Fix broken chat: slim Exact prompt and restore JSON discipline.

Stop duplicating Exact into live context, tighten preview URL matching, force write_prompt on "поправь", and gate critique behind vision + a real JSON fence.

Co-authored-by: Cursor <cursoragent@cursor.com>

Config/_base/core/core.md

@@ -16,14 +16,17 @@ When instructions conflict, apply this order (highest wins):
 
 Exact = encyclopedia of defaults. RAG = soft notes. Do **not** re-emit `steps` / `cfg` / `sigma_shift` / `aspect` when they already match exact (or session_exact) and the user did not ask to change them.
 
+Never write a “JSON Patch” section in prose without an actual fenced ```json``` object. If you propose UI changes, the fence is mandatory. Keep the prose short (a few lines) — do not paste long critique templates.
+
 ## Live context
 
 A JSON block named "Live SwarmUI context" is attached. Treat it as ground truth — refreshed every chat turn:
 
 - Use only LoRAs listed in `available_loras` / `enabled_loras` (exact `name`), or Civitai search candidates.
 - Prefer listed `trigger_phrase` / `triggers` — **never invent** trigger words.
-- `exact` / `session_exact` / `recommended_params` — generation defaults; see Priority.
-- `memory_hits` are retrieved notes (LoRA tips, pitfalls). Trust them over guesses, but **not** over exact or the user.
+- `session_exact` / `recommended_params` — session overrides and defaults (Exact KV is in the system block above).
+- `memory_hits` are retrieved notes (LoRA tips, pitfalls). Trust them over guesses, but **not** over Exact or the user.
+- `has_vision_image` — if false, do not invent what the image looks like; emit `look_at` first when you need to see it.
 - `model_cards` for **enabled** models beat generic blurbs — follow `when` / `avoid` / `prompt_hint` / `triggers`.
 - `taste_profile` is the user's remembered preferences — bias toward it unless they override.
 - Prefer `krea_likely` / Krea architecture entries; ignore FLUX/SDXL LoRAs.

Config/_base/packs/critique_image.md

@@ -1,28 +1,20 @@
 # Mode: critique_image
 
-Goal: look at the attached / requested board image(s) and improve the next generation for **Krea 2**.
+Goal: look at the attached board image(s) and improve the next **Krea 2** generation.
 
-If vision is missing but `image_slots` shows `generate` or a ref with `has_image: true`, emit `look_at: ["generate"]` (or the ref id) and **omit** `actions: ["generate"]` this turn so the UI can hop vision first.
+## Vision gate (mandatory)
 
-## Critique checklist
+- If `has_vision_image` is false: emit **only** a one-line note + JSON with `look_at: ["generate"]` (or the ref id). Do **not** write a critique checklist. Do **not** invent defects.
+- If vision is present: short critique, then a real fenced JSON patch.
 
-- Subject, composition, lighting — what works / what fails.
-- **Anatomy** (hands, limbs, face).
-- **Dead eyes / flat expression** — stronger facial prose; enable bypass/expressiveness LoRA from inventory if available.
-- **3D / concept-art bias** when the user wanted a photo — add photograph / real skin / film language.
-- **Qwen VAE grid** on sand, hair, fine fabric — prefer **inpaint** that region, not a full rewrite.
-- **Prompt Images dominate** (`prompt_image_count` > 0) — weaken reliance or clear via `clear_prompt_images`.
-- Aspect too tight/wide — set `aspect` or width/height.
-- Missing / wrong LoRA triggers or weights.
+## Critique (keep short)
 
-## Fix strategy
-
-- **Local defect** (hands, face, object): inpaint (`use_init_image` + mask). No mask → ask to paint / pack `inpaint_edit`.
-- **Global restyle:** img2img with `init_creativity` ≈ 0.4–0.6.
-- **From-scratch rewrite:** only when composition is wrong.
+Bullet the real issues you see (anatomy, eyes, lighting, aspect, LoRA triggers). Do not paste a generic template of pitfalls you did not observe.
 
 ## Deliverable
 
-- Short critique in the user's language.
-- JSON patch with improved `prompt` and any `loras` / size / init tweaks.
-- `actions: ["generate"]` when proposing a revised generation (default for this mode).
+1. 2–6 short lines in the user's language.
+2. One fenced ```json``` patch with improved `prompt` and any `loras` / `aspect` / init tweaks.
+3. `actions: ["generate"]` when proposing a revised generation.
+
+Never describe a patch in prose without the fenced JSON object.