financial_metrics`: direct metric lookups (revenue, market cap, etc.).
+- `get_financials`: US financial statements, metrics, segments (Financial Datasets). Routes multi-company/metric queries internally.
+- `get_market_data`: US prices, company news, insider trades, 13F holdings.
 - `read_filings`: SEC filing reader for 10-K, 10-Q, 8-K documents.
-- `web_search`: general web search (Exa if `EXASEARCH_API_KEY` set, else Tavily if `TAVILY_API_KEY` set).
+- `stock_screener`: screen US stocks by financial criteria (P/E, growth, margins).
+- Korean tools (`src/tools/finance-kr/`, all take 6-digit tickers): `get_financials_kr`, `get_filings_kr`, `get_large_holders_kr`, `get_insider_trades_kr`, `read_filings_kr` (DART, gated on `DART_API_KEY`); `get_market_data_kr`, `get_foreign_ownership_kr` (Naver, keyless); `get_short_balance_kr` (KRX login); `get_nps_holdings` (data.go.kr).
+- `web_search`: general web search; provider fallback chain Exa → Perplexity → Tavily → LangSearch (registered if ≥1 key set).
 - `browser`: Playwright-based web scraping for reading pages the agent discovers.
 - `skill`: invokes SKILL.md-defined workflows (e.g. DCF valuation). Each skill runs at most once per query.
-- Tool registry: `src/tools/registry.ts`. Tools are conditionally included based on env vars.
+- Tool registry: `src/tools/registry.ts`. US finance + screener tools register unconditionally (they 401 at call-time without `FINANCIAL_DATASETS_API_KEY`); KR, web_search, and X tools are gated by their env vars.
 
 ## Skills
 
 - Skills live as `SKILL.md` files with YAML frontmatter (`name`, `description`) and markdown body (instructions).
-- Built-in skills: `src/skills/dcf/SKILL.md`.
+- Built-in skills (`src/skills/<name>/SKILL.md`): `dcf` (DCF valuation; branches US/KR internally), `kr-spinoff` (물적/인적분할 analysis), `write-memo`, `x-research`.
 - Discovery: `src/skills/registry.ts` scans for SKILL.md files at startup.
 - Skills are exposed to the LLM as metadata in the system prompt; the LLM invokes them via the `skill` tool.
 
@@ -73,15 +76,16 @@
 - Agent loop: `src/agent/agent.ts`. Iterative tool-calling loop with configurable max iterations (default 10).
 - Scratchpad: `src/agent/scratchpad.ts`. Single source of truth for all tool results within a query.
 - Context management: Anthropic-style. Full tool results kept in context; oldest results cleared when token threshold exceeded.
-- Final answer: generated in a separate LLM call with full scratchpad context (no tools bound).
+- Final answer: the text of the turn where the model stops emitting tool calls (`handleDirectResponse`). Tools stay bound on every call — there is no separate no-tools finalization pass, so answer quality is governed by the system prompt.
 - Events: agent yields typed events (`tool_start`, `tool_end`, `thinking`, `answer_start`, `done`, etc.) for real-time UI updates.
 
 ## Environment Variables
 
-- LLM keys: `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `GOOGLE_API_KEY`, `XAI_API_KEY`, `OPENROUTER_API_KEY`
+- LLM keys: `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `GOOGLE_API_KEY`, `XAI_API_KEY`, `OPENROUTER_API_KEY`, `MOONSHOT_API_KEY`, `DEEPSEEK_API_KEY`
 - Ollama: `OLLAMA_BASE_URL` (default `http://127.0.0.1:11434`)
-- Finance: `FINANCIAL_DATASETS_API_KEY`
-- Search: `EXASEARCH_API_KEY` (preferred), `TAVILY_API_KEY` (fallback)
+- US finance: `FINANCIAL_DATASETS_API_KEY`
+- Korean data: `DART_API_KEY` (재무·공시), `KRX_ID`+`KRX_PW` or `KRX_COOKIE` (공매도), `DATA_GO_KR_SERVICE_KEY` (국민연금). Naver tools (외국인·현재가) need no key.
+- Search: `EXASEARCH_API_KEY`, `PERPLEXITY_API_KEY`, `TAVILY_API_KEY`, `LANGSEARCH_API_KEY` (fallback chain in that order)
 - Tracing: `LANGSMITH_API_KEY`, `LANGSMITH_ENDPOINT`, `LANGSMITH_PROJECT`, `LANGSMITH_TRACING`
 - Never commit `.env` files or real API keys.