Compressed PDA Programs

Build Solana programs with compressed accounts via manual CPI to the Light System Program. Compressed accounts cost ~100x less than regular accounts (no rent-exemption required).

When to use compressed PDAs

• Per-user state (profiles, game state, credentials)
• DePIN device registrations
• Nullifier-based double-spend prevention
• Infrequently accessed accounts

Choosing approach

If you just want rent-free Anchor accounts, use the light-sdk skill instead. This skill is for programs that require manual CPI to the Light System Program (custom compressed state, ZK proofs, address derivation).

Client-program interaction flow

 ├─ Client
 │  ├─ Get ValidityProof from RPC.
 │  ├─ pack accounts with PackedAccounts into PackedAddressTreeInfo and PackedStateTreeInfo.
 │  ├─ pack CompressedAccountMeta.
 │  ├─ Build Instruction from PackedAccounts and CompressedAccountMetas.
 │  └─ Send transaction.
 │
 └─ Custom Program
    ├─ CpiAccounts parse accounts consistent with PackedAccounts.
    ├─ LightAccount instantiates from CompressedAccountMeta.
    │
    └─ Light System Program CPI
       ├─ Verify ValidityProof.
       ├─ Update State Merkle tree.
       ├─ Update Address Merkle tree.
       └─ Complete atomic state transition.

Domain references

Reference repos

/home/tilo/Workspace/program-examples/
├── create-and-update/     # Basic create + update pattern
├── close-account/         # Close, burn, reinit patterns
├── token-escrow/          # Token + compressed account patterns
├── airdrop-implementations/distributor/  # Merkle + compressed
└── zk/                    # ZK proof patterns

/home/tilo/Workspace/examples-zk-compression/
/home/tilo/Workspace/example-token-distribution/
/home/tilo/Workspace/example-nodejs-client/

Workflow

1. Clarify intent
   • Recommend plan mode, if it's not activated
   • Use AskUserQuestion to resolve blind spots
   • All questions must be resolved before execution
2. Identify references
   • Match task to domain references and reference repos
   • Locate relevant documentation and examples
3. Write plan file (YAML task format)
   • Use AskUserQuestion for anything unclear — never guess or assume
   • Identify blockers: permissions, dependencies, unknowns
   • Plan must be complete before execution begins
4. Execute
   • Use Task tool with subagents for parallel research
   • Subagents load skills via Skill tool
   • Track progress with TodoWrite
5. When stuck: spawn subagent with Read, Glob, Grep, DeepWiki MCP access and load skills/ask-mcp

Build and test

Required commands

Anchor programs:

anchor build
anchor test

Native programs:

cargo build-sbf
cargo test-sbf

Forbidden shortcuts

• Do NOT use cargo build (must use cargo build-sbf)
• Do NOT use cargo test (must use cargo test-sbf)
• Do NOT skip SBF compilation
• Tests MUST run against real BPF bytecode

Failure recovery

On failure, spawn debugger agent with error context.

Loop rules:

1. Each debugger gets fresh context + previous debug reports
2. Each attempt tries something DIFFERENT
3. NEVER GIVE UP - keep spawning until fixed
4. Max 5 attempts per error

Do NOT proceed until all tests pass.

SDK references

DeepWiki fallback

If no matching pattern in reference repos:

mcp__deepwiki__ask_question("Lightprotocol/light-protocol", "How to {operation}?")