gsd-trae 1.0.0 → 1.0.2

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/design.md

@@ -1,111 +0,0 @@
-# Design: VibeCodingBench Architecture
-
-## Context
-Building a comprehensive benchmark for coding agents (Claude Code, Gemini, Codex, DeepSeek, etc.) that measures real-world developer task performance. Must support both local execution and hosted leaderboard.
-
-## Goals
-- Reproducible evaluation across different agents
-- Fair comparison with isolated Docker execution
-- Multi-dimensional scoring (not just pass/fail)
-- Easy task contribution workflow
-- Support polyglot: TypeScript, Python, Go, Rust, Java
-
-## Non-Goals
-- Real-time collaboration features
-- IDE integrations (agents run headless)
-- Training data generation
-
-## Decisions
-
-### 1. Monorepo Structure
-```
-vibecodingbench/
-├── packages/
-│   ├── cli/              # Task runner CLI
-│   ├── evaluator/        # Scoring engine
-│   └── leaderboard/      # Web service
-├── tasks/
-│   ├── saas-core/        # 30% weight
-│   ├── glue-code/        # 20% weight
-│   ├── ai-integration/   # 20% weight
-│   ├── frontend/         # 15% weight
-│   └── api-integrations/ # 15% weight
-├── templates/            # Starter codebases
-│   ├── nextjs-supabase/
-│   ├── fastapi-postgres/
-│   ├── go-fiber/
-│   └── rust-axum/
-└── docker/               # Base images
-```
-**Rationale:** Single repo simplifies versioning, CI, and contributions.
-
-### 2. Task Definition Format
-Each task is a directory:
-```
-tasks/saas-core/auth/supabase-oauth/
-├── task.yaml           # Metadata, prompt, constraints
-├── docker-compose.yaml # Services (DB, mock APIs)
-├── template/           # Starter code (optional)
-├── tests/              # Evaluation tests
-│   ├── functional/     # Must pass
-│   ├── security/       # OWASP checks
-│   └── visual/         # Screenshot diff (frontend)
-└── golden/             # Reference implementation
-```
-**Rationale:** Self-contained, versionable, easy to add.
-
-### 3. Execution Model
-```
-┌─────────────┐     ┌──────────────┐     ┌─────────────┐
-│   CLI       │────▶│  Task Env    │────▶│  Evaluator  │
-│  (host)     │     │  (Docker)    │     │  (Docker)   │
-└─────────────┘     └──────────────┘     └─────────────┘
-      │                    │                    │
-      │   mount workspace  │   run agent        │   run tests
-      │   inject prompt    │   capture output   │   compute scores
-      └────────────────────┴────────────────────┘
-```
-- Agent runs inside container with network access (for package installs)
-- Evaluation runs in separate container (no agent access)
-- Time/token limits enforced by CLI
-
-### 4. Scoring Dimensions
-| Dimension | Weight | Method |
-|-----------|--------|--------|
-| Functional | 40% | Test pass rate |
-| Code Quality | 20% | ESLint/Ruff + complexity metrics |
-| Security | 20% | Semgrep OWASP rules |
-| Efficiency | 20% | Tokens used + wall time |
-
-### 5. Agent Interface
-Agents connect via stdio or HTTP:
-```yaml
-# task.yaml
-agent_interface:
-  type: stdio  # or http
-  prompt_file: PROMPT.md
-  workspace: /workspace
-  timeout: 300s
-  token_limit: 100000
-```
-
-## Alternatives Considered
-
-### Task Registry (rejected)
-- Pros: Smaller local footprint
-- Cons: More infrastructure, harder offline use
-- Decision: Start monorepo, can extract registry later
-
-### VM-per-task (rejected)
-- Pros: Better isolation
-- Cons: 10x cost, slower iteration
-- Decision: Docker sufficient, VMs for hosted tier only
-
-## Risks & Mitigations
-
-| Risk | Mitigation |
-|------|------------|
-| Task contamination in training data | Version tasks, rotate variants |
-| Agent gaming metrics | Multiple equivalent tasks per category |
-| Unfair time comparisons | Normalize by model speed tier |
-| Docker escape | Rootless containers, seccomp profiles |

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/proposal.md

@@ -1,15 +0,0 @@
-# Change: Initialize VibeCodingBench
-
-## Why
-Existing coding benchmarks (HumanEval, SWE-bench) focus on algorithmic puzzles or isolated bug fixes. Real developers spend 40% of time on SaaS boilerplate, integrations, and glue code. We need a benchmark that measures what coding agents actually do in production.
-
-## What Changes
-- Create monorepo structure with task runner CLI
-- Define task specification format (YAML + Docker)
-- Implement multi-dimensional evaluation (functional, quality, security, efficiency)
-- Build leaderboard service for hosted evaluation
-- Add 200+ tasks across 5 categories: SaaS, Glue Code, AI Integration, Frontend, API
-
-## Impact
-- Affected specs: task-runner, task-definition, evaluation, leaderboard (all new)
-- Affected code: Greenfield project

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/specs/evaluation/spec.md

@@ -1,105 +0,0 @@
-## ADDED Requirements
-
-### Requirement: Multi-Dimensional Scoring
-The system SHALL compute scores across five dimensions with configurable weights.
-
-#### Scenario: Default weights
-- **WHEN** no custom weights specified
-- **THEN** system uses: Functional 40%, Visual 20%, Quality 20%, Cost 10%, Speed 10%
-
-#### Scenario: Custom weights
-- **WHEN** user specifies `--weights func=50,visual=0,quality=30,cost=10,speed=10`
-- **THEN** system applies custom weight distribution
-
-### Requirement: Functional Correctness (Pass@k)
-The system SHALL measure functional correctness via execution-based testing.
-
-#### Scenario: Pass@1
-- **WHEN** test suite runs once and passes
-- **THEN** functional score = 100%
-
-#### Scenario: Pass@n with retries
-- **WHEN** task allows n attempts and any attempt passes
-- **THEN** functional score = 100% but efficiency penalty applied
-
-#### Scenario: Fail-to-Pass validation
-- **WHEN** task is bug-fix type
-- **THEN** system verifies agent's test fails before fix and passes after
-
-### Requirement: Visual Fidelity
-The system SHALL measure UI accuracy via screenshot comparison.
-
-#### Scenario: Pixel diff scoring
-- **WHEN** task has `reference.png` in golden/
-- **THEN** system captures screenshot and computes pixel match percentage
-
-#### Scenario: Responsive breakpoints
-- **WHEN** task specifies `breakpoints: [375, 768, 1440]`
-- **THEN** system tests at each width and averages scores
-
-#### Scenario: Tolerance threshold
-- **WHEN** pixel mismatch < 5%
-- **THEN** visual score = 100% (allows font rendering variance)
-
-### Requirement: Code Quality
-The system SHALL measure code hygiene via static analysis.
-
-#### Scenario: Linter errors
-- **WHEN** generated code has linter errors
-- **THEN** quality score reduced by error count (max -50 points)
-
-#### Scenario: Cyclomatic complexity
-- **WHEN** average complexity > 10
-- **THEN** quality score reduced proportionally
-
-#### Scenario: Security scan
-- **WHEN** Semgrep finds Critical/High vulnerabilities
-- **THEN** task auto-fails regardless of other scores
-
-### Requirement: Hallucination Detection
-The system SHALL detect fabricated dependencies.
-
-#### Scenario: Import validation
-- **WHEN** agent imports package not in npm/PyPI/Go modules
-- **THEN** hallucination flag raised, quality score -20
-
-### Requirement: Cost Efficiency
-The system SHALL track token usage and compute costs.
-
-#### Scenario: Token tracking
-- **WHEN** task completes
-- **THEN** system records input_tokens, output_tokens, total_cost
-
-#### Scenario: Cost per solved task (CPST)
-- **WHEN** computing leaderboard
-- **THEN** CPST = total_cost / passed_tasks
-
-#### Scenario: Context pollution rate
-- **WHEN** agent reads files
-- **THEN** pollution_rate = (files_read - files_edited) / files_read
-
-### Requirement: Speed Metrics
-The system SHALL track execution time and reasoning efficiency.
-
-#### Scenario: Wall-clock time
-- **WHEN** task completes
-- **THEN** system records start_time, end_time, duration_seconds
-
-#### Scenario: Step efficiency
-- **WHEN** agent completes task
-- **THEN** system counts LLM round-trips (fewer = better)
-
-#### Scenario: Self-correction rate
-- **WHEN** agent encounters error and retries
-- **THEN** system tracks retry_count (target < 2)
-
-### Requirement: Final Score Calculation
-The system SHALL compute weighted final score with penalties.
-
-#### Scenario: Score formula
-- **WHEN** all dimensions computed
-- **THEN** final_score = (func * w1) + (visual * w2) + (quality * w3) - (cost_penalty) - (speed_penalty)
-
-#### Scenario: Leaderboard ranking
-- **WHEN** displaying results
-- **THEN** rank by final_score descending, show all dimensions in spider chart

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/specs/leaderboard/spec.md

@@ -1,68 +0,0 @@
-## ADDED Requirements
-
-### Requirement: Submission API
-The system SHALL accept evaluation submissions via REST API.
-
-#### Scenario: Submit run results
-- **WHEN** POST /api/submissions with run results JSON
-- **THEN** system validates, stores, and queues for leaderboard update
-
-#### Scenario: Agent identification
-- **WHEN** submission includes `agent_id` and `model_version`
-- **THEN** system groups results by agent for comparison
-
-### Requirement: Leaderboard Display
-The system SHALL display ranked agents with multi-dimensional scores.
-
-#### Scenario: Overall leaderboard
-- **WHEN** GET /api/leaderboard
-- **THEN** system returns agents ranked by final_score with all dimension breakdowns
-
-#### Scenario: Category leaderboard
-- **WHEN** GET /api/leaderboard?category=saas-core
-- **THEN** system returns agents ranked by performance in that category only
-
-#### Scenario: Spider chart data
-- **WHEN** GET /api/leaderboard/:agent_id/chart
-- **THEN** system returns 5-axis radar chart data (func, visual, quality, cost, speed)
-
-### Requirement: Historical Tracking
-The system SHALL track agent performance over time.
-
-#### Scenario: Version comparison
-- **WHEN** same agent submits new model version
-- **THEN** system shows delta vs previous version
-
-#### Scenario: Trend graphs
-- **WHEN** viewing agent detail page
-- **THEN** system displays score trends over last 30 days
-
-### Requirement: Live Demo Dashboard
-The system SHALL provide real-time task execution viewing.
-
-#### Scenario: Active runs
-- **WHEN** tasks are running
-- **THEN** dashboard shows live terminal streams and browser recordings
-
-#### Scenario: Replay recordings
-- **WHEN** user selects completed run
-- **THEN** system plays back asciinema recording synced with browser video
-
-#### Scenario: Side-by-side comparison
-- **WHEN** user selects 2+ agents for same task
-- **THEN** system shows parallel playback of each agent's execution
-
-### Requirement: Fairness Controls
-The system SHALL enforce fair comparison conditions.
-
-#### Scenario: Docker isolation
-- **WHEN** submitting results
-- **THEN** system verifies run was in fresh Docker container (via attestation)
-
-#### Scenario: Held-out validation
-- **WHEN** task is marked `held_out: true`
-- **THEN** system only accepts submissions from last 14 days (prevents training contamination)
-
-#### Scenario: Standardized scaffolding
-- **WHEN** displaying leaderboard
-- **THEN** system shows which agent tooling was used (raw API vs Claude Code CLI vs Codex CLI)

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/specs/task-definition/spec.md

@@ -1,45 +0,0 @@
-## ADDED Requirements
-
-### Requirement: Task Schema
-The system SHALL validate task definitions against a JSON Schema.
-
-#### Scenario: Valid task.yaml
-- **WHEN** task.yaml contains all required fields (id, name, category, prompt, timeout)
-- **THEN** system loads task without errors
-
-#### Scenario: Invalid task.yaml
-- **WHEN** task.yaml is missing required fields
-- **THEN** system reports validation errors with line numbers
-
-### Requirement: Task Structure
-Each task SHALL be a self-contained directory with standardized layout.
-
-#### Scenario: Minimal task
-- **WHEN** task directory contains `task.yaml` and `tests/`
-- **THEN** system can execute and evaluate the task
-
-#### Scenario: Full task with template
-- **WHEN** task directory contains `task.yaml`, `template/`, `tests/`, `golden/`
-- **THEN** system uses template as starter code and golden for reference comparison
-
-### Requirement: Task Metadata
-Task definitions SHALL include metadata for filtering and scoring.
-
-#### Scenario: Category and weight
-- **WHEN** task.yaml specifies `category: saas-core` and `weight: 1.5`
-- **THEN** system applies weight multiplier to final score
-
-#### Scenario: Difficulty level
-- **WHEN** task.yaml specifies `difficulty: hard`
-- **THEN** system adjusts timeout and token limits accordingly
-
-### Requirement: Prompt Specification
-Tasks SHALL define agent prompts with clear success criteria.
-
-#### Scenario: Prompt file
-- **WHEN** task.yaml specifies `prompt_file: PROMPT.md`
-- **THEN** system reads prompt from that file with variable substitution
-
-#### Scenario: Inline prompt
-- **WHEN** task.yaml contains `prompt:` field directly
-- **THEN** system uses inline prompt text

package/refs/vbenchmark/openspec/changes/init-vibecodingbench/specs/task-runner/spec.md

@@ -1,49 +0,0 @@
-## ADDED Requirements
-
-### Requirement: Task Discovery
-The system SHALL discover tasks from the `tasks/` directory by scanning for `task.yaml` files.
-
-#### Scenario: List all tasks
-- **WHEN** user runs `vibecodingbench list`
-- **THEN** system displays all tasks grouped by category with metadata
-
-#### Scenario: Filter by category
-- **WHEN** user runs `vibecodingbench list --category saas-core`
-- **THEN** system displays only tasks in that category
-
-### Requirement: Task Execution
-The system SHALL execute tasks in isolated Docker containers with configurable timeouts.
-
-#### Scenario: Run single task
-- **WHEN** user runs `vibecodingbench run <task-id> --agent claude-code`
-- **THEN** system spawns Docker container, injects prompt, captures agent output
-
-#### Scenario: Timeout enforcement
-- **WHEN** agent exceeds task timeout (default 300s)
-- **THEN** system kills container and records timeout failure
-
-#### Scenario: Token limit enforcement
-- **WHEN** agent exceeds token limit (default 100k)
-- **THEN** system stops agent and records token limit failure
-
-### Requirement: Agent Interface
-The system SHALL support multiple agent connection methods.
-
-#### Scenario: Stdio agent
-- **WHEN** task.yaml specifies `agent_interface.type: stdio`
-- **THEN** system communicates via stdin/stdout pipes
-
-#### Scenario: HTTP agent
-- **WHEN** task.yaml specifies `agent_interface.type: http`
-- **THEN** system communicates via REST API on localhost:8080
-
-### Requirement: Live Demo Mode
-The system SHALL support live streaming of task execution for demos.
-
-#### Scenario: Stream execution
-- **WHEN** user runs `vibecodingbench run <task-id> --live`
-- **THEN** system streams agent actions, terminal output, and browser (if applicable) to web UI
-
-#### Scenario: Record session
-- **WHEN** user runs `vibecodingbench run <task-id> --record`
-- **THEN** system saves asciinema recording and browser video to `results/<run-id>/`