APEX — Adaptive Prompt Engineering eXtension

Structured reasoning, rich context engineering, and prompt amplification — complementing Copilot with zero external API keys.

APEX is a VS Code extension that enhances GitHub Copilot Chat by adding intelligent model routing, workspace-aware context injection, and task-specific prompt engineering. It works entirely within Copilot's ecosystem — no external API keys, no additional billing.

Features

Slash Commands

Command	Strategy	Description
`@apex /debug`	ReAct	Observe → Think → Act → Verify debugging
`@apex /refactor`	Chain-of-Thought	Step-by-step refactoring with dependency analysis
`@apex /design`	Tree-of-Thought	Multiple approaches → compare → recommend
`@apex /test`	Few-Shot	Test generation matching your project's patterns
`@apex /explain`	Layered	High-level → implementation → gotchas
`@apex /plan`	Decomposition	Break complex tasks into sub-steps
`@apex /verify`	Cross-Model Critique	Review via a different-vendor model
`@apex /init`	Stack Detection	Generate `.github/copilot-instructions.md`

Core Capabilities

Intent Engine — Zero-PRU-cost heuristic classifier that detects task type and complexity from freeform queries
Context Maximizer — Collects active file, diagnostics, symbols, git diff, imports, and user #file: references
Prompt Amplification — Task-specific reasoning templates (CoT, ToT, ReAct, Few-Shot) injected via @vscode/prompt-tsx
Model Router — Budget-aware selection: respects your dropdown choice by default, picks optimal models when opted in
PRU Governor — Local PRU estimate tracking with 3 budget modes (Eco / Balanced / Performance) and auto-downgrade
Cross-Model Critique — Sends your response to a different-vendor model for review (when complexity warrants it)
Agent Mode Tools — 4 tools available to all Copilot participants: workspace context, related files, conventions, test patterns

Budget Modes

Mode	Icon	Behavior
Eco	`$(circle-filled)`	Free models only, prompt engineering focus
Balanced	`$(warning)`	Tier 1 reviewers for complex tasks
Performance	`$(flame)`	Full model routing, cross-model critique

Requirements

VS Code ^1.100.0
GitHub Copilot Chat extension installed and active
A GitHub Copilot subscription (Free, Pro, or Student plan)

Installation

# Clone and build
git clone <repo-url>
cd apex
npm install
npm run build

# For development
npm run build:dev

# Package for distribution
npm run package

Extension Settings

Setting	Type	Default	Description
`apex.budgetMode`	`eco\\|balanced\\|performance`	`eco`	How aggressively APEX uses premium models
`apex.monthlyBudget`	`number` (0-300)	`100`	Max estimated PRU per month
`apex.modelOverride`	`never\\|always`	`never`	`never`: use your model choice. `always`: APEX picks optimal models
`apex.enableCritique`	`boolean`	`true`	Enable cross-model critique pipeline
`apex.critiqueThreshold`	`number` (0-100)	`30`	Min complexity score to trigger critique
`apex.enableDecomposition`	`boolean`	`true`	Enable multi-step task decomposition
`apex.decompositionThreshold`	`number` (0-100)	`60`	Min complexity to trigger decomposition
`apex.pipelineDelayMs`	`number` (0-5000)	`500`	Delay between pipeline steps (rate limit protection)
`apex.modelCacheTTLHours`	`number` (1-168)	`24`	Model discovery cache TTL
`apex.enableTelemetry`	`boolean`	`true`	Anonymized usage telemetry

Commands

Command	Description
`APEX: Show Budget Details`	View current PRU usage and budget status
`APEX: Reset PRU Budget Estimate`	Reset the monthly PRU estimate to 0
`APEX: Verify PRU Multipliers`	Show discovered models and their PRU multipliers

Agent Mode Tools

These tools are available to all Copilot participants in agent mode:

Tool	Reference	Description
`apex_collectWorkspaceContext`	`#apexContext`	Diagnostics, symbols, git changes
`apex_findRelatedFiles`	`#apexRelated`	Import graph and naming pattern analysis
`apex_detectConventions`	`#apexConventions`	Naming, error handling, test patterns
`apex_extractTestPatterns`	`#apexTests`	Framework, assertions, mocks

Architecture

User → @apex /debug TypeError
         │
         ▼
  apexParticipant.ts
    ├── ModelRegistry.ensureDiscovered()   (lazy, first call only)
    ├── IntentEngine.classify()            (zero-PRU heuristics)
    ├── ModelRouter.select()               (budget-aware routing)
    ├── resolvePromptReferences()          (#file: attachments)
    ├── ContextMaximizer.collect()         (workspace context)
    ├── renderPrompt(BasePrompt)           (prompt-tsx with priorities)
    ├── sendWithFallback()                 (retry + fallback chain)
    ├── performCritique()                  (optional cross-model review)
    └── PruGovernor.record()               (budget tracking)

Key Design Decisions

No selectChatModels() on activate — consent dialog requires user-initiated action
request.model respected by default — APEX only overrides when user opts in via modelOverride: "always"
PRU tracking is a local estimate — the extension API cannot query actual remaining balance
@vscode/prompt-tsx alpha — pinned version with renderPrompt() fallback if it breaks
All model family strings are unverified — Phase 1 go/no-go gate must confirm via real selectChatModels() call

Development

# Type check
npm run lint

# Build (production)
npm run build

# Build (development, with source maps)
npm run build:dev

# Watch mode
npm run watch

# Package VSIX
npm run package

Testing

# Run tests (requires VS Code test infrastructure)
# Tests use mocked vscode APIs — no real Copilot calls
npx tsc -p tsconfig.test.json --noEmit

Project Structure

src/
├── extension.ts              # Entry point
├── core/                     # Intent, routing, context, amplification, budget
├── prompts/                  # prompt-tsx components (BasePrompt, History, Context, Critique)
├── participant/              # @apex handler, follow-ups, slash commands
├── tools/                    # Agent mode tools (4 tools)
├── instructions/             # /init: stack detection, convention inference
├── ui/                       # Status bar, notifications
└── utils/                    # Logger, config, retry, throttle, git, LSP
templates/                    # 10 prompt templates + copilot-instructions.hbs
test/                         # Mocha test suite

Known Limitations

PRU estimates are approximations — actual PRU consumption may differ
Model family strings are unverified until Phase 1 runtime testing on a real account
Cross-model critique and task decomposition are hypotheses pending A/B testing
@vscode/prompt-tsx is pre-1.0 alpha — API surface may change

License

MIT

APEX AI

Nikhil Pise