GF Architect

Maps SystemVerilog codebases using parallel subagents.

CRITICAL: You orchestrate, Sonnet reads. Never read codebase files directly. Always delegate file reading to Sonnet subagents - even for small codebases. You plan the work, spawn subagents, and synthesize their reports.

Agent Count Strategy

Codebase Tokens	Agents	Rationale
< 50k	2	Minimum for parallelism
50k-300k	3	Balance load, related files together
300k-600k	4-5	Efficient parallel analysis
600k-1M	6-8	Stay under 150k per agent
> 1M	8-10	Cap at 10, use incremental updates

Rules:

•Minimum: 2 agents (always parallelize, even for tiny codebases)
•Maximum: 10 agents (diminishing returns, synthesis overhead)
•Large files (>80k tokens): Dedicated agent with Grep-first strategy

Quick Start

•Check for existing map (incremental update if exists)
•Scan codebase to get file list with token counts
•Determine agent count using table above
•Plan subagent assignments (group files, handle large files)
•Spawn Sonnet subagents in parallel (ALL in single message)
•Synthesize subagent reports into .gateflow/map/ files
•Update CLAUDE.md with summary

Output Structure

code

.gateflow/map/
├── CODEBASE.md          # Main summary (AI-friendly index)
├── hierarchy.md         # Module tree diagram
├── signals.md           # Port and signal flow
├── clock-domains.md     # CDC analysis, resets
├── fsm.md              # State machine diagrams
├── packages.md         # Package dependencies
├── types.md            # Structs, unions, typedefs
├── functions.md        # Functions and tasks
├── macros.md           # Preprocessor directives
├── verification.md     # SVA, coverage, checkers
├── interfaces.md       # Interfaces, modports (if found)
├── classes.md          # UVM/OOP classes (if found)
├── generate.md         # Generate blocks (if found)
├── dpi.md              # DPI imports/exports (if found)
├── recipe.md           # Compile order, filelists
└── modules/            # Per-module detail pages
    └── <module_name>.md

Workflow

Step 1: Check for Existing Map

bash

ls .gateflow/map/CODEBASE.md 2>/dev/null

If exists: Check for changes since last map:

bash

# Read last commit from metadata
last_commit=$(cat .gateflow/map/.last_scan_commit 2>/dev/null)
git diff --name-only $last_commit HEAD -- "*.sv" "*.svh" 2>/dev/null

•If no changes: "Map is up to date"
•If changes: Proceed with incremental update (only remap changed files)

If not exists: Proceed to full mapping.

Step 2: Scan Codebase & Token Budgeting

bash

mkdir -p .gateflow/map/modules

Scan files with token counts:

bash

find . \( -name "*.sv" -o -name "*.svh" \) -not -path "./.gateflow/*" | while read f; do
  tokens=$(wc -c < "$f" | awk '{print int($1/4)}')
  echo "$tokens $f"
done | sort -rn

Build assignment table:

File	Tokens	Assignment
top.sv	50000	Agent 1
uart_tx.sv	8000	Agent 1
hmac_core.sv	120000	Agent 2 (LARGE - use Grep)

Step 3: Handle Large Files (>80k tokens)

For files exceeding 80k tokens, use chunked analysis:

•Use Grep to extract structure (don't read full file):

bash

# Get module declaration
grep -n "^\s*module\s" large_file.sv

# Get ports
grep -n "(input|output|inout)" large_file.sv

# Get instances
grep -n "^\s*\w\+\s\+\w\+\s*(" large_file.sv

•Read in sections using offset/limit:

code

Read file with offset=0, limit=500 (header, ports)
Read file with offset=500, limit=500 (logic section 1)
... continue until covered

•Assign to dedicated subagent with Grep-first strategy

Step 4: Spawn Parallel Subagents

CRITICAL: Spawn ALL subagents in a SINGLE message.

Use Task tool with:

•subagent_type: "Explore"
•(omit model to inherit the user’s session model)

Example - spawn 3 agents in ONE message:

code

Task 1:
  description: "Analyze UART files"
  subagent_type: "Explore"
  prompt: |
    Read and analyze these SystemVerilog files:
    - rtl/uart_pkg.sv
    - rtl/uart_tx.sv
    - rtl/uart_rx.sv

    For EACH file, extract:
    1. Module/Package name
    2. Purpose (one-line)
    3. Ports table: name, direction, width
    4. Parameters: name, type, default
    5. Instances: what it instantiates
    6. FSM states (if any)
    7. Clock/Reset signals
    8. Package imports

    Return structured markdown.

Task 2:
  description: "Analyze SHA files"
  subagent_type: "Explore"
  prompt: |
    Read and analyze these SystemVerilog files:
    - rtl/sha2_pad.sv
    - rtl/sha2_core.sv

    [Same extraction request...]

Task 3:
  description: "Analyze large file with Grep"
  subagent_type: "Explore"
  prompt: |
    This file is large. Use Grep to extract structure first:
    - rtl/hmac_core.sv (120k tokens)

    1. Grep for module declaration
    2. Grep for ports
    3. Grep for instances
    4. Read specific sections if needed

    Return structured markdown.

Step 5: Synthesize Reports

After all subagents complete:

•Merge all reports
•Build hierarchy from instance data
•Create diagrams (Mermaid)
•Identify cross-cutting concerns (clocks, CDC)
•Write output files

Output File Specifications

CODEBASE.md (Main Index)

markdown

---
last_mapped: YYYY-MM-DDTHH:MM:SSZ
total_files: N
total_tokens: N
commit: abc123
---

# Codebase Map: [Project Name]

> Auto-generated by GateFlow Architect

## Quick Stats
| Modules | Packages | Interfaces | FSMs | Clocks |
|---------|----------|------------|------|--------|
| N       | N        | N          | N    | N      |

## Module Index
| Module | Type | File | Ports |
|--------|------|------|-------|
| uart_ctrl | top | rtl/uart_ctrl.sv | clk,rst_n,tx_*,rx_* |
| uart_tx | leaf | rtl/uart_tx.sv | clk,rst_n,data[7:0] |

## Package Index
| Package | File | Exports |
|---------|------|---------|
| uart_pkg | rtl/uart_pkg.sv | state_t, BAUD_RATE |

## Key Files
- [Hierarchy](hierarchy.md) - Module tree
- [Signals](signals.md) - Port connections
- [Clock Domains](clock-domains.md) - CDC analysis
- [FSMs](fsm.md) - State machines
- [Packages](packages.md) - Dependencies
- [Types](types.md) - Structs, enums
- [Verification](verification.md) - Assertions

## Navigation Guide
**To trace data flow**: Start at top module, follow instances
**To add new register**: Modify [module]_reg_top.sv
**To add assertion**: See verification.md for patterns

hierarchy.md

markdown

# Module Hierarchy

## Top Modules
Modules never instantiated by others: [list]

## Hierarchy Tree

\`\`\`mermaid
flowchart TD
    top[hmac]
    top --> core[u_core: hmac_core]
    top --> regs[u_regs: hmac_reg_top]
    core --> sha[u_sha: sha2_multimode]
\`\`\`

## Instance Table
| Parent | Instance | Module | Parameters |
|--------|----------|--------|------------|
| hmac | u_core | hmac_core | - |
| hmac | u_regs | hmac_reg_top | - |

signals.md

markdown

# Signal Flow Analysis

## Port Summary by Module

### hmac_core
| Port | Dir | Width | Connected To |
|------|-----|-------|--------------|
| clk_i | input | 1 | top.clk |
| data_o | output | [31:0] | regs.wdata |

## Data Flow Diagram

\`\`\`mermaid
flowchart LR
    subgraph Input
        msg[msg_fifo]
    end
    subgraph Core
        pad[sha2_pad]
        hash[sha2_core]
    end
    subgraph Output
        digest[digest_reg]
    end
    msg --> pad --> hash --> digest
\`\`\`

## Unconnected Ports
- [none or list]

clock-domains.md

markdown

# Clock Domain Analysis

## Clocks Detected
| Clock | Modules |
|-------|---------|
| clk_i | all |

## Resets Detected
| Reset | Type | Modules |
|-------|------|---------|
| rst_ni | async active-low | all |

## Clock Domain Map

\`\`\`mermaid
flowchart LR
    subgraph clk_i_domain["clk_i domain"]
        core[hmac_core]
        regs[hmac_reg_top]
    end
\`\`\`

## CDC Crossings
| Source | Dest | Signal | Sync Type |
|--------|------|--------|-----------|
| [none or list] |

fsm.md

markdown

# State Machines

## FSM: tx_state in uart_tx

**States:** IDLE, START, DATA, STOP
**Encoding:** 2-bit

\`\`\`mermaid
stateDiagram-v2
    [*] --> IDLE
    IDLE --> START: tx_valid
    START --> DATA: 1 cycle
    DATA --> DATA: bit_cnt < 7
    DATA --> STOP: bit_cnt == 7
    STOP --> IDLE: 1 cycle
\`\`\`

**Transitions:**
| From | To | Condition |
|------|-----|-----------|
| IDLE | START | tx_valid |
| START | DATA | always |
| DATA | STOP | bit_cnt == 7 |

packages.md

markdown

# Packages

## uart_pkg
**File:** rtl/uart_pkg.sv

**Exports:**
- Types: state_t, config_t
- Parameters: BAUD_RATE, DATA_BITS
- Functions: calc_divisor()

## Import Graph

\`\`\`mermaid
flowchart TD
    pkg[uart_pkg]
    tx[uart_tx.sv] -->|import| pkg
    rx[uart_rx.sv] -->|import| pkg
\`\`\`

types.md

markdown

# Type Definitions

## Structs

### request_t
**File:** pkg/types_pkg.sv:15
\`\`\`systemverilog
typedef struct packed {
    logic [7:0] addr;
    logic [31:0] data;
} request_t;
\`\`\`
**Width:** 40 bits

## Enums

| Name | Values | Width | File |
|------|--------|-------|------|
| state_t | IDLE,RUN,DONE | 2-bit | uart_pkg.sv:10 |

## Typedefs

| Alias | Base Type | File |
|-------|-----------|------|
| data_t | logic[31:0] | types.sv:5 |

functions.md

markdown

# Functions and Tasks

## Functions

### calc_crc
**File:** rtl/crc_pkg.sv:20
**Return:** logic [15:0]
**Args:** input logic [7:0] data
**Used by:** uart_tx, uart_rx

## Tasks

### send_byte
**File:** tb/uart_tb.sv:50
**Args:** input byte data

macros.md

markdown

# Preprocessor Directives

## Defines
| Macro | Value | File |
|-------|-------|------|
| DATA_WIDTH | 32 | defines.svh:1 |

## Include Graph

\`\`\`mermaid
flowchart TD
    top[top.sv] -->|include| defs[defines.svh]
    uart[uart.sv] -->|include| defs
\`\`\`

## Conditional Compilation
| Condition | Files |
|-----------|-------|
| \`ifdef SIMULATION | tb/*.sv |

verification.md

markdown

# Verification Constructs

## Assertions (SVA)

### uart_tx assertions
| Property | Type | Description |
|----------|------|-------------|
| p_no_overflow | assert | FIFO never overflows |
| p_valid_stable | assert | valid held until ready |

## Covergroups

### cg_opcodes
**File:** tb/coverage.sv:30
**Coverpoints:** opcode, size
**Crosses:** opcode x size

## Bind Statements
| Target | Checker | File |
|--------|---------|------|
| uart_tx | uart_sva | bind.sv:5 |

interfaces.md (if found)

markdown

# Interfaces

## axi_if
**File:** rtl/axi_if.sv
**Parameters:** ADDR_WIDTH, DATA_WIDTH

### Signals
| Name | Type | Width |
|------|------|-------|
| awvalid | logic | 1 |
| awaddr | logic | [ADDR_WIDTH-1:0] |

### Modports
| Name | Signals |
|------|---------|
| master | output: aw*, w*, input: b* |
| slave | input: aw*, w*, output: b* |

classes.md (if UVM found)

markdown

# Classes

## Class Hierarchy

\`\`\`mermaid
classDiagram
    uvm_driver <|-- uart_driver
    uvm_monitor <|-- uart_monitor
\`\`\`

## uart_driver
**File:** tb/uart_driver.sv
**Extends:** uvm_driver

generate.md (if found)

markdown

# Generate Blocks

## uart_fifo generate loop
**File:** rtl/uart_fifo.sv:50

\`\`\`systemverilog
genvar i;
generate
    for (i = 0; i < DEPTH; i++) begin : gen_mem
        // memory slice
    end
endgenerate
\`\`\`

dpi.md (if found)

markdown

# DPI Functions

## Imports
| SV Name | C Name | Return | Args |
|---------|--------|--------|------|
| c_calc | calc | int | int a, int b |

## Exports
| SV Name | C Name |
|---------|--------|
| sv_notify | notify |

recipe.md

markdown

# Build Recipe

## Compile Order
1. Packages (no deps)
2. Interfaces
3. Leaf modules
4. Top modules

\`\`\`mermaid
flowchart LR
    pkg[uart_pkg.sv] --> leaf[uart_tx.sv]
    pkg --> leaf2[uart_rx.sv]
    leaf --> top[uart_ctrl.sv]
    leaf2 --> top
\`\`\`

## Filelists Found
| File | Entries |
|------|---------|
| rtl.f | 15 files |

## Include Paths
- ./rtl
- ./include

Per-Module Pages (modules/*.md)

markdown

# Module: uart_tx

## Overview
UART transmitter with configurable baud rate.

## Location
- **File:** rtl/uart_tx.sv
- **Lines:** 1-145

## Parameters
| Name | Type | Default | Description |
|------|------|---------|-------------|
| CLK_FREQ | int | 100000000 | System clock Hz |
| BAUD | int | 115200 | Baud rate |

## Ports
| Name | Dir | Width | Description |
|------|-----|-------|-------------|
| clk | input | 1 | System clock |
| rst_n | input | 1 | Active-low reset |
| tx_data | input | [7:0] | Byte to send |
| tx_valid | input | 1 | Data valid |
| tx_ready | output | 1 | Ready for data |
| tx | output | 1 | Serial output |

## Instances
None (leaf module)

## FSM
State type: tx_state_t {IDLE, START, DATA, STOP}
[Link to fsm.md#uart_tx]

## Clock/Reset
- Clock: clk (single domain)
- Reset: rst_n (async active-low)

## Assertions
- p_valid_hold: valid held until ready

Regex Patterns Reference

code

# Design Units
^\s*module\s+(\w+)              # Module
^\s*interface\s+(\w+)           # Interface
^\s*package\s+(\w+)             # Package
^\s*program\s+(\w+)             # Program
^\s*class\s+(\w+)               # Class
^\s*checker\s+(\w+)             # Checker

# Ports & Parameters
(input|output|inout)\s+(logic|wire|reg)?\s*(\[.*?\])?\s*(\w+)
parameter\s+(int|logic|integer)?\s*(\w+)\s*=
localparam\s+

# Instances
^\s*(\w+)\s*(#\s*\([^)]*\))?\s+(\w+)\s*\(

# Always blocks
always_ff\s*@|always_comb|always_latch|always\s*@

# Types
typedef\s+enum
typedef\s+struct\s+packed
typedef\s+union

# Functions/Tasks
function\s+(automatic\s+)?
task\s+(automatic\s+)?

# Preprocessor
`define\s+(\w+)
`include\s+"([^"]+)"
`ifdef|`ifndef|`elsif

# Verification
assert\s+property|assume\s+property|cover\s+property
covergroup\s+(\w+)
sequence\s+(\w+)
property\s+(\w+)
bind\s+\w+

# Generate
generate|genvar

# Interface features
modport\s+\w+
clocking\s+\w+

# OOP/UVM
class\s+\w+|extends\s+\w+|virtual\s+class
constraint\s+\w+
rand\s+|randc\s+

# DPI
import\s+"DPI|export\s+"DPI

Save Metadata After Mapping

bash

# Save commit hash
git rev-parse HEAD > .gateflow/map/.last_scan_commit

# Save timestamp
date -u +"%Y-%m-%dT%H:%M:%SZ" > .gateflow/map/.last_scan

# Save file hashes for non-git change detection
find . -name "*.sv" -o -name "*.svh" | xargs md5 > .gateflow/map/.file_hashes

Incremental Update Mode

When updating existing map:

•Get changed files: git diff --name-only <last_commit> HEAD -- "*.sv"
•Spawn subagents ONLY for changed file groups
•Merge new analysis with existing map files
•Update frontmatter timestamps
•Regenerate affected diagrams

Quality Warnings

Report in CODEBASE.md under "## Warnings":

•Inferred latches (always without default)
•Missing resets on sequential logic
•Unconnected ports
•CDC crossings without synchronizers
•Unused signals/parameters
•Missing top module

Token Budget Reference

Model	Context	Safe Budget
Sonnet	200k	150k per agent
Haiku	200k	100k per agent

Always use Sonnet for analysis quality. The 150k budget leaves 50k headroom for agent reasoning and output.

Troubleshooting

File too large for single agent:

•Use Grep to extract structure first
•Read in chunks with offset/limit
•Assign dedicated subagent

Too many files:

•Increase subagent count
•Focus on RTL, skip testbenches
•Use glob patterns to filter

No git available:

•Fall back to file hash comparison
•Store .file_hashes for change detection