claude-mcp-workflow 0.1.7 → 0.1.9

package/templates/skills/domain-reid/SKILL.md

@@ -11,7 +11,7 @@ BNNeck = BatchNorm before classifier. The trick:
 - **Training**: Euclidean distance in triplet loss (features BEFORE BNNeck)
 - **Inference**: Cosine similarity (features AFTER BNNeck)
 
-Sonnet and others commonly reverse this — getting the direction wrong silently degrades results.
+This direction is commonly reversed — getting it wrong silently degrades results.
 
 ## Loss Recipe
 
@@ -22,19 +22,21 @@ The center loss weight (lambda=0.0005) is critical — too high destabilizes tra
 ## CLIP-ReID
 
 Two-stage approach (NOT generic CLIP fine-tuning):
-1. Learn **text tokens per identity** (freeze visual encoder)
-2. Fine-tune **visual encoder** with learned text supervision
+1. Learn **text tokens per identity** — BOTH encoders frozen; only the id-specific text tokens train
+2. Fine-tune **visual encoder** with learned text supervision (text side stays fixed)
 
-SOTA on MSMT17: **~86.7% mAP** (ViT-B). Sonnet underestimates at 70-75%.
+MSMT17 **~86.7% mAP** is the SIE+OLP **+ re-ranking** configuration — plain ViT-B CLIP-ReID lands in the low-to-mid 70s mAP. Don't quote 86.7 as the vanilla-model number.
 
 ## SOTA Performance Reference
 
 | Method | Market-1501 R1/mAP | MSMT17 R1/mAP |
 |--------|---------------------|----------------|
-| CLIP-ReID (ViT-B) | 96.4 / 93.3 | 91.1 / 86.7 |
+| CLIP-ReID (ViT-B, SIE+OLP, +re-rank) | 96.4 / 93.3 | 91.1 / 86.7 |
 | TransReID (ViT-B) | 95.2 / 89.5 | 86.2 / 69.4 |
 | BoT (ResNet-50) | 94.5 / 85.9 | 77.5 / 47.5 |
 
+The CLIP-ReID row already includes k-reciprocal re-ranking — do NOT add the re-ranking boost from the section below on top of these numbers (double-count).
+
 ## Key Datasets
 
 | Dataset | IDs | Images | Notes |
@@ -49,3 +51,35 @@ Note: DukeMTMC-reID was **retracted** due to privacy concerns — avoid citing i
 ## k-Reciprocal Re-Ranking
 
 Parameters: k1=20, k2=6, lambda=0.3. Boosts mAP by 5-10%. Use for offline, skip for real-time.
+
+## torchreid install — from git, NOT PyPI
+
+`pip install torchreid` fetches a **stale 0.2.5** (2019) with a different API — `import torchreid`
+often outright fails on a modern torch/numpy. The real KaiyangZhou/deep-person-reid is **1.4.x, git-only**. Its `setup.py` imports numpy/Cython
+at build time, so a plain `pip install git+...` **fails under PEP-517 build isolation**
+(`error: getting requirements to build wheel` → `ModuleNotFoundError: No module named 'torchreid'`
+/ numpy). Two working installs (both need numpy+Cython in the env):
+`pip install --no-build-isolation git+https://github.com/KaiyangZhou/deep-person-reid.git`, or the
+official `git clone … && cd deep-person-reid && pip install -r requirements.txt && python setup.py develop`
+(its requirements.txt does NOT pin torch, so a preinstalled CUDA torch survives).
+So a "package is installed but import raises" state ≠ missing — detect it with
+`importlib.util.find_spec(m) is not None` + a real import attempt, and surface the actual exception
+(don't report "missing" and re-suggest the PyPI name that caused it).
+
+## torchreid (deep-person-reid) — query/gallery need DIFFERENT camids
+
+`torchreid.metrics.evaluate_rank` (run by every engine at the final epoch, and whenever
+`eval_freq` fires) drops, **for each query, every gallery sample sharing that query's
+`(pid, camid)`**, then asserts `num_valid_q > 0` → `AssertionError: all query identities do not
+appear in gallery`. A query is "valid" only if the same pid appears in gallery under a *different*
+camid.
+
+Trap when registering a custom `ImageDataset` for fine-tuning: if you build a throwaway
+query/gallery by splitting one identity's crops but leave them under one camid, every query is
+invalid → training crashes at the FINAL-EPOCH eval (not at data load, so it looks like a late,
+unrelated failure). Fix: assign query camid=0, gallery camid=1 (or any distinct pair) with the
+same pids on both sides.
+
+Note the value gap vs cosine ReID: the evaluator ranks by that removal rule (Market1501 protocol);
+it is NOT a plain cosine-margin metric, so its CMC/mAP on a tiny throwaway split is meaningless —
+do the real margin eval separately.

package/templates/skills/domain-yolo/SKILL.md

@@ -5,19 +5,30 @@ description: YOLO object detection model selection
 
 # YOLO — Verified Gotchas
 
+## YOLO26 — Current Default
+
+YOLO26 is NMS-free end-to-end (no NMS post-processing step), drops DFL, and delivers ~43% faster CPU inference. Same Ultralytics API — change the `.pt` file. Prefer it for new projects, especially CPU/edge deployment.
+
 ## YOLO11 vs YOLOv8
 
-YOLO11: 22% fewer params than v8m at same mAP. Refined C3k2 blocks. **Drop-in replacement** — change `.pt` file, code identical. Same Ultralytics API.
+YOLO11: higher mAP than v8m with 22% fewer params. Refined C3k2 blocks. **Drop-in replacement** — change `.pt` file, code identical. Same Ultralytics API.
+
+## YOLOv10 — Legacy NMS-Free
+
+YOLOv10 pioneered NMS-free inference, but its headline "1.8x faster" figure is vs an RT-DETR-R18 baseline, not vs other YOLOs. Superseded by YOLO26 for the NMS-free niche.
 
-## YOLOv10 — NMS-Free
+## Zero-Shot: YOLOE Supersedes YOLO-World
 
-YOLOv10 unique feature: NMS-free inference (1.8x faster end-to-end). No `iou` parameter needed. Use when latency-critical.
+YOLOE covers text prompts, visual prompts, AND a prompt-free mode at real-time speed. Reach for YOLO-World only if a dependency pins it.
 
 ## Model Selection Quick Reference
 
 | Model | When | Key Feature |
 |-------|------|-------------|
-| YOLO11 | Default choice | 22% fewer params than v8 |
-| YOLOv8 | Mature, proven production | Anchor-free, unified API |
-| YOLOv10 | Latency-critical | NMS-free |
-| YOLO-World | Zero-shot | Text prompt detection |
+| YOLO26 | Default choice | NMS-free end-to-end, ~43% faster CPU inference |
+| YOLO11 | Proven/stable production | Higher mAP than v8 with 22% fewer params |
+| YOLO12 | Attention-based research | Attention-centric (area attention) |
+| YOLOv8 | Mature, widest ecosystem | Anchor-free, unified API |
+| YOLOv10 | Legacy | First NMS-free (1.8x claim is vs RT-DETR-R18) |
+| YOLOE | Zero-shot / open-vocabulary | Text/visual prompts + prompt-free mode |
+| YOLO-World | Superseded by YOLOE | Text prompt detection |

package/templates/skills/ide-zed/SKILL.md

@@ -0,0 +1,23 @@
+---
+name: ide-zed
+description: Zed IDE integration rules
+---
+
+# Zed IDE Integration
+
+1. **`Read` can't read binary files** (PNG, etc.) — Zed bug #48133 (open, P2). Use display tree for state, not screenshots.
+2. **Stale Playwright processes accumulate** → `pkill -f playwright 2>/dev/null` if browser behaves unexpectedly.
+3. **clangd relative `-I` paths in `.clangd` resolve from the source file's directory, not the project root** (when no `compile_commands.json` exists). Files in subdirectories (e.g. `project/common/foo.cpp`) will fail to find includes specified as `-I./some-dir` in `.clangd` `CompileFlags.Add`.
+
+   WRONG — `.clangd`:
+   ```yaml
+   CompileFlags:
+     Add: [-I./native, -I./project/common]
+   ```
+
+   RIGHT — `compile_flags.txt` in the project root:
+   ```
+   -I./native
+   -I./project/common
+   ```
+   clangd uses the `compile_flags.txt` parent directory as CWD, so paths resolve correctly from the root. Keep `.clangd` for non-path flags (warnings, diagnostics, `If`/`PathMatch` overrides) only.

package/templates/skills/lang-as3/SKILL.md

@@ -3,19 +3,21 @@ name: lang-as3
 description: AS3 / AIR 51 language gotchas — optional chaining bugs, immutability limits
 ---
 
-## AS3 / AIR 51 Language Gotchas
+# AS3 / AIR 51 — Verified Gotchas
 
-### Optional chaining `?.`
+## Optional chaining `?.`
 
 - **Property access**: `obj?.prop` — works, returns `undefined` if null
 - **Method calls**: `obj?.method()` — **BROKEN**. Calls `null()` instead of short-circuiting. Use `if (obj != null) obj.method()`
-- Codebase marks reverted `?.` with comments `// ?. air bugfix`
+- When reverting a broken `?.` call to an explicit null-check, leave a marker comment at the site (e.g. `// ?. AIR bugfix — don't restore optional chaining`) so later edits don't reintroduce the bug
 
-### Immutability
+## Immutability
 
 - `final` only works on **methods and classes**, NOT on fields — `final var` is a compile error
+- Use `const` for never-reassigned fields instead: `private const _items:Array = [];` — assign at the declaration, or once in the constructor
+- `const` is shallow — the binding is fixed, the object's contents stay mutable
 
-### `as` operator with primitive types
+## `as` operator with primitive types
 
 - `context.call(...) as int` works **only** if the native extension returns `FRENewObjectFromInt32` (value arrives as a boxed `int`). If the C++ side returns `FRENewObjectFromDouble`, `as int` silently returns `null`, which coerces to `0`.
 - Prefer explicit cast: `int(context.call(...))` — works regardless of whether the native side returns `int`, `uint`, `Number`, or `double`.