agent_compositor_test/AGENTS.md

AGENTS.md — Prism

Project Identity

Prism is a node-based image compositing tool. Images flow through a directed acyclic graph (DAG) of processing nodes (read, blend, colour-grade, etc.) and are rendered via GPU shaders.

• Language: C11 / C++11 (dual-mode, like src/wapp/)
• GPU API: Vulkan, abstracted behind a Rendering Hardware Interface (RHI)
• Shading language: Slang, stored in external .slang files under src/shaders/
• Build: justfile (Just) as a task runner
• Dependencies: src/wapp/ (local utility library, already vendored)

Coding Conventions

All code follows the patterns established in src/wapp/. The project prefix is pr (functions/macros) / Pr (types).

Naming

Category	Convention	Examples
Types	Pr + PascalCase	PrNode, PrGraph, PrImage
Functions	pr + PascalCase	prGraphCreate, prNodeEval
Macros	pr + PascalCase	prArrayOf, prNodeType
Enum constants	PR_ + SCREAMING_SNAKE	PR_NODE_BLEND, PR_OK
Internal/static funcs	_ + camelCase	_resolveTopology, _execCmd
File-scope globals	_ + camelCase	_default_allocator
File-scope constants	SCREAMING_SNAKE	MAX_NODE_NAME
Variables	snake_case	vertex_count, edge_node

• Variable names: Prefer descriptive names (e.g. vertex_count over vc, edge_node over en). Short names (i, j, n) are acceptable only in tight loop counters or trivial index variables. Avoid single-letter names in non-trivial scopes, and avoid abbreviated names that require the reader to hold a mental dictionary.

Formatting

• Tabs for indentation, 8-column tab width.
• Braces on the same line as control statements (Attach style).
• Braces on single-line if statements: Always use braces, even for single-line bodies:

  // correct
  if (!buffer) { return; }
  if (!texture) { _abort("alloc failed"); }
  
  // wrong
  if (!buffer) return;
  if (!texture) _abort("alloc failed");
  

• Pointers: * against the name, not the type (PrRhiBuffer *buf, not PrRhiBuffer* buf).
• Line width: 120 columns.
• Continuation lines align to the opening parenthesis.
• Return-type alignment: Within each // ===== section, align function declaration names so the first letter of every function occupies the same column. For pointer return types, place * directly against the function name (no space) and put all alignment padding between the type name and *.

  // correct — * against fn name, padding before *
  PrRhiSwapchain      *prRhiCreateSwapchain(…);
  void                 prRhiDestroySwapchain(…);
  PrRhiSwapchainResult prRhiAcquireNextImage(…);
  

Storage qualifiers

Use the wp_extern / wp_intern / wp_persist aliases from wapp:

wp_extern void prGraphInit(PrGraph *g);     // extern linkage
wp_intern void _helper(void);               // static (file-scope)
wp_persist u32 _counter;                     // static (file-scope, mutable)

Structs

POD structs with typedef, following wapp/base/ patterns. Functions operate on structs by pointer rather than through vtables:

typedef struct PrNode PrNode;
struct PrNode {
    PrNodeType  type;
    u32         input_count;
    PrNode    **inputs;
    // ...
};

void prNodeInit(PrNode *n, PrAllocator *alloc);
void prNodeEval(PrNode *n, PrRhiCmdList *cl);
void prNodeDestroy(PrNode *n, PrAllocator *alloc);

Memory management

Use wapp allocators (WpAllocator, arena-based). Stack-allocate where possible; pass allocators explicitly.

Never use libc for memory or file I/O. wapp always takes precedence:

• wpMemAllocatorAlloc / wpMemAllocatorFree instead of malloc / free
• wpFileOpen / wpFileRead / wpFileClose instead of fopen / fread / fclose

For one-shot loads (e.g. SPIR-V at init), use &_G_RHI_CONTEXT.allocator.

PrGraph *prGraphCreate(PrAllocator *alloc);
void     prGraphDestroy(PrGraph *g, PrAllocator *alloc);

File organisation

One .h / .c pair per module, grouped in subdirectories by subsystem. A top-level umbrella header re-exports everything (see src/wapp/wapp.h).

src/prism/prism.h        ← umbrella include
src/prism/core/          ← graph, node types, DAG
src/prism/rhi/           ← rendering hardware interface
src/prism/node/          ← node implementations
src/prism/app/           ← entry point, window, main loop

Shaders live in a flat-ish src/shaders/ tree:

src/shaders/blit.slang
src/shaders/blend.slang
src/shaders/common/       ← shared include files

Data-oriented design

For hot paths (per-frame graph evaluation, image transfers, shader dispatch) prefer SoA layouts, batch processing, and minimise pointer chasing. Keep the DAG in contiguous arrays (e.g. adjacency lists packed in flat buffers) rather than individually allocated linked structures.

Frame-by-frame command batching

Commands that execute every frame must avoid arena allocation. Use stack arrays with a while-loop to batch operations in fixed-size chunks:

while (count > 0) {
	VkTypeArray batch = wpArrayWithCapacity(VkType, 16, WP_ARRAY_INIT_FILLED);
	u32 batch_size = count < wpArrayCapacity(batch) ? count : (u32)wpArrayCapacity(batch);
	// ... process batch ...
	vkCmd*(cb->handle, ...);
	count -= batch_size;
	first += batch_size;
}

This avoids per-frame arena churn while handling arbitrarily large inputs.

Graph / adjacency lists

Adjacency list nodes must be separately allocated from the vertex array. Never use the vertex struct itself as a linked-list node in another vertex's adjacency chain — that shares the next pointer between two roles (vertex state vs. edge linking) and causes edges to splice into unintended chains.

// correct — per-edge copy on the arena
static void addEdge(PrGraph *g, const WpAllocator *alloc, u64 from, u64 to) {
	PrVertex *src = &g->vertices[from];
	PrVertex *dst = wpMemAllocatorAlloc(alloc, sizeof(PrVertex));
	if (!dst) { /* handle OOM */ return; }
	dst->id    = g->vertices[to].id;
	dst->value = g->vertices[to].value;
	dst->next  = src->next;
	src->next  = dst;
}

// wrong — reuses the destination vertex as the list node
static void addEdge_bad(PrGraph *g, u64 from, u64 to) {
	PrVertex *src  = &g->vertices[from];
	PrVertex *dest = &g->vertices[to];
	dest->next = src->next;
	src->next  = dest;   // overwrites dest->next used elsewhere
}

Arena bump allocation (wpMemAllocatorAlloc) is the natural fit for building graph structures: per-edge nodes live in the arena and are freed in one shot when the arena is destroyed.

Always NULL-check the result of wpMemAllocatorAlloc — even arena allocators can fail if the backing buffer is exhausted.

WpArray usage

Scenario	API	Size must be…
Stack (local, short-lived)	wpArrayWithCapacity	compile-time constant
Heap (arena-backed)	wpArrayAllocCapacity	runtime value

wpArrayWithCapacity creates a VLA-like compound literal on the stack — passing a runtime variable triggers undefined behaviour and compiler warnings. Use wpArrayAllocCapacity with an arena allocator for runtime sizes.

Always use typed array aliases (WpU64Array, PrNodeIdArray, etc.) rather than raw pointers when declaring array variables. Follow the existing typedef pattern in the module (typedef Type *TypeArray).

Typedef pattern:

• Opaque handles use ** (pointer-to-pointer)
• Value types use * (contiguous block)

typedef PrRhiBuffer         **PrRhiBufferArray;           // opaque handles → **
typedef PrRhiColorAttachment *PrRhiColorAttachmentArray;  // value types → *

Group opaque handle arrays first, value type arrays second, separated by a blank line.

Use named init flags (WP_ARRAY_INIT_NONE, WP_ARRAY_INIT_FILLED) instead of bare 0 — they make the initialisation policy explicit.

• WP_ARRAY_INIT_FILLED: sets count = capacity on allocation. Required when you plan to index into the array directly (not via append/push), since the array's count must reflect valid elements for any downstream use.
• WP_ARRAY_INIT_NONE: leaves count = 0. Use when you'll fill the array incrementally via wpArrayAppendCapped / wpArrayAppendAlloc.

Use wpArrayCapacity(arr), wpArrayCount(arr), wpArraySetCount(arr, n) to query and control array state rather than computing sizes manually.

Local/scratch arenas

For function-local scratch allocations, use wpMemArenaAllocatorInitZero with a fixed size rather than a stack buffer + InitWithBuffer:

WpAllocator scratch = wpMemArenaAllocatorInitZero(KiB(16));

Documentation

Save research notes and implementation plans as markdown in documents/:

documents/
├── ARCHITECTURE.md
├── RENDERING_HARDWARE_INTERFACE.md
├── NODE_SYSTEM.md
├── ROADMAP.md
├── research/
│   └── <topic>.md
└── session-logs/
    └── YYYY-MM-DD.md

Skills

Domain-specific conventions are stored as skills in .opencode/skills/<name>/SKILL.md. These are loaded on-demand by the AI agent when a task matches their description, keeping AGENTS.md lean.

• prism-rhi — RHI backend dispatch, by-value descs, file layout
• prism-dag — DAG adjacency lists, arena allocation, Kahn's algorithm

Workflows for AI agents

Research / planning

1. Save findings to documents/research/<topic>.md.
2. When asked for a plan, write it to documents/<PLAN_NAME>.md and present a summary; iterate on the plan before writing any code.
3. Only start implementing after the plan is approved.

Skill maintenance

When you observe the user correcting your output (e.g. formatting, conventions), infer the rule and add it to the relevant skill's SKILL.md. If no skill matches, add a new one. This keeps AGENTS.md focused on project identity and critical workflow rules rather than accumulating domain details.

Committing

Only commit when explicitly asked. When asked:

• Stage only intended files.
• Write a short, conventional commit message in present tense.
• Never amend, force-push, or create PRs without a request.