2.4.1: model picker - honest sizing, nothing hidden, upgrade-break detection

CHANGELOG.md

@@ -2,6 +2,21 @@
 
 ## Unreleased
 
+## 2.4.1
+
+**Model picker fixes.** The Models card is one card again, the picker follows the app's styling, no model is hidden from it, and an active model invalidated by a FastFlowLM upgrade is now called out instead of failing silently.
+
+### Fixed
+
+- **No model is hidden from the picker any more.** Suggestions used to drop every candidate the size heuristic judged unfit, which made real catalog models invisible — `qwen3.6-moe:35b-a3b` never appeared and had to be pulled by hand. Every model FastFlowLM offers is now listed; oversized ones are shown with a "too large" marker and still ask for confirmation before downloading.
+- **Mixture-of-Experts models are sized correctly.** `…:35b-a3b` was read as a 35B model when only ~3B parameters are active per token. Sizing now prefers FastFlowLM's own measured **footprint** (so `gemma4-it:e4b`, which reads as "4B" by name but is really 8B / 9.1 GB, is also judged correctly) and falls back to active-parameter counts for MoE tags. Fit labels now show the real memory cost, e.g. "~24.3 GB needs most of ~25.6 GB usable".
+- **A model broken by a FastFlowLM upgrade is now reported.** FLM 0.9.45 rejects models whose local weights were stamped for an older version and reports them as not installed; nothing surfaced that, so the next hotkey failed with an opaque provider error. Config → Models now shows a warning naming the model and the remedy.
+
+### Changed
+
+- **"Installed models" and "Pull a new model" are a single "Models" card** with `Installed` / `Add a model` / `FastFlowLM runtime` sections, instead of two separate cards.
+- **The model picker is styled by the app.** It replaced a native `<datalist>` dropdown and a native `<select size>` list box — browsers draw datalist popups in their own chrome and ignore page CSS entirely, so that control could never match the dashboard. The new combobox and list are ordinary elements: they follow the light/dark theme, and support arrow-key/Enter/Escape navigation and type-to-filter.
+
 ## 2.4.0
 
 **Streaming chat.** The dashboard Chat tab renders replies token-by-token instead of waiting for the whole answer, so the response starts appearing at warm time-to-first-token rather than after the full completion.

README.md

@@ -4,10 +4,11 @@ Flowkey is a Windows desktop assistant that adds local-LLM hotkeys for grammar f
 
 Everything runs locally through [FastFlowLM](https://fastflowlm.com) (AMD Ryzen AI NPU) or, on machines without the NPU, through [Ollama](https://ollama.com) (CPU/GPU) as a secondary provider. No cloud service, analytics, or telemetry is used by the app.
 
-Current version: `2.4.0`
+Current version: `2.4.1`
 
 ## What's new in 2.4
 
+- **The model picker shows every model, sized honestly.** Config → **Models** is now one card, styled like the rest of the dashboard (the old picker used a native browser dropdown that ignored the app's theme). Nothing is hidden from it: models too big for your machine are marked rather than silently dropped, and Mixture-of-Experts models are sized by the parameters actually active per token plus FastFlowLM's own measured memory footprint — so a 35B-total / 3B-active model is no longer mistaken for a 35B one. If a FastFlowLM upgrade invalidates your active model, the card now says so instead of letting the next hotkey fail with an opaque error.
 - **Chat streams as it thinks.** The dashboard **Chat** tab now fills the reply in token-by-token over a live stream instead of waiting on a blank "Thinking…" until the whole answer is ready — the first words typically appear in ~1.6 s (warm) rather than after the full response. It works on both FastFlowLM and Ollama, and if anything interrupts the stream the partial reply is kept and the error is shown. Grammar, `prompt:`, and the other hotkeys are unchanged (they still paste the finished result).
 
 ## What's new in 2.3

SPEC.md

@@ -89,6 +89,9 @@ Caveman-encoded (compression, not amputation). Paths / ids / action names / numb
 - V35: default v2 → exactly 1 short LLM draft call; ⊥ anti-echo/rescue calls; V33 finalizer supplies surfaced 4-section contract
 - V36: pre-2.3 prompt_builder identity cfg (⊥ `prompt_version`, legacy default fields) → migrate v2+concise; any custom field → preserve
 - V37: `chat_send_stream` streams reply deltas as `text/event-stream` (SSE `data:`/`event:` frames), persists the full-or-partial turn exactly once under the daemon write-lock; CSRF `X-FFP-API` + Host gates still apply; dashboard falls back to `chat_send` iff the stream ⊥ opens (never after tokens arrive); `send()`/AHK/grammar/prompt paths unchanged (⊥ streamed)
+- V38: `recommend_models` returns EVERY candidate tagged yes/tight/no/unknown + `fit_reason`; UI ⊥ hide any (oversized → marked + confirm-before-pull). Fit signal precedence: provider `footprint_gb` → effective params (MoE active `a<N>b` < total, else total) vs budget
+- V39: configured active model ⊥ installed → `model_recommendations.active_model.status="not_installed"` + dashboard warning naming model/remedy; listing error/exception → `"unknown"` (⊥ claim broken)
+- V40: model picker + installed list = app-styled elements (⊥ native `<datalist>`/`<select size>`, ∵ browser chrome ignores page CSS); theme-aware + keyboard-navigable (↑↓/Enter/Esc)
 
 ## §T tasks
 
@@ -121,6 +124,7 @@ T24|x|prompt-v2 default + v1 rollback selector + 240/320/420 caps|V24,V27,V28,V3
 T25|x|FastFlowLM startup+idle keep-warm + cold/warm measurement support|V31
 T26|x|2.3.0 release evidence + version/docs rebaseline after A/B gate passes|V18,V29,V33,V34,V35,V36
 T27|x|streaming Chat tab (SSE): `ffp_chat.stream_send` + daemon `_STREAM_ACTIONS`/`_sse_frame`/`_stream_action` + `app.js` fetch-reader w/ `chat_send` fallback; live FLM first-token 1.58s|V37
+T28|x|model picker 2.4.1: footprint/MoE sizing + never-hide + active-model health + merged single Models card + app-styled combobox|V38,V39,V40
 ```
 
 ## §B bugs
@@ -155,4 +159,10 @@ B25|2026-07-10|broad V35 patch changed runtime/system branches instead of retry
 B26|2026-07-10|default v2 input ≥ routing threshold still made compression subcalls; 1 huge clause could exceed V27|V27,V35; bypass routing + bound grounded sections
 B27|2026-07-10|upgraded 2.2 cfg retained `detail_level=balanced` ∴ v1 selector missed new concise-default identity|V24,V30; `prompt_version=v1` authoritative legacy XML path
 B28|2026-07-10|2.2 persisted identity cfg lacked `prompt_version` + kept `balanced` ∴ upgrade bypassed v2 path|V36; narrow legacy-identity migration
+B29|2026-07-27|`qwen3.6-moe:35b-a3b` invisible in picker: `parse_params_b` read MoE total 35B (⊥ 3B active) → `fits=no` → `app.js` `continue` silently dropped it ∴ manual `flm pull`|V38; footprint-first sizing + MoE active parse + never-hide (marked + confirm)
+B30|2026-07-27|FLM 0.9.45 invalidated locally-pulled `qwen3.5:4b` (stamped 0.9.43) → FLM reports ⊥ installed; ⊥ detection ∴ next hotkey failed opaquely (Ollama absent ∴ ⊥ fallback)|V39; `_active_model_health` + dashboard warning
+B31|2026-07-27|model picker ⊥ followed app styling ∵ native `<datalist>` popup drawn by browser chrome (page CSS inert) + native `<select size>`|V40; custom app-styled combobox + list
+B32|2026-07-27|`app.js` set `models-title` = "Installed models — <provider>" ∴ overwrote merged card heading at runtime|V40 caught in browser; title → "Models — <provider>"
+B33|2026-07-27|new CSS block referenced undefined `--muted`/`--card`/`--bad` (pre-existing pattern elsewhere in stylesheet: 6+2+3 refs, never defined ∴ inert)|use defined tokens `--text-muted`/`--surface`/`--warn`; pre-existing refs left for separate cleanup
+B34|2026-07-27|self-caught: `_active_model_health` tested membership vs `_provider_list("all")`; ollama "all" = installed+suggested (`ffp_provider_runtime:80`) ∴ never-pulled model → false `installed=True` (⊥ warn)|V39; trust `details` (unfiltered, authoritative) else re-list w/ `installed` filter
 ```

docs/prompt-mode-speed-quality-plan.md

@@ -157,10 +157,39 @@ like the 2.2.0 prompt-builder eval.
 - **Phase 4 (optional) — Faster route. ⏸ TRIGGER NOT MET — deferred.** The plan
   gates this on "p50 still > 15 s after phases 1–2," but 2.3.0 put prompt p50 at
   **3.38 s warm** and Phase 3 streaming already delivers the *perceived* speed
-  this phase was chasing. A second NPU-decode provider (LM Studio/Lemonade) also
-  can't co-serve the NPU with FastFlowLM (both wedge — see the benchmark
-  evidence), so it would add real complexity for no measured need. Revisit only
-  if a concrete slow surface appears.
+  this phase was chasing. Revisit only if a concrete slow surface appears.
+
+  **Feasibility note (2026-07-27, FLM 0.9.45).** The original blocker recorded
+  here — "a second runtime can't co-serve the NPU with FastFlowLM (both wedge)" —
+  applies to another *NPU* runtime (Lemonade/LM Studio on the NPU). It does **not**
+  apply to an NPU + **discrete/iGPU** split, which is different hardware with no
+  NPU contention. FLM 0.9.45 adds **per-round KV cache checks**: a conversation
+  can alternate FLM-on-NPU and llama.cpp-on-GPU rounds and the NPU now recognises
+  its own already-cached earlier rounds instead of re-prefilling the whole
+  history from scratch. That makes a hybrid route genuinely practical for the
+  first time. It changes the *feasibility*, not the decision — the speed goal is
+  already met, so this stays deferred rather than becoming newly attractive.
+
+### Does FLM 0.9.45's KV cache work change anything today? (assessed 2026-07-27)
+
+**For prompt mode and the other hotkeys: no.** Per-round KV cache checks pay off
+only when rounds of *one* conversation alternate between backends. Flowkey never
+does that:
+
+- `prompt:`, grammar, summarize, explain, tone are **single-shot and stateless** —
+  one system+user pair, no prior rounds to reuse. `prompt:` is also decode-bound
+  (TTFT ≈1.5 s of a ≈3.4 s total), so prefill is not the cost.
+- **Chat** is multi-turn but always replays its 12-turn window to a *single*
+  backend — ordinary sequential prefix caching, which FLM already did. The new
+  gap-tolerance never triggers.
+- **Provider fallback** is all-or-nothing between FastFlowLM and Ollama, which are
+  separate servers with separate caches — nothing is shared to reuse.
+- **Meetings digests** are one large single-shot prefill.
+
+So there is nothing to change in the prompt pipeline for 0.9.45. The upgrade's
+real impact on users was operational, not architectural: it **invalidates
+already-pulled models** stamped for an older FLM (see 2.4.1 / SPEC V39, which
+added detection for exactly that).
 
 ## Integration with the existing prompt-builder (2.2.0)
 

installer/installer.iss

@@ -41,7 +41,7 @@
 #define AppURL        "https://github.com/agr77one/Fastflow"
 #define AppExeName    "Flowkey.exe"  ; symbolic — actual launchers below
 ; Keep in lockstep with scripts\_version.py.
-#define AppVersion    "2.4.0"
+#define AppVersion    "2.4.1"
 
 [Setup]
 AppId={{8A4F1E6C-9B3D-4E62-9F7A-FASTFLOW140}}

installer/sign.ps1

@@ -52,7 +52,7 @@
 .EXAMPLE
     # Sign the installer
     $env:FFP_SIGN_PFX_PASSWORD = "ChangeMe!"
-    .\sign.ps1 -FilePath ..\out\Flowkey-Setup-2.4.0.exe
+    .\sign.ps1 -FilePath ..\out\Flowkey-Setup-2.4.1.exe
 #>
 
 [CmdletBinding(DefaultParameterSetName = "Sign")]

pyproject.toml

@@ -17,7 +17,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "fastflowprompt"
-version = "2.4.0"
+version = "2.4.1"
 description = "Local-LLM-powered grammar fix, prompt rewrite, chat, and dashboard for Windows."
 readme = "README.md"
 requires-python = ">=3.11"

scripts/_version.py

@@ -1,3 +1,3 @@
 """Single source of truth for the app version. Read by grammar_fix.py."""
 
-__version__ = "2.4.0"
+__version__ = "2.4.1"

scripts/ffp_daemon.py

@@ -340,16 +340,92 @@ def _act_models_not_installed(_args: dict) -> dict:
 
 def _act_model_recommendations(_args: dict) -> dict:
     """Hardware-aware model suggestions for the active provider: detected
-    RAM/VRAM, a size budget, and candidate models tagged fits yes/tight/no."""
+    RAM/VRAM, a size budget, and candidate models tagged fits yes/tight/no.
+
+    On FastFlowLM the catalog's own per-model metadata (measured footprint,
+    install state, minimum FLM version) is forwarded so fit is decided from real
+    numbers rather than guessed from the tag name."""
     import ffp_hardware
     provider = grammar_fix.LLM_PROVIDER
     if provider == "ollama":
-        candidates = list(ffp_hardware.OLLAMA_CATALOG)
+        candidates: list = list(ffp_hardware.OLLAMA_CATALOG)
     else:
         listing = grammar_fix._provider_list("all")
-        names = listing.get("models") or []
-        candidates = [(n, ffp_hardware.parse_params_b(n)) for n in names]
-    return ffp_hardware.recommend_models(provider, candidates)
+        details = listing.get("details") or []
+        if details:
+            candidates = [dict(entry) for entry in details if isinstance(entry, dict)]
+        else:
+            # Older FLM without per-model JSON metadata: fall back to tag parsing.
+            candidates = [(n, ffp_hardware.parse_params_b(n)) for n in (listing.get("models") or [])]
+    out = ffp_hardware.recommend_models(provider, candidates)
+    out["active_model"] = _active_model_health(listing if provider != "ollama" else None)
+    return out
+
+
+def _active_model_health(listing: dict | None = None) -> dict:
+    """Report whether the configured active model is actually usable.
+
+    A FastFlowLM upgrade can invalidate an already-pulled model (FLM 0.9.45
+    rejects models stamped for 0.9.43), and FLM then reports them as
+    NOT installed. Nothing used to surface that, so the next hotkey failed with
+    an opaque provider error. This gives the dashboard a specific, actionable
+    state: which model is configured, whether it is installed, and the remedy.
+
+    ``listing`` may be a already-fetched provider listing to avoid a second CLI
+    call; it is only trusted when it carries per-model ``details`` (whose
+    ``installed`` flags are authoritative and filter-independent). Without
+    details we must re-list with the ``installed`` filter, because an
+    ``all`` listing includes not-yet-pulled suggestions and would report a
+    missing model as installed."""
+    provider = grammar_fix.LLM_PROVIDER
+    model = str(grammar_fix.FLM_MODEL or "")
+    health = {"provider": provider, "model": model, "installed": None, "status": "unknown", "detail": ""}
+    if not model:
+        health.update(status="unknown", detail="no active model configured")
+        return health
+
+    details = list((listing or {}).get("details") or [])
+    if not details:
+        try:
+            listing = grammar_fix._provider_list("installed")
+        except Exception as exc:
+            health.update(status="unknown", detail=f"could not list models: {exc}")
+            return health
+        if listing.get("error"):
+            health.update(status="unknown", detail=str(listing.get("error")))
+            return health
+        details = list(listing.get("details") or [])
+    elif listing and listing.get("error"):
+        health.update(status="unknown", detail=str(listing.get("error")))
+        return health
+
+    entry = next((d for d in details if isinstance(d, dict) and d.get("name") == model), None)
+    if entry is not None:
+        installed = bool(entry.get("installed"))
+        health["installed"] = installed
+        health["flm_min_version"] = entry.get("flm_min_version") or ""
+        if installed:
+            health.update(status="ok", detail="")
+        else:
+            health.update(
+                status="not_installed",
+                detail=f"'{model}' is in the catalog but not installed for this "
+                       f"{provider} version - run a pull to (re)install it.",
+            )
+        return health
+
+    # No per-model metadata (Ollama, or an older FLM): the names here are
+    # installed-only, so absence genuinely means "not installed".
+    names = set((listing or {}).get("models") or [])
+    if model in names:
+        health.update(installed=True, status="ok", detail="")
+    else:
+        health.update(
+            installed=False,
+            status="not_installed",
+            detail=f"'{model}' is not installed for {provider} - pull it or pick another model.",
+        )
+    return health
 
 
 def _act_pull_model(args: dict) -> str:

scripts/ffp_flm_server.py

@@ -252,6 +252,14 @@ def flm_list(filter_kind: str, model: str, no_window: int) -> dict:
     bare names. The JSON mode (FLM >= 0.9.x) carries an authoritative per-model
     `installed` boolean instead, so we parse that and filter client-side. We
     force UTF-8 decoding because the human-readable variant embeds emoji.
+
+    Also returns ``details``: the per-model metadata the catalog already carries
+    — ``footprint_gb`` (FLM's own measured memory cost, the honest sizing signal
+    that name-parsing cannot supply: ``gemma4-it:e4b`` reads as "4B" but is 8B /
+    9.1 GB, and ``qwen3.6-moe:35b-a3b`` is 35B total / 3B active / 24.3 GB),
+    ``installed``, ``flm_min_version``, ``parameter_size`` and
+    ``context_length``. ``details`` is unfiltered so callers can tell "not in
+    the catalog" from "in the catalog but not installed".
     """
     if filter_kind not in {"installed", "not-installed", "all"}:
         return {"error": f"bad filter: {filter_kind}", "models": [], "active": model}
@@ -276,19 +284,36 @@ def flm_list(filter_kind: str, model: str, no_window: int) -> dict:
         return {"error": "could not parse 'flm list --json' output", "models": [], "active": model}
 
     names: list[str] = []
+    details: list[dict] = []
     for entry in payload.get("models") or []:
         if not isinstance(entry, dict):
             continue
         name = str(entry.get("model") or entry.get("name") or "").strip()
         if not name:
             continue
         is_installed = bool(entry.get("installed"))
+        info = entry.get("details") if isinstance(entry.get("details"), dict) else {}
+        details.append({
+            "name": name,
+            "installed": is_installed,
+            "footprint_gb": _as_float(entry.get("footprint")),
+            "flm_min_version": str(entry.get("flm_min_version") or "").strip(),
+            "parameter_size": str(info.get("parameter_size") or "").strip(),
+            "context_length": entry.get("default_context_length"),
+        })
         if filter_kind == "installed" and not is_installed:
             continue
         if filter_kind == "not-installed" and is_installed:
             continue
         names.append(name)
-    return {"models": names, "active": model}
+    return {"models": names, "active": model, "details": details}
+
+
+def _as_float(value: object) -> float | None:
+    try:
+        return float(value)  # type: ignore[arg-type]
+    except (TypeError, ValueError):
+        return None
 
 
 def flm_version(no_window: int) -> str: