romdev-mcp 0.1.1

package/src/platforms/nes/TROUBLESHOOTING.md

@@ -0,0 +1,321 @@
+# NES — symptom → fix
+
+Find your symptom below; each entry has the 1-line diagnosis + the
+MCP tool call that confirms it. Run these BEFORE you start bisecting
+your C source.
+
+## "Screen is grey / random / never changes"
+
+**Cause #1: palette never written.** Power-on palette RAM is
+undefined. Symptom: a single ugly color (often $0F → grey-ish or
+$00) fills the screen. Fix: `palette_load(palette)` BEFORE
+`ppu_on_all()`.
+```js
+loadCategory({category:"memory"})
+readMemory({region:"nes_palette", offset:0, length:32})
+// Expect non-random bytes 0..0x3F.
+```
+
+**Cause #2: PPU rendering disabled.** PPUMASK = 0 means the PPU
+doesn't render. `ppu_on_all()` sets bit 3 (BG) + bit 4 (sprites) =
+$1E.
+```js
+loadCategory({category:"debug"})
+getRenderingContext({})
+// Look for "bgVisible: true, spritesVisible: true".
+```
+
+**Cause #3: PPU not warmed up.** PPU ignores writes for ~29 658
+cycles after reset. Our crt0 waits for two vblanks. If you replaced
+the crt0, restore the two `bit $2002; bpl` waits.
+
+## "Sprite doesn't appear"
+
+Check in this order:
+
+1. **OAM byte 0 is the Y position. Y=$EF..$FF hides the sprite.**
+   Hardware Y = screen Y - 1, so the top edge of the screen is
+   OAM Y=0 (i.e. screen Y=1).
+   ```js
+   readMemory({region:"nes_oam", offset:0, length:16})
+   // bytes 0..3 = slot 0: [Y, tile, attr, X]
+   ```
+
+2. **Sprite pattern table mismatch.** Default PPUCTRL has bit 3 = 0,
+   so sprites fetch from $0000-$0FFF. If your sprite tile is at
+   $1000+ you'll render BG tiles as sprites (= corrupt).
+
+3. **PPUMASK.4 (OBJ enable) is off.** `ppu_on_all()` turns it on.
+
+4. **Sprite uploaded but slot Y still $FF.** `oam_clear()` sets Y to
+   $FF and resets `oam_index` to 0. You must call `oam_spr` AFTER
+   `oam_clear`.
+   ```js
+   inspectSprites({})    // shows which slots are visible
+   ```
+
+5. **The Y off-by-one bug.** `oam_spr` subtracts 1 from Y to convert
+   "screen coordinate" to "OAM byte 0". If you write shadow_oam
+   directly, do this yourself: `shadow_oam[0] = (screen_y - 1)`.
+
+## "Wrong colors"
+
+1. **Universal backdrop ≠ what you set.** `$3F00` is the backdrop.
+   The PPU also reads $3F04/$3F08/$3F0C as backdrop if a sprite is
+   transparent above a BG-color-0 cell — there's mirroring weirdness.
+   Set them all to the same value to avoid surprises.
+
+2. **Attribute table boundaries are 2×2 tile groups.**
+   `tile_set_palette(nt, x, y, p)` changes the palette of a 2×2
+   group. Setting one tile changes its 3 neighbors. There's no way
+   to give adjacent 2×2 groups different palettes without
+   accepting the boundary.
+
+3. **PPUMASK.0 is "greyscale mode."** Set it accidentally and every
+   color is mapped to nearest grey. Check via `getRenderingContext`.
+   ```js
+   getRenderingContext({})    // "grayscale" should be false
+   ```
+
+## "Game freezes / locks up"
+
+1. **NMI vector points at $0000.** If you replaced the crt0 and
+   didn't define `nmi:`, the CPU jumps to garbage and hangs.
+   ```js
+   readMemory({region:"system_ram", ...})    // or:
+   xxd file.nes | tail -1                    // last 6 bytes = NMI, RESET, IRQ
+   ```
+
+2. **`ppu_wait_nmi()` spinning forever.** That means NMI never fires.
+   Check that PPUCTRL bit 7 is set:
+   ```js
+   getRenderingContext({})    // "nmiEnabled" should be true
+   ```
+   If not, `ppu_on_all()` or `ppu_on_bg()` should set it. If you wrote
+   to PPUCTRL manually with bit 7 off, you've disabled NMI.
+
+3. **`oam_clear()` + `oam_spr()` mid-frame race.** If you call these
+   AFTER `ppu_wait_nmi()` (i.e. just after vblank started), the cc65-
+   generated code is slow enough that you may not finish before the
+   next vblank, and the NMI's DMA captures shadow_oam mid-update.
+   Fix: stage sprites BEFORE `ppu_wait_nmi()`, not after.
+
+## "Sprites flicker on and off every other frame"
+
+Your loop order is `ppu_wait_nmi → oam_clear → oam_spr` instead of
+the canonical `oam_clear → oam_spr → ppu_wait_nmi`. The NMI fires
+at vblank start and DMAs whatever shadow_oam contains AT THAT
+MOMENT. With the wrong order, you populate shadow_oam AFTER the
+NMI has already run, so half the frames DMA the cleared
+(all-`$FF`) shadow_oam.
+
+Diagnostic: sample `nes_oam[0..3]` over 8 consecutive frames.
+Wrong-order produces alternating `sprite/ff000000/sprite/ff000000`.
+Correct-order produces `sprite/sprite/sprite/sprite`.
+
+Fix: stage shadow_oam BEFORE waiting.
+```c
+for (;;) {
+  /* update game state */
+  oam_clear(); oam_spr(...);   // stage
+  tile_set(...);                // queue BG writes
+  ppu_wait_nmi();               // sleep — NMI DMAs + flushes
+}
+```
+
+## "BG tile visible in inspectPatternTiles but not on screen"
+
+The NES PPU renders 240 scanlines but most TVs (and our
+framebuffer output) crop to 224. Nametable row 0 (PPU $2000-$201F,
+top 8 px) and row 29 (PPU $23A0-$23BF, bottom 8 px) get cut. If
+your score lives at row 0 it'll be in the nametable but invisible.
+
+Fix: move HUD to row 2+ and keep "press start"-style bottom UI at
+row 27 or lower.
+
+## "State corrupts / mystery crashes as the game grows" — RAM/BSS overflow
+
+The NES has only **2 KB of RAM** ($0000-$07FF), and the chr-ram linker
+config carves most of it up: zeropage, the stack, and `shadow_oam`
+(256 B at $0200) leave roughly **~512 B for your BSS+DATA** (globals).
+Overflow it and there's no error — your globals quietly collide with
+the stack or shadow OAM → corrupted state, sprites that flicker to
+garbage, random crashes.
+
+**Check the `ramUsage` field in the buildSource/runSource response** —
+it lists your BSS / DATA / ZEROPAGE segment sizes from the linker map.
+If BSS+DATA is approaching the config's RAM region, shrink your state:
+prefer `uint8_t` over `int`, bit-pack flags, use small fixed arrays,
+avoid large `static` buffers. (This is why "NES-shaped C" uses bitmasks
+and tiny structs — it's not style, it's the 512 B ceiling.)
+
+## "runSource screenshot looks one frame behind my sprites"
+
+On NES, the NMI handler DMAs `shadow_oam` → real OAM at the *start* of
+each vblank, so sprites you stage on frame N first appear when frame
+N+1 renders. `runSource` now steps one extra frame on NES before the
+screenshot so it matches your staged OAM — but if you script frames
+manually (`stepFrames` then `screenshot`), add one extra `stepFrames(1)`
+after staging to see the current sprite positions.
+
+## "readMemory(nes_chr) returns same bytes for offset 0 and offset 4096"
+
+Was a real bug in R59 of the fceumm patch — `readMemory(nes_chr)`
+collapsed all 8 1KB pages into copies of the first page on NROM.
+Fixed in R61 (2026-05-27). If you still see this, your MCP server
+is running a stale WASM:
+```sh
+# Check the timestamp of the bundled fceumm wasm in your install:
+stat -c '%y' node_modules/romdev-core-fceumm/wasm/fceumm_libretro.wasm
+```
+Workaround if you can't restart: `inspectPatternTiles` reads CHR
+via a different path and was unaffected throughout.
+
+## "Globals read garbage / `_nmi_counter` never advances"
+
+Your `.cfg` puts the BSS region (`RAM:`) at $6000-$7FFF, but your
+cart is NROM with no battery WRAM (iNES flags6 bit 1 = 0). That
+region is UNMAPPED on the hardware — reads return open-bus
+(parasitic capacitance from the last bus value). Globals appear
+to "kind of work" because the open bus briefly retains what you
+just wrote, but `_nmi_counter` increments are lost and any
+"wait until counter == target" loop spins forever.
+
+Diagnostic:
+```js
+readMemory({region:"system_ram", offset:0x6000, length:16})
+// All $FF or matching the last value you wrote = open bus.
+```
+
+Fix: edit your `.cfg`:
+```
+RAM:    file = "", start = $0300, size = $0200, define = yes;
+```
+(512 bytes of real internal RAM between OAM at $0200 and the C
+stack at $0500.) For larger BSS, opt into a mapper that maps
+PRG-RAM at $6000 and set iNES flags6 bit 1.
+
+The bundled `chr-ram-runtime.cfg` already does this correctly as
+of 2026-05-27.
+
+## "OAM is full of $FF — DMA must be broken"
+
+Probably not. The crt0's init loop writes `$FF` to all 256 bytes
+of `_shadow_oam @ $0200` at boot (canonical sprite-Y off-screen
+sentinel). If your game only populates a few slots, the rest stay
+`$FF` and OAM DMA faithfully copies them. `readMemory(nes_oam)`
+showing mostly `$FF` is the EXPECTED state, not a bug.
+
+Sentinel test that proves DMA works:
+```js
+pause()
+writeMemory({region:"system_ram", offset:0x0200, hex:"42".repeat(256)})
+resume(); stepFrames({frames:1})
+readMemory({region:"nes_oam", offset:0, length:16})
+// All $42 → DMA copies the source page faithfully. Working as designed.
+// All $FF → real DMA bug. Escalate.
+```
+
+## "retro_load_game failed"
+
+The build pipeline writes a valid iNES header (mapper 0, vertical
+mirroring, CHR-RAM=0) for every C build via our chr-ram crt0. If
+you're seeing this error, you're either:
+
+1. Loading a hand-built ROM that skipped the bundled crt0.
+2. The ROM has stale CHR-ROM bytes that don't match the header's
+   "CHR-RAM" declaration. Re-build with `linkerConfig:"chr-ram"`
+   (or just leave linkerConfig blank — the C path defaults to chr-ram).
+
+Verify the header:
+```sh
+xxd -l 16 file.nes
+# Expect: 4E 45 53 1A 02 00 01 00 00 00 00 00 00 00 00 00
+#         N  E  S  ^  PRG CHR FL FL
+#                  16 banks=2 0=CHR-RAM vmirror standard
+```
+
+## "Pad input does nothing"
+
+1. **Wrong bit layout.** Our `pad_poll` uses the Shiru/neslib layout:
+   `PAD_A = 0x80`, `PAD_RIGHT = 0x01`. **Not** the cc65 `joy.h`
+   layout. Use the PAD_* masks from `nes_runtime.h`.
+
+2. **Strobed wrong.** `pad_poll` does the strobe internally. Don't
+   call it from inside an NMI handler — the strobe sequence collides
+   with the controller pollers some emulators run during DMA.
+
+3. **Controller 1 vs 2.** `pad_poll(0)` reads $4016 (controller 1).
+   `pad_poll(1)` reads $4017 (controller 2). Most games only use 0.
+
+## "BG renders garbage / random tiles"
+
+1. **CHR-RAM wasn't initialised.** The crt0 zeros all 8 KB at boot.
+   If you skipped the crt0, do it yourself: `chr_ram_upload(0x0000,
+   zeros, 0x2000)` at boot before `palette_load`.
+
+2. **PPUCTRL bit 4 wrong.** Bit 4 = 0 means BG fetches from $0000;
+   bit 4 = 1 means BG fetches from $1000. Default is bit 4 set
+   ($1000). If you upload BG tiles to $0000 (sprite area) and forget
+   to change PPUCTRL, BG won't see them.
+
+3. **Nametable wasn't written.** A nametable of all zeros = all tile
+   0. If your tile 0 is blank (which is the convention), you get a
+   black screen. Use `vram_unsafe_set` during PPU-off or `tile_set`
+   from C (queued).
+
+## "Scroll glitches / flickers"
+
+1. **NMI didn't reset PPUADDR/PPUSCROLL.** The bundled NMI handler
+   does this. If you replaced it, restore the sequence:
+   `bit $2002; lda #$20; sta $2006; sta $2006; lda scroll_x; sta $2005;
+   lda scroll_y; sta $2005; lda ppuctrl_value; sta $2000`.
+
+2. **You wrote to PPUADDR outside vblank.** $2006 writes during
+   rendering corrupt the internal scroll latch. Always queue VRAM
+   writes via `vram_set` / `tile_set` so they flush in NMI.
+
+## Debug recipes
+
+A few high-leverage tools you might not know exist:
+
+- **`inspectSprites({})`** — pretty-prints all 64 OAM slots with
+  visible/hidden status, tile, attr, position.
+- **`inspectBackgroundMap({render:true})`** — composites the active
+  nametable as a PNG.
+- **`inspectPalette({})`** — shows the loaded palette as RGB colors,
+  not raw bytes.
+- **`getRenderingContext({})`** — decodes PPUCTRL/PPUMASK/PPUSTATUS
+  + "what bank are BG/sprites fetching from right now".
+- **`getCPUState({})`** — PC + flags. Use when you suspect a hang.
+- **`watchMemory({region:"nes_oam", offset:0, length:4})`** — trace
+  every write to OAM slot 0, returns the PC that wrote it.
+- **`stepFrames({frames:3600})`** — runs 1 minute of game time in
+  milliseconds. Don't be conservative.
+
+## Mental model + boot order
+
+See [`MENTAL_MODEL.md`](MENTAL_MODEL.md) for the architecture
+overview (fetch via `getPlatformDoc({platform:"nes", name:"mental_model"})`).
+The canonical "boots cleanly + shows a sprite" sequence is:
+
+```
+1. ppu_off()                       // safe to write VRAM
+2. chr_ram_upload(0x0000, ...)     // sprite tiles
+3. chr_ram_upload(0x1000, ...)     // BG tiles
+4. palette_load(palette)            // 32-byte palette at $3F00
+5. render initial nametable        // vram_unsafe_set during PPU-off
+6. oam_clear() + oam_spr(...)      // initial sprite list
+7. ppu_on_all()                    // enable rendering + NMI
+8. for(;;) {
+     stage sprites + tile_set      // for upcoming frame
+     ppu_wait_nmi()                // NMI auto-DMAs + commits queue
+     pad = pad_poll(0)             // input
+     update game state
+   }
+```
+
+## Adding a new symptom
+
+Hit something not on this list? Open an issue with a 5-line repro and
+the symptom at https://github.com/monteslu/romdev/issues.

package/src/platforms/nes/UPSTREAM_SOURCES.md

@@ -0,0 +1,31 @@
+# NES — source you can read
+
+Trust hierarchy:
+
+1. **Bundled examples** (`examples/nes/templates/*.c`).
+2. **Bundled runtime** — `src/platforms/nes/lib/c/nes_runtime.c`,
+   asm helpers in `lib/*.s`, FamiTone2 source at `lib/asm/famitone2.s`.
+3. **Bundled cc65 NES libsrc** (R58) at `src/platforms/nes/lib/
+   cc65-src/`. Full cc65 NES target source: joystick driver, conio,
+   PPU helpers, header builder, etc.
+4. **Upstream GitHub**:
+
+   | What | Upstream |
+   |---|---|
+   | cc65 | https://github.com/cc65/cc65 |
+   | famitone2 + tools | https://shiru.untergrund.net/code.shtml |
+   | fceumm libretro core | https://github.com/libretro/libretro-fceumm |
+
+## NES hardware docs
+
+- **NESdev Wiki** (canonical, everything you'd need): https://www.nesdev.org/wiki
+- **NESdev guide series**: https://www.nesdev.org/wiki/Tutorials
+- **6502 instruction reference**: https://www.masswerk.at/6502/6502_instruction_set.html
+
+## When to use what
+
+- "How does cc65's `joy_read(JOY_1)` poll the controller?" →
+  `lib/cc65-src/joystick/nes-stdjoy.s`
+- "Why does my CHR-RAM write not appear?" → NESdev wiki "PPU memory map"
+- "FamiTone2 song format?" → `lib/asm/famitone2.s` header comment + Shiru's docs
+- "fceumm doesn't expose register X" → libretro-fceumm GitHub

package/src/platforms/nes/image-to-chr.js

@@ -0,0 +1,195 @@
+// Convert a PNG (or raw RGBA image) into NES CHR bytes.
+//
+// NES tiles are 8×8 pixels, 2 bits per pixel (4 colors per tile, from a
+// 4-entry palette). CHR ROM packs each tile into 16 bytes: 8 bytes of bit
+// plane 0 followed by 8 bytes of bit plane 1.
+//
+// To convert arbitrary art:
+//   1. For each 8×8 tile in the input image:
+//      a. Pick the 4 most common colors in that tile.
+//      b. Map each pixel to the index of its nearest color in that
+//         per-tile palette.
+//   2. Encode the indices as bitplanes.
+//
+// Returns the CHR bytes plus a recommended NES master palette that's the
+// closest match to the colors actually used across all tiles. The agent
+// can then write both into a ROM via patchRom.
+
+import { PNG } from "pngjs";
+import { NES_PALETTE } from "./palette.js";
+
+/**
+ * @typedef {Object} ConvertResult
+ * @property {Uint8Array} chr           CHR bytes (16 × number-of-tiles)
+ * @property {Uint8Array} palette       suggested 32-byte NES palette (16 BG + 16 sprite mirrored)
+ * @property {number} tilesAcross
+ * @property {number} tilesDown
+ * @property {number} totalTiles
+ * @property {Object[]} tilePalettes    one per tile: {indices: [4 NES master indices]}
+ */
+
+/**
+ * Convert a PNG buffer to CHR bytes.
+ *
+ * @param {Buffer | Uint8Array} pngBytes      PNG-encoded image (any source — agent-generated, ripped from another ROM)
+ * @param {Object} [opts]
+ * @param {number} [opts.maxTiles] hard cap on tile count (e.g. 256 = one pattern table)
+ * @returns {ConvertResult}
+ */
+export function imageToChr(pngBytes, opts = {}) {
+  const png = PNG.sync.read(Buffer.from(pngBytes));
+  return rgbaToChr({
+    width: png.width,
+    height: png.height,
+    pixels: png.data,
+    ...opts,
+  });
+}
+
+/**
+ * Convert a raw RGBA image to CHR. Same logic as imageToChr but works on
+ * an already-decoded buffer.
+ *
+ * @param {Object} args
+ * @param {number} args.width
+ * @param {number} args.height
+ * @param {Uint8Array | Buffer} args.pixels   RGBA, length = width*height*4
+ * @param {number} [args.maxTiles]
+ * @returns {ConvertResult}
+ */
+export function rgbaToChr(args) {
+  const { width, height, pixels } = args;
+  if (width % 8 !== 0) {
+    throw new Error(`width must be a multiple of 8, got ${width}`);
+  }
+  if (height % 8 !== 0) {
+    throw new Error(`height must be a multiple of 8, got ${height}`);
+  }
+
+  const tilesAcross = width / 8;
+  const tilesDown = height / 8;
+  let totalTiles = tilesAcross * tilesDown;
+  if (args.maxTiles && totalTiles > args.maxTiles) {
+    // Truncate to fit. We render row-major, so this drops bottom tiles.
+    totalTiles = args.maxTiles;
+  }
+
+  const chr = new Uint8Array(totalTiles * 16);
+  /** @type {{indices: number[]}[]} */
+  const tilePalettes = [];
+
+  // Track NES color usage frequency across all tiles, so we can suggest a
+  // single 32-byte palette that covers the art.
+  const nesUsage = new Int32Array(64);
+
+  for (let t = 0; t < totalTiles; t++) {
+    const tx = t % tilesAcross;
+    const ty = Math.floor(t / tilesAcross);
+
+    // Sample the 64 pixels of this tile.
+    /** @type {[number,number,number][]} */
+    const tilePixels = [];
+    for (let y = 0; y < 8; y++) {
+      for (let x = 0; x < 8; x++) {
+        const px = tx * 8 + x;
+        const py = ty * 8 + y;
+        const o = (py * width + px) * 4;
+        tilePixels.push([pixels[o], pixels[o + 1], pixels[o + 2]]);
+      }
+    }
+
+    // Quantize the tile to a 4-color palette by finding the 4 most-distinct
+    // representative colors. Simple approach: pick the 4 most-common colors,
+    // or if fewer unique colors exist, pad.
+    const palette = chooseTilePalette(tilePixels);
+    // Map each NES master index used to the global usage table.
+    for (const nesIdx of palette) {
+      if (nesIdx >= 0 && nesIdx < 64) nesUsage[nesIdx]++;
+    }
+
+    tilePalettes.push({ indices: palette });
+
+    // Encode the tile as bitplanes.
+    for (let y = 0; y < 8; y++) {
+      let lo = 0;
+      let hi = 0;
+      for (let x = 0; x < 8; x++) {
+        const rgb = tilePixels[y * 8 + x];
+        const palIdx = nearestPalIdx(rgb, palette);
+        if (palIdx & 1) lo |= 1 << (7 - x);
+        if (palIdx & 2) hi |= 1 << (7 - x);
+      }
+      chr[t * 16 + y] = lo;
+      chr[t * 16 + y + 8] = hi;
+    }
+  }
+
+  // Build a recommended 32-byte palette: pick the most-used NES master indices.
+  const sortedNes = [];
+  for (let i = 0; i < 64; i++) sortedNes.push({ idx: i, count: nesUsage[i] });
+  sortedNes.sort((a, b) => b.count - a.count);
+  const palette = new Uint8Array(32);
+  // The universal BG entry is the most-used color overall, ideally background.
+  palette[0] = sortedNes[0]?.idx ?? 0x0f;
+  // Fill in 3 BG palettes + 4 sprite palettes with the next most-used colors.
+  for (let p = 0; p < 4; p++) {
+    for (let i = 1; i < 4; i++) {
+      const top = sortedNes[(p * 3 + i) % sortedNes.length];
+      palette[p * 4 + i] = top ? top.idx : 0x0f;
+    }
+    // Mirror universal BG into sprite half too.
+    palette[16 + p * 4] = palette[0];
+    for (let i = 1; i < 4; i++) {
+      palette[16 + p * 4 + i] = palette[p * 4 + i];
+    }
+  }
+
+  return { chr, palette, tilesAcross, tilesDown, totalTiles, tilePalettes };
+}
+
+/** Choose 4 NES master-palette indices that best represent the tile's colors. */
+function chooseTilePalette(tilePixels) {
+  // Quantize each pixel to its nearest NES master index, then take the 4 most-common.
+  const counts = new Map();
+  for (const rgb of tilePixels) {
+    const idx = nearestNesIndex(rgb);
+    counts.set(idx, (counts.get(idx) ?? 0) + 1);
+  }
+  const sorted = [...counts.entries()].sort((a, b) => b[1] - a[1]);
+  const palette = sorted.slice(0, 4).map((e) => e[0]);
+  while (palette.length < 4) palette.push(0x0f); // pad with NES black
+  return palette;
+}
+
+/** Map an RGB to the nearest entry in the NES 64-color master palette. */
+function nearestNesIndex(rgb) {
+  const [r, g, b] = rgb;
+  let bestIdx = 0;
+  let bestDist = Infinity;
+  for (let i = 0; i < 64; i++) {
+    const [pr, pg, pb] = NES_PALETTE[i];
+    const dr = r - pr, dg = g - pg, db = b - pb;
+    const d = dr * dr + dg * dg + db * db;
+    if (d < bestDist) {
+      bestDist = d;
+      bestIdx = i;
+    }
+  }
+  return bestIdx;
+}
+
+/** Pick the 0..3 palette slot whose NES color is closest to the pixel. */
+function nearestPalIdx(rgb, palette) {
+  let bestIdx = 0;
+  let bestDist = Infinity;
+  for (let i = 0; i < palette.length; i++) {
+    const [pr, pg, pb] = NES_PALETTE[palette[i] & 0x3f];
+    const dr = rgb[0] - pr, dg = rgb[1] - pg, db = rgb[2] - pb;
+    const d = dr * dr + dg * dg + db * db;
+    if (d < bestDist) {
+      bestDist = d;
+      bestIdx = i;
+    }
+  }
+  return bestIdx;
+}