minima-cli 0.7.0__tar.gz → 0.7.1__tar.gz

minima_cli-0.7.1/.github/workflows/catalog-snapshot.yml

@@ -0,0 +1,100 @@
+name: Catalog — refresh vendored snapshot
+
+# Keeps the vendored cold-start model catalog fresh so it stops drifting by hand.
+#
+# WHAT/WHY: src/minima/catalog/data/capability_priors.json is the cold-start
+# fallback the server loads when the live cost feed (LiteLLM/OpenRouter) is
+# unreachable — see src/minima/catalog/store.py (load_snapshot_cards) and the
+# /v1/models router. It used to be hand-maintained and drifted stale (e.g.
+# missing the gpt-5 family). This job pulls the *live* served catalog from
+# api.minima.sh/v1/models — the source-of-truth that already merges curated
+# capability priors with live cost — normalizes it (via jq) into the EXACT
+# shape the loader reads, and opens a PR only when it actually changed.
+#
+# We deliberately DO NOT commit straight to main and DO NOT hand-edit the JSON:
+# a reviewer sees the diff, and the automation (not a human) keeps it current.
+on:
+  schedule:
+    - cron: "17 6 * * *" # daily (~every 24h); minute offset avoids the top-of-hour rush
+  workflow_dispatch: {}
+
+concurrency:
+  group: catalog-snapshot
+  cancel-in-progress: false
+
+# Least privilege: only what's needed to push a chore branch and open a PR.
+permissions:
+  contents: write
+  pull-requests: write
+
+env:
+  # Source-of-truth: the live served catalog. Override via repo/environment var
+  # CATALOG_SOURCE_URL if the prod host ever changes.
+  CATALOG_URL: ${{ vars.CATALOG_SOURCE_URL || 'https://api.minima.sh/v1/models' }}
+  SNAPSHOT_PATH: src/minima/catalog/data/capability_priors.json
+  BRANCH: chore/catalog-snapshot-auto
+
+jobs:
+  refresh:
+    name: Fetch live catalog → PR if changed
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v5
+
+      - name: Fetch + normalize live catalog into the vendored snapshot shape
+        run: |
+          set -euo pipefail
+          # /v1/models returns {models:[ModelCard...], catalog_version, stale}.
+          # ModelCard is a superset of the snapshot object, so we project it back
+          # to the exact fields load_snapshot_cards() reads and drop the runtime
+          # metadata (cost_source/cost_fetched_at/cost_stale/capability_source),
+          # which the loader always re-stamps as "fallback-snapshot".
+          curl -fsSL --retry 3 --retry-delay 5 --max-time 60 "$CATALOG_URL" -o live.json
+
+          count="$(jq '.models | length' live.json)"
+          if [ "$count" -lt 5 ]; then
+            echo "::error::live catalog returned only $count models (<5); refusing to overwrite snapshot"
+            exit 1
+          fi
+
+          version="fallback-snapshot-$(date -u +%Y-%m)"
+          note="Cold-start fallback only. Auto-refreshed by .github/workflows/catalog-snapshot.yml from the live /v1/models catalog — do not hand-edit. Live cost is overlaid from LiteLLM/OpenRouter at runtime; capability_by_task_type are normalized [0,1] heuristic priors, corrected by accumulated memory. Prices are approximate USD per million tokens."
+
+          jq \
+             --arg version "$version" \
+             --arg note "$note" '{
+               version: $version,
+               note: $note,
+               models: (.models | sort_by(.model_id) | map({
+                 model_id, provider, display_name,
+                 input_cost_per_mtok, output_cost_per_mtok, cache_read_cost_per_mtok,
+                 supports_prompt_caching, context_window, max_output_tokens,
+                 capability_priors, capability_by_task_type
+               }))
+             }' live.json > "$SNAPSHOT_PATH.tmp"
+          mv "$SNAPSHOT_PATH.tmp" "$SNAPSHOT_PATH"
+          rm -f live.json
+          echo "Normalized $count models into $SNAPSHOT_PATH"
+
+      - name: Commit + open PR only if the snapshot changed
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          set -euo pipefail
+          if git diff --quiet -- "$SNAPSHOT_PATH"; then
+            echo "Snapshot already up to date; nothing to do."
+            exit 0
+          fi
+
+          git config user.name "github-actions[bot]"
+          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
+          git checkout -B "$BRANCH"
+          git add "$SNAPSHOT_PATH"
+          git commit -m "chore(catalog): refresh vendored model snapshot from live /v1/models"
+          git push -u origin "$BRANCH" --force-with-lease
+
+          gh pr create --base main --head "$BRANCH" \
+            --title "chore(catalog): refresh vendored model snapshot" \
+            --body "Automated refresh of \`$SNAPSHOT_PATH\` from the live catalog at \`$CATALOG_URL\` (normalized to the loader's shape). Opened by \`.github/workflows/catalog-snapshot.yml\`. Review the diff — this is the cold-start fallback the server serves when the live cost feed is unreachable." \
+            || echo "PR already open for $BRANCH; pushed update."

minima_cli-0.7.1/.github/workflows/ci.yml

@@ -0,0 +1,69 @@
+name: CI — PR checks
+
+# Gate on every PR to main (and on push to main) with the exact local test
+# conventions: server (ruff + offline pytest) and TUI (bun test + tsc). Two
+# independent jobs so a red on one surfaces without masking the other. No
+# deploy, no secrets — the suites are offline (`-m "not live and not eval"`).
+# mypy runs after ruff/pytest: src/minima is now type-clean, so it gates too.
+on:
+  pull_request:
+    branches: [main]
+  push:
+    branches: [main]
+
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+env:
+  BUN_VERSION: "1.3.14"
+
+jobs:
+  server:
+    name: Server — ruff + pytest + mypy
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v5
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v6
+
+      - name: Sync deps (dev extra)
+        run: uv sync --extra dev
+
+      - name: Ruff
+        run: uv run ruff check src client_sdk tests
+
+      - name: Pytest (offline)
+        run: uv run pytest -m "not live and not eval" -q
+
+      - name: Mypy
+        run: uv run mypy src/minima
+
+  tui:
+    name: TUI — bun test + tsc
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: packages/tui
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v5
+
+      - name: Set up Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: ${{ env.BUN_VERSION }}
+
+      - name: Install deps
+        run: bun install --frozen-lockfile
+
+      - name: Test
+        run: bun test
+
+      - name: Typecheck
+        run: bun run check

{minima_cli-0.7.0 → minima_cli-0.7.1}/CHANGELOG.md

@@ -4,6 +4,37 @@ All notable changes to Minima are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/), and the project aims to follow
 [Semantic Versioning](https://semver.org/).
 
+## [0.7.1] - 2026-07-05
+
+### Changed (API behavior — note for integrators)
+- **`max_cost_per_call` is now a true hard filter.** When no model fits the budget,
+  `POST /v1/recommend` returns **422** (`"no model within max_cost_per_call budget"`)
+  instead of 200 with the cheapest over-budget model plus a `no_model_within_cost_budget`
+  warning. This matches the documented "hard filter" contract and the existing
+  no-candidates behavior. Callers who want "cheapest possible regardless of ceiling"
+  should use the cost-quality slider, not `max_cost_per_call`.
+- **Auth rejects malformed bearer tokens.** A bearer token that is not in the Mubit key
+  format (`mbt_…`) now returns **401** before any work, instead of being accepted and
+  served from priors. A missing bearer with a server-configured `MUBIT_API_KEY` fallback
+  is unaffected.
+
+### Fixed
+- **Cold-start catalog prices were stale.** The vendored fallback snapshot
+  (`capability_priors.json`) is refreshed from the live catalog (e.g. `claude-opus-4-8`
+  `15/75` → `5/25`), and a scheduled workflow now keeps it current so it stops drifting
+  by hand. (The auto-refresh updates prices for the curated model list; it does not add
+  new models — expanded provider coverage such as the gpt-5 family remains a separate
+  curation.)
+- **Server version string can no longer drift.** `minima.__version__` (reported by
+  `/v1/health` and `/v1/capabilities`) now derives from installed package metadata instead
+  of a hardcoded constant that had silently lagged across releases.
+- **Feedback reconciliation types the quality value honestly** as `float | None` — unjudged
+  rows keep `NULL`, no fabricated default (preserves the M-J2 fix).
+
+### Internal
+- **PR CI added** — every PR now runs server (`ruff` + `pytest` + `mypy`) and TUI
+  (`bun test` + `tsc`) checks; the recommender is mypy-clean.
+
 ## [0.7.0] - 2026-07-05
 
 ### Added

{minima_cli-0.7.0 → minima_cli-0.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: minima-cli
-Version: 0.7.0
+Version: 0.7.1
 Summary: Minima CLI: cost-aware LLM model routing — recommend cheaper models, backed by Mubit memory.
 Project-URL: Homepage, https://docs.minima.sh
 Project-URL: Documentation, https://docs.minima.sh

{minima_cli-0.7.0 → minima_cli-0.7.1}/packages/tui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "minima-tui",
-  "version": "0.7.0",
+  "version": "0.7.1",
   "description": "Minima harness + TUI — the `minima` CLI (TypeScript/Ink; successor to the removed Python harness). Talks to the Python FastAPI recommender service over the /v1/* contract.",
   "type": "module",
   "license": "FSL-1.1-Apache-2.0",

{minima_cli-0.7.0 → minima_cli-0.7.1}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "hatchling.build"
 # package stays `minima`. This package ships the server + client SDK only; the `minima` CLI
 # is the TypeScript binary installed via Homebrew (brew tap mubit-ai/minima).
 name = "minima-cli"
-version = "0.7.0"
+version = "0.7.1"
 description = "Minima CLI: cost-aware LLM model routing — recommend cheaper models, backed by Mubit memory."
 readme = "README.md"
 requires-python = ">=3.11"

{minima_cli-0.7.0 → minima_cli-0.7.1}/src/minima/api/auth.py

@@ -20,6 +20,20 @@ def bearer_key(request: Request) -> str | None:
     return None
 
 
+def is_mubit_key_format(key: str) -> bool:
+    """Cheap client-side shape check for a canonical Mubit key.
+
+    Canonical keys are ``mbt_<instance>_...`` (see tenancy.passthrough._org_id). This is
+    a front-door format gate only — it rejects tokens that clearly are not Mubit keys
+    (e.g. ``not-a-key``) before any work is done, without a Mubit round-trip. It does NOT
+    authenticate the key; a well-formed key still resolves normally (and works offline
+    from local priors). Kept deliberately permissive to avoid locking out real callers:
+    the only requirement is an ``mbt_`` prefix followed by a non-empty instance segment.
+    """
+    prefix, sep, rest = key.partition("_")
+    return prefix == "mbt" and sep == "_" and bool(rest)
+
+
 def configured_key(request: Request) -> str | None:
     settings = getattr(request.app.state, "settings", None)
     key = getattr(settings, "mubit_api_key", None)
@@ -27,7 +41,14 @@ def configured_key(request: Request) -> str | None:
 
 
 async def get_tenant(request: Request) -> TenantContext:
-    key = bearer_key(request) or configured_key(request)
+    token = bearer_key(request)
+    if token is not None and not is_mubit_key_format(token):
+        raise ApiError(
+            401,
+            "Unauthorized",
+            "Authorization bearer token is not a valid Mubit API key (expected 'mbt_...' format)",
+        )
+    key = token or configured_key(request)
     if not key:
         raise ApiError(
             401,
@@ -39,7 +60,10 @@ async def get_tenant(request: Request) -> TenantContext:
 
 async def get_tenant_optional(request: Request) -> TenantContext | None:
     """Like get_tenant but returns None instead of 401 (for health probes)."""
-    key = bearer_key(request) or configured_key(request)
+    token = bearer_key(request)
+    if token is not None and not is_mubit_key_format(token):
+        return None
+    key = token or configured_key(request)
     if not key:
         return None
     return request.app.state.passthrough_runtime.resolve(key)

minima_cli-0.7.1/src/minima/catalog/data/capability_priors.json

@@ -0,0 +1,330 @@
+{
+  "version": "fallback-snapshot-2026-07",
+  "note": "Cold-start fallback only.",
+  "models": [
+    {
+      "model_id": "claude-haiku-4-5",
+      "provider": "anthropic",
+      "display_name": "Claude Haiku 4.5",
+      "input_cost_per_mtok": 1.0,
+      "output_cost_per_mtok": 5.0,
+      "cache_read_cost_per_mtok": 0.09999999999999999,
+      "supports_prompt_caching": true,
+      "context_window": 200000,
+      "max_output_tokens": 64000,
+      "capability_priors": {
+        "intelligence_index": 0.5
+      },
+      "capability_by_task_type": {
+        "code": 0.7,
+        "summarization": 0.85,
+        "extraction": 0.88,
+        "qa": 0.8,
+        "reasoning": 0.6,
+        "classification": 0.9,
+        "translation": 0.85,
+        "creative": 0.75,
+        "rag": 0.82,
+        "tool_use": 0.75,
+        "other": 0.72
+      }
+    },
+    {
+      "model_id": "claude-opus-4-8",
+      "provider": "anthropic",
+      "display_name": "Claude Opus 4.8",
+      "input_cost_per_mtok": 5.0,
+      "output_cost_per_mtok": 25.0,
+      "cache_read_cost_per_mtok": 0.5,
+      "supports_prompt_caching": true,
+      "context_window": 1000000,
+      "max_output_tokens": 64000,
+      "capability_priors": {
+        "intelligence_index": 0.95
+      },
+      "capability_by_task_type": {
+        "code": 0.96,
+        "summarization": 0.93,
+        "extraction": 0.94,
+        "qa": 0.95,
+        "reasoning": 0.97,
+        "classification": 0.95,
+        "translation": 0.93,
+        "creative": 0.94,
+        "rag": 0.94,
+        "tool_use": 0.95,
+        "other": 0.93
+      }
+    },
+    {
+      "model_id": "claude-sonnet-4-6",
+      "provider": "anthropic",
+      "display_name": "Claude Sonnet 4.6",
+      "input_cost_per_mtok": 3.0,
+      "output_cost_per_mtok": 15.0,
+      "cache_read_cost_per_mtok": 0.3,
+      "supports_prompt_caching": true,
+      "context_window": 1000000,
+      "max_output_tokens": 64000,
+      "capability_priors": {
+        "intelligence_index": 0.72
+      },
+      "capability_by_task_type": {
+        "code": 0.88,
+        "summarization": 0.9,
+        "extraction": 0.92,
+        "qa": 0.9,
+        "reasoning": 0.85,
+        "classification": 0.93,
+        "translation": 0.9,
+        "creative": 0.88,
+        "rag": 0.9,
+        "tool_use": 0.9,
+        "other": 0.86
+      }
+    },
+    {
+      "model_id": "gemini-2.5-flash",
+      "provider": "google",
+      "display_name": "Gemini 2.5 Flash",
+      "input_cost_per_mtok": 0.3,
+      "output_cost_per_mtok": 2.5,
+      "cache_read_cost_per_mtok": 0.03,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.55
+      },
+      "capability_by_task_type": {
+        "code": 0.74,
+        "summarization": 0.88,
+        "extraction": 0.88,
+        "qa": 0.82,
+        "reasoning": 0.66,
+        "classification": 0.9,
+        "translation": 0.88,
+        "creative": 0.78,
+        "rag": 0.86,
+        "tool_use": 0.8,
+        "other": 0.76
+      }
+    },
+    {
+      "model_id": "gemini-2.5-flash-lite",
+      "provider": "google",
+      "display_name": "Gemini 2.5 Flash-Lite",
+      "input_cost_per_mtok": 0.09999999999999999,
+      "output_cost_per_mtok": 0.39999999999999997,
+      "cache_read_cost_per_mtok": 0.01,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.4
+      },
+      "capability_by_task_type": {
+        "code": 0.62,
+        "summarization": 0.8,
+        "extraction": 0.84,
+        "qa": 0.74,
+        "reasoning": 0.5,
+        "classification": 0.86,
+        "translation": 0.82,
+        "creative": 0.68,
+        "rag": 0.78,
+        "tool_use": 0.68,
+        "other": 0.64
+      }
+    },
+    {
+      "model_id": "gemini-2.5-pro",
+      "provider": "google",
+      "display_name": "Gemini 2.5 Pro",
+      "input_cost_per_mtok": 1.25,
+      "output_cost_per_mtok": 10.0,
+      "cache_read_cost_per_mtok": 0.125,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.8
+      },
+      "capability_by_task_type": {
+        "code": 0.9,
+        "summarization": 0.92,
+        "extraction": 0.92,
+        "qa": 0.91,
+        "reasoning": 0.88,
+        "classification": 0.93,
+        "translation": 0.92,
+        "creative": 0.89,
+        "rag": 0.91,
+        "tool_use": 0.9,
+        "other": 0.88
+      }
+    },
+    {
+      "model_id": "gemini-3-flash-preview",
+      "provider": "google",
+      "display_name": "Gemini 3 Flash (Preview)",
+      "input_cost_per_mtok": 0.5,
+      "output_cost_per_mtok": 3.0,
+      "cache_read_cost_per_mtok": 0.049999999999999996,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.62
+      },
+      "capability_by_task_type": {
+        "code": 0.8,
+        "summarization": 0.87,
+        "extraction": 0.88,
+        "qa": 0.83,
+        "reasoning": 0.7,
+        "classification": 0.9,
+        "translation": 0.87,
+        "creative": 0.8,
+        "rag": 0.85,
+        "tool_use": 0.84,
+        "other": 0.78
+      }
+    },
+    {
+      "model_id": "gemini-3.1-flash-lite",
+      "provider": "google",
+      "display_name": "Gemini 3.1 Flash-Lite",
+      "input_cost_per_mtok": 0.25,
+      "output_cost_per_mtok": 1.5,
+      "cache_read_cost_per_mtok": 0.024999999999999998,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.48
+      },
+      "capability_by_task_type": {
+        "code": 0.68,
+        "summarization": 0.83,
+        "extraction": 0.86,
+        "qa": 0.78,
+        "reasoning": 0.58,
+        "classification": 0.88,
+        "translation": 0.84,
+        "creative": 0.72,
+        "rag": 0.8,
+        "tool_use": 0.72,
+        "other": 0.68
+      }
+    },
+    {
+      "model_id": "gemini-3.1-pro-preview",
+      "provider": "google",
+      "display_name": "Gemini 3.1 Pro (Preview)",
+      "input_cost_per_mtok": 2.0,
+      "output_cost_per_mtok": 12.0,
+      "cache_read_cost_per_mtok": 0.19999999999999998,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.9
+      },
+      "capability_by_task_type": {
+        "code": 0.93,
+        "summarization": 0.93,
+        "extraction": 0.93,
+        "qa": 0.93,
+        "reasoning": 0.95,
+        "classification": 0.94,
+        "translation": 0.93,
+        "creative": 0.91,
+        "rag": 0.93,
+        "tool_use": 0.92,
+        "other": 0.91
+      }
+    },
+    {
+      "model_id": "gemini-3.5-flash",
+      "provider": "google",
+      "display_name": "Gemini 3.5 Flash",
+      "input_cost_per_mtok": 1.5,
+      "output_cost_per_mtok": 9.0,
+      "cache_read_cost_per_mtok": 0.15,
+      "supports_prompt_caching": true,
+      "context_window": 1048576,
+      "max_output_tokens": 65536,
+      "capability_priors": {
+        "intelligence_index": 0.65
+      },
+      "capability_by_task_type": {
+        "code": 0.85,
+        "summarization": 0.88,
+        "extraction": 0.9,
+        "qa": 0.84,
+        "reasoning": 0.72,
+        "classification": 0.91,
+        "translation": 0.88,
+        "creative": 0.82,
+        "rag": 0.87,
+        "tool_use": 0.88,
+        "other": 0.8
+      }
+    },
+    {
+      "model_id": "gpt-4o",
+      "provider": "openai",
+      "display_name": "GPT-4o",
+      "input_cost_per_mtok": 2.5,
+      "output_cost_per_mtok": 10.0,
+      "cache_read_cost_per_mtok": 1.25,
+      "supports_prompt_caching": true,
+      "context_window": 128000,
+      "max_output_tokens": 16384,
+      "capability_priors": {
+        "intelligence_index": 0.7
+      },
+      "capability_by_task_type": {
+        "code": 0.85,
+        "summarization": 0.9,
+        "extraction": 0.9,
+        "qa": 0.89,
+        "reasoning": 0.82,
+        "classification": 0.92,
+        "translation": 0.9,
+        "creative": 0.87,
+        "rag": 0.89,
+        "tool_use": 0.89,
+        "other": 0.85
+      }
+    },
+    {
+      "model_id": "gpt-4o-mini",
+      "provider": "openai",
+      "display_name": "GPT-4o mini",
+      "input_cost_per_mtok": 0.15,
+      "output_cost_per_mtok": 0.6,
+      "cache_read_cost_per_mtok": 0.075,
+      "supports_prompt_caching": true,
+      "context_window": 128000,
+      "max_output_tokens": 16384,
+      "capability_priors": {
+        "intelligence_index": 0.45
+      },
+      "capability_by_task_type": {
+        "code": 0.68,
+        "summarization": 0.84,
+        "extraction": 0.86,
+        "qa": 0.78,
+        "reasoning": 0.55,
+        "classification": 0.9,
+        "translation": 0.83,
+        "creative": 0.72,
+        "rag": 0.8,
+        "tool_use": 0.72,
+        "other": 0.7
+      }
+    }
+  ]
+}

{minima_cli-0.7.0 → minima_cli-0.7.1}/src/minima/recommender/decisionlog.py

@@ -119,7 +119,10 @@ class Reconciliation:
 
     model_id: str
     outcome: str
-    quality: float
+    # None for UNJUDGED feedback (judged=False): never fabricate a quality for
+    # rows the harness didn't judge (M-J2). _apply stores it straight through to
+    # DecisionRecord.realized_quality, which is likewise `float | None` (NULL).
+    quality: float | None
     cost_usd: float | None = None
     latency_ms: int | None = None
     ts: float = 0.0

{minima_cli-0.7.0 → minima_cli-0.7.1}/src/minima/recommender/engine.py

@@ -108,11 +108,17 @@ class Recommender:
         warnings: list[str] = []
 
         classification = classify_details(req.task)
+        # classify_details always populates `profile`; bind it locally so the type
+        # checker sees a non-None ClassificationProfile (narrowing on the attribute
+        # is otherwise dropped across the awaits below). Mutations still apply to the
+        # same object, so behavior is unchanged.
+        class_profile = classification.profile
+        assert class_profile is not None
         task_type = classification.task_type
         difficulty = classification.difficulty
         profile.mark("classify")
-        classification.profile.final_task_type = task_type
-        classification.profile.final_difficulty = difficulty
+        class_profile.final_task_type = task_type
+        class_profile.final_difficulty = difficulty
         signature = (
             salient_signature(req.task.task, settings.minima_cluster_signature_tokens)
             if settings.minima_cluster_granularity.lower() == "fine"
@@ -152,16 +158,16 @@ class Recommender:
                 [(ev.record.task_type, ev.score) for ev in evidence if ev.record is not None]
             )
             if neighbor_estimate is not None:
-                classification.profile.neighbor_support = neighbor_estimate.neighbor_support
-                classification.profile.neighbor_count = neighbor_estimate.neighbor_count
+                class_profile.neighbor_support = neighbor_estimate.neighbor_support
+                class_profile.neighbor_count = neighbor_estimate.neighbor_count
                 if (
                     neighbor_estimate.task_type != task_type
-                    and classification.profile.caller_task_type is None
+                    and class_profile.caller_task_type is None
                     and task_type == TaskType.other
                 ):
                     task_type = neighbor_estimate.task_type
-                    classification.profile.task_type_source = "neighbor_vote"
-                    classification.profile.final_task_type = task_type
+                    class_profile.task_type_source = "neighbor_vote"
+                    class_profile.final_task_type = task_type
                 cluster = task_cluster(task_type, difficulty, signature)
                 warnings.append("neighbor_classified")
         profile.mark("neighbor_classify")
@@ -224,7 +230,11 @@ class Recommender:
             if affordable:
                 scored = affordable
             else:
-                warnings.append("no_model_within_cost_budget")
+                # max_cost_per_call is a hard filter: if no model fits the budget the
+                # eligible set is empty, which is the same "nothing fits" condition as an
+                # impossible candidate/exclusion set — surface it as the same 422 instead
+                # of serving an over-budget model.
+                raise NoCandidatesError("no model within max_cost_per_call budget")
 
         if req.constraints.max_latency_ms is not None:
             # Only exclude candidates with OBSERVED latency evidence above the budget —