fix(dashboard): status-text shrinks correctly when sibling text grows

.env.public

@@ -0,0 +1,29 @@
+# llm-relay public environment defaults (git tracked)
+# Shared by Docker (docker-compose.yml) and systemd (llm-relay-api.service).
+# Local overrides go in .env.local (gitignored).
+
+# ── Core ──────────────────────────────────────────────────────────
+LLM_RELAY_UPSTREAM=https://api.anthropic.com
+
+# ── Feature toggles ──────────────────────────────────────────────
+LLM_RELAY_GB_INTERCEPT=0
+LLM_RELAY_CACHE_FIX=0
+LLM_RELAY_SCAN_ENABLED=0
+LLM_RELAY_TOKPRESS=0
+LLM_RELAY_SSE_PARSE_USAGE=1
+
+# ── Session history (composition pie chart, SNR, duplicate reads) ─
+LLM_RELAY_HISTORY=1
+LLM_RELAY_HISTORY_RAW=0
+
+# ── Internationalization (en/ko, browser auto-detect on web) ─────
+LLM_RELAY_LANG=en
+
+# ── Turn counter thresholds (CC) ─────────────────────────────────
+# Zone A — absolute token count, aligned with Zone B (50/70/90/100% of 1M)
+LLM_TOKEN_A_YELLOW=500000
+LLM_TOKEN_A_ORANGE=700000
+LLM_TOKEN_A_RED=900000
+LLM_TOKEN_A_HARD=1000000
+# Zone B — context window ceiling (for ratio calculation)
+LLM_TOKEN_CEILING=1000000

.github/workflows/ci.yml

@@ -0,0 +1,52 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install ruff
+      - run: ruff check src/
+
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Install dependencies
+        run: |
+          if python -c "import sys; sys.exit(0 if sys.version_info >= (3, 10) else 1)"; then
+            pip install -e ".[dev]"
+          else
+            pip install -e ".[cli,proxy]" pytest ruff
+          fi
+      - run: pytest tests/ -v --tb=short
+
+  build:
+    runs-on: ubuntu-latest
+    needs: [lint, test]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install build
+      - run: python -m build

.github/workflows/docker.yml

@@ -0,0 +1,49 @@
+name: Docker
+
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: "Image tag (e.g., 0.8.0)"
+        required: true
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: docker/setup-qemu-action@v3
+      - uses: docker/setup-buildx-action@v3
+
+      - uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - uses: docker/metadata-action@v5
+        id: meta
+        with:
+          images: ghcr.io/arknill/llm-relay
+          tags: |
+            type=semver,pattern={{version}}
+            type=semver,pattern={{major}}.{{minor}}
+            type=raw,value=latest
+            type=raw,value=${{ inputs.tag }},enable=${{ inputs.tag != '' }}
+
+      - uses: docker/build-push-action@v5
+        with:
+          context: .
+          platforms: linux/amd64,linux/arm64
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max

.gitignore

@@ -6,9 +6,30 @@ dist/
 build/
 .venv/
 .env
+.env.local
 *.db
 *.sqlite
 .pytest_cache/
 .ruff_cache/
 .mypy_cache/
 vendor/
+
+# Project instruction files (LOCAL-ONLY — NEVER COMMIT)
+CLAUDE.md
+PLAN.txt
+
+# GrowthBook intercept (LOCAL-ONLY — DO NOT COMMIT)
+src/llm_relay/proxy/growthbook_intercept.py
+
+# Investigation/evidence/audit files (local-only analysis)
+docs/investigation-*
+docs/audit-*
+docs/evidence-*/
+
+# Backup files
+docker-compose.yml.bak*
+
+# Credential files
+*.key
+*.pem
+oauth_creds.json

CHANGELOG.md

@@ -0,0 +1,105 @@
+# Changelog
+
+All notable changes to llm-relay are documented here.
+
+## [0.9.2] - 2026-04-30
+
+### Added
+- **I18n**: Lightweight en/ko locale support via `i18n.py` message catalog (27 keys)
+  - Backend: `t()` translation function wired into all zone classifiers (routes.py, display.py, mcp/server.py)
+  - Frontend: `msg()` helper with browser `navigator.language` auto-detection (dashboard, display, history)
+  - API: `/api/v1/i18n?lang=ko` endpoint for frontend message loading
+  - Env: `LLM_RELAY_LANG` (default `en`)
+- **MCP tool**: `session_history` — conversation replay with turn filtering (8 tools total)
+- **Auto-load .env**: Native uvicorn startup now reads `.env.public` → `.env.local` automatically (Docker `env_file` parity)
+
+### Fixed
+- Hardcoded Korean zone labels replaced with i18n `t()` calls (PR #12/#13 by @cnighswonger)
+- Composition pie chart missing after native restart (LLM_RELAY_HISTORY not loaded without .env autoload)
+
+## [0.9.1] - 2026-04-29
+
+### Added
+- **Zone A↔B alignment**: Recalibrated CC/Codex dual-zone thresholds
+- **Codex 400K unification**: Official context window constants
+- **Display enhancements**: Placeholders, font normalization, dead code removal
+- 3-provider status tiles (Anthropic/OpenAI/Gemini) on dashboard
+- `term_name` surfacing on `/turns` + dashboard render
+- GH App token injection for Codex via `cli_delegate`
+- API parameter validation (`_ParamError` → 400)
+
+## [0.7.1] - 2026-04-26
+
+### Added
+- **`llm-relay init`**: One-command setup — auto-detects CLIs, configures Claude Code proxy + MCP, initializes DB, starts server, runs health check. Options: `--dry-run`, `--skip-server`, `--port`
+- Updated README with `llm-relay init` as primary quick start
+- 18 new tests for init module
+
+## [0.7.0] - 2026-04-26
+
+### Added
+- **Quota monitoring**: `/api/v1/quota` endpoint surfaces Q5h/Q7d utilization and overage status from stored ratelimit headers
+- **Error rate tracking**: `/api/v1/errors` endpoint with 2xx/4xx/5xx/429 breakdown and error rate percentage
+- **Cache hit rate**: `/api/v1/cache` endpoint and per-session `cache_hit_rate` field in `/turns/{id}` and `/display` responses
+- **TTL tier detection**: `/api/v1/ttl` endpoint detects 1h/5m/mixed ephemeral cache tiers from SSE `message_start` events
+- **Dashboard API Health section**: 4-card grid showing quota, error rate, cache hit rate, and TTL tier with 30s auto-refresh
+- **Display page badges**: Cache hit rate and TTL tier badges on session cards
+- Ephemeral token extraction from SSE `cache_creation.ephemeral_1h/5m_input_tokens` (streaming and non-streaming)
+- DB migration: `ephemeral_1h_tokens` and `ephemeral_5m_tokens` columns on `requests` table
+- 28 new tests (quota 5, error 5, cache 6, TTL 7, log_request 2, API endpoints 3)
+
+## [0.6.0] - 2026-04-26
+
+### Added
+- **Dashboard Context Health section**: Real-time summary of SNR, duplicate reads, and tool_result% across all active sessions with per-session health cards
+- **Duplicate read details**: `duplicate_reads` dict in composition API returns `{filepath: count}` instead of count-only; top file basenames shown in Display, Dashboard, and TUI
+- **Duplicate read warning**: `DUPLICATE_READ_WARN_THRESHOLD` env var (default 5) triggers visual warning when any file exceeds threshold
+- **SNR recommendation**: `CC_SNR_WARNING` env var (default 0.3) adds session-split recommendation message when SNR drops below threshold
+- **Per-turn composition chart**: SVG stacked area chart on `/history/` session detail page showing how context composition evolves across turns, with compaction markers and hover tooltips
+- **Per-turn composition API**: `GET /api/v1/history/{session_id}/composition` endpoint with automatic sampling for large sessions (>50 turns)
+- `composition` field added to `/api/v1/turns` response (was only on `/api/v1/display`)
+- 13 new tests (composition, turns, history)
+
+### Fixed
+- History test mock leaks: `discover_external_cli_sessions` was not mocked, causing tests to find real session files on disk
+- Codex session file fallback test: path detection now works with `patch()` instead of monkey-patching
+- 2 ruff lint issues in `scripts/context_composition.py`
+
+## [0.5.0] - 2026-04-24
+
+### Added
+- **Context composition analysis**: Real-time 6-category breakdown (user/assistant/tool_use/tool_result/thinking/system) with SNR metrics and duplicate read tracking
+- **`llm-relay top`**: btop-style terminal monitor using Rich Live — works over SSH without a browser
+- **Connection type detection**: Automatically detects SSH, tmux, screen, mosh, tailscale, native (and combinations like ssh+tmux) from `/proc/PID/environ` + parent process tree
+- **SVG pie chart** on `/display/` page showing context composition with popover tooltips
+- **Connection type badges** on session cards
+- **i18n support** (contributed by [@cnighswonger](https://github.com/cnighswonger)): Browser locale detection with en/ko; server override via `LLM_RELAY_LANG`
+- `/api/v1/i18n` endpoint for locale-specific messages
+- `scripts/context_composition.py` CLI analysis tool
+- 51 new tests (composition 24, connection type 14, TUI 13)
+
+### Changed
+- Display page: prompt moved to top of session cards
+- Display page: border-radius 4px, border-left 2px
+- Development status upgraded from Alpha to Beta
+
+## [0.4.0] - 2026-04-23
+
+### Added
+- **Session history capture**: Proxy-level conversation recording for CC/Codex/Gemini
+- Delta/full storage with compaction detection
+- `/history/` web replay viewer with compaction timeline
+- 3 history API endpoints (`/api/v1/history`, `/{session_id}`, `/{session_id}/compactions`)
+- `session_history` MCP tool (8th tool)
+- Alive filter for `/api/v1/turns` endpoint with shared liveness helpers
+
+### Changed
+- MCP tools: 7 → 8
+
+## [0.3.0] - 2026-04-15
+
+### Changed
+- Clean public release: removed internal references and legacy naming
+- Unified branding to llm-relay
+- Sanitized design documentation for public release
+- Removed FeatureFlags detector (internal-only)

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 QuartzUnit
+Copyright (c) 2026 hmj / QuartzUnit
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.ko.md

@@ -0,0 +1,93 @@
+# llm-relay
+
+LLM 사용 통합 관리 — API 프록시 + 세션 진단 + 멀티 CLI 오케스트레이션
+
+[English](README.md) | [llms.txt](llms.txt)
+
+## 기능
+
+- **Proxy**: API 투명 프록시 — 캐시/토큰 모니터링 + 12전략 pruning
+- **Detect**: 7종 디텍터 (orphan, stuck, bloat, synthetic, cache, resume, microcompact)
+- **Recover**: 세션 복구 + doctor (7개 건강 검사)
+- **Guard**: 4-tier 임계값 데몬 — dual-zone(절대+비율) 분류
+- **Cost**: per-1% 비용 산출 + rate-limit 헤더 분석
+- **Orch**: 멀티 CLI 오케스트레이션 (Claude Code, Codex CLI, Gemini CLI)
+- **Display**: 멀티 CLI 세션 모니터 — provider 배지 + 프로세스 생존 감지
+- **I18n**: 다국어 지원 (영어/한국어) — 브라우저 자동 감지 + `LLM_RELAY_LANG` 환경변수
+- **MCP**: stdio 전송 8개 도구 (cli_delegate, cli_status, cli_probe, orch_delegate, orch_history, relay_stats, session_turns, session_history)
+
+## 설치
+
+```bash
+# CLI 전용 (진단, 복구, 오케스트레이션)
+pip install llm-relay
+
+# 프록시 + 웹 대시보드
+pip install llm-relay[proxy]
+
+# MCP 서버 (Python 3.10 이상)
+pip install llm-relay[mcp]
+
+# 전부
+pip install llm-relay[all]
+```
+
+## 빠른 시작
+
+### CLI 진단 (서버 불필요)
+
+```bash
+llm-relay scan              # 세션 건강 검사 (7종 디텍터)
+llm-relay doctor            # 설정 건강 검사 (7개 항목)
+llm-relay recover           # 세션 컨텍스트 추출 (재개용)
+```
+
+### 웹 대시보드
+
+```bash
+# 방법 1: 직접 실행
+pip install llm-relay[proxy]
+uvicorn llm_relay.proxy.proxy:app --host 0.0.0.0 --port 8083
+
+# 방법 2: Docker
+cp .env.public .env         # 필요에 따라 수정
+docker compose up -d
+```
+
+접속 주소:
+- `/dashboard/` — CLI 상태, 비용, 위임 히스토리, Turn Monitor (alive 세션만; `?include_dead=1` 로 우회)
+- `/display/` — 턴 카운터 + CC/Codex/Gemini 세션 카드 (alive 필터: CC=cc_pid+TTY fallback, Codex/Gemini=fd-open)
+
+### MCP 서버
+
+```bash
+llm-relay-mcp               # stdio 전송, 8개 도구
+```
+
+### Claude Code API 프록시
+
+```bash
+# Claude Code에서 설정
+export ANTHROPIC_BASE_URL=http://localhost:8080
+```
+
+## CLI 지원 현황
+
+| CLI | 상태 |
+|-----|------|
+| Claude Code | 전체 지원 |
+| OpenAI Codex | 전체 지원 |
+| Gemini CLI | Display 지원, oauth-personal 서버사이드 403 버그 ([#25425](https://github.com/google-gemini/gemini-cli/issues/25425)) |
+
+## 요구 사항
+
+- Python >= 3.9
+- MCP 도구는 Python >= 3.10
+
+## 라이선스
+
+MIT
+
+## 생태계
+
+[QuartzUnit](https://github.com/QuartzUnit) 오픈소스 생태계의 일부입니다.