a2a-source-edit: write agent.py

agent.py

@@ -1,189 +1,588 @@
-"""quote-judge-studio-v1 agent.
-
-Starter stack:
-  - DeepAgents for tool-calling orchestration
-  - Caller-provided LLM credentials via ctx.llm
-  - A tiny model-call middleware hook you can replace with tracing,
-    routing, rate limits, or policy checks
-"""
+"""QuoteJudge Studio v1: deterministic quote comparison with durable per-user state."""
 from __future__ import annotations
 
+import base64
+import csv
+import hashlib
+import io
 import json
-from pathlib import Path
-from typing import Any
+import os
+import re
+from datetime import datetime, timezone
+from decimal import Decimal, ROUND_HALF_UP
+from typing import Annotated, Any
 
-from pydantic import BaseModel
+from pydantic import BaseModel, Field, field_validator
 
 import a2a_pack as a2a
 from a2a_pack import (
     A2AAgent,
-    LLMProvisioning,
-    {{ auth_type }},
+    AgentDatabase,
+    AgentDatabaseEnv,
+    AgentDatabaseMigrations,
+    AgentPlatformResources,
+    PlatformUserAuth,
     Pricing,
+    Resources,
     RunContext,
+    State,
     WorkspaceAccess,
     WorkspaceMode,
 )
-from a2a_pack.context import LLMCreds
+from a2a_pack.workspace import FileUpload, UploadedFile
+
+try:  # psycopg is installed in the deployed image via requirements.txt.
+    import psycopg
+    from psycopg.rows import dict_row
+    from psycopg.types.json import Jsonb
+except Exception:  # pragma: no cover - lets `a2a card` run before deps install.
+    psycopg = None
+    dict_row = None
+    Jsonb = None
+
+
+COMPARISON_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9_.-]{0,79}$")
+MAX_QUOTES = 10
+MAX_BROWSER_DOCUMENTS = 4
+MAX_UPLOAD_BYTES = 64 * 1024
+ALLOWED_UPLOAD_MEDIA_TYPES = {
+    "application/json",
+    "text/json",
+    "text/csv",
+    "application/csv",
+    "application/vnd.ms-excel",
+    "text/plain",
+}
+
+# Local/dev fallback only. Hosted production uses DATABASE_URL and managed Postgres.
+_LOCAL_COMPARISONS: dict[tuple[str, str], dict[str, Any]] = {}
+_LOCAL_RECEIPTS: list[dict[str, Any]] = []
 
 
 class QuoteJudgeStudioV1Config(BaseModel):
-    pass
+    """Runtime config is intentionally empty; the app uses platform auth + DB."""
 
 
-SYSTEM_PROMPT = """\
-You are a compact tool-calling agent.
+class QuoteInput(BaseModel):
+    vendor: str = Field(..., min_length=1, max_length=120)
+    quantity: int = Field(..., ge=1, le=1_000_000)
+    unit_price: float = Field(..., ge=0, le=1_000_000_000)
+    delivery_days: int = Field(..., ge=0, le=3650)
+    warranty_months: int = Field(..., ge=0, le=240)
 
-Use the text_stats tool when the user asks about text, counts, summaries,
-or anything where exact length/word numbers would help. Mention tool results
-briefly instead of dumping raw JSON.
-"""
-
-RUNTIME_SKILLS_DIR = "quote-judge-studio-v1/.deepagents/skills/"
-DEEPAGENTS_RECURSION_LIMIT = 500
+    @field_validator("vendor")
+    @classmethod
+    def clean_vendor(cls, value: str) -> str:
+        cleaned = " ".join(str(value or "").split())
+        if not cleaned:
+            raise ValueError("vendor is required")
+        return cleaned
 
 
-class QuoteJudgeStudioV1(A2AAgent[QuoteJudgeStudioV1Config, {{ auth_type }}]):
+class QuoteWeights(BaseModel):
+    price: float = Field(..., ge=0, le=1000)
+    delivery: float = Field(..., ge=0, le=1000)
+    warranty: float = Field(..., ge=0, le=1000)
+
+    @field_validator("warranty")
+    @classmethod
+    def at_least_one_weight(cls, value: float, info: Any) -> float:
+        values = dict(info.data)
+        total = float(values.get("price") or 0) + float(values.get("delivery") or 0) + float(value or 0)
+        if total <= 0:
+            raise ValueError("at least one weight must be greater than zero")
+        return value
+
+
+class BrowserDocument(BaseModel):
+    filename: str = Field(..., min_length=1, max_length=160)
+    media_type: str = Field(..., min_length=1, max_length=120)
+    data_base64: str = Field(..., min_length=1, max_length=90_000)
+
+    @field_validator("filename")
+    @classmethod
+    def safe_filename(cls, value: str) -> str:
+        cleaned = str(value or "").strip().replace("\\", "/")
+        if not cleaned or "/" in cleaned or cleaned in {".", ".."}:
+            raise ValueError("filename must be a simple file name")
+        return cleaned
+
+    @field_validator("media_type")
+    @classmethod
+    def known_media_type(cls, value: str) -> str:
+        cleaned = str(value or "").split(";", 1)[0].strip().lower()
+        if cleaned not in ALLOWED_UPLOAD_MEDIA_TYPES:
+            raise ValueError("unsupported media type")
+        return cleaned
+
+
+class QuoteJudgeStudioV1(A2AAgent[QuoteJudgeStudioV1Config, PlatformUserAuth]):
     name = "quote-judge-studio-v1"
-    description = "QuoteJudge compares vendor quotes with weighted price, delivery, and warranty scoring, persists per-user comparisons, supports browser uploads, and serves a polished packed product UI."
+    description = (
+        "QuoteJudge compares vendor quotes with weighted price, delivery, and warranty "
+        "scoring, persists per-user comparisons, supports browser uploads, and serves "
+        "a polished packed product UI."
+    )
     version = "0.1.0"
 
     config_model = QuoteJudgeStudioV1Config
-    auth_model = {{ auth_type }}
+    auth_model = PlatformUserAuth
 
-    # Hosted generated agents read the caller's saved LLM credential through
-    # ctx.llm. The platform may proxy that credential through LiteLLM, but agent
-    # code never reads provider keys, LiteLLM master keys, or OPENAI_API_KEY
-    # directly.
-    llm_provisioning = LLMProvisioning.PLATFORM
+    state = State.DURABLE
     pricing = Pricing(
         price_per_call_usd=0.0,
-        caller_pays_llm=True,
-        notes="Starter agent uses the caller's saved LLM credential via ctx.llm.",
+        caller_pays_llm=False,
+        notes="Deterministic quote scoring; no LLM credential or provider key is required.",
     )
+    resources = Resources(cpu="500m", memory="512Mi", max_runtime_seconds=600)
+    tools_used = ("postgres", "mcp", "packed-react-frontend")
     workspace_access = WorkspaceAccess.dynamic(
-        max_files=64,
+        max_files=8,
         allowed_modes=(WorkspaceMode.READ_ONLY, WorkspaceMode.READ_WRITE_OVERLAY),
         require_reason=False,
+        max_total_size_bytes=MAX_UPLOAD_BYTES * MAX_BROWSER_DOCUMENTS,
+    )
+    platform_resources = AgentPlatformResources(
+        databases=(
+            AgentDatabase(
+                name="quote-judge-studio-v1-data",
+                scope="user",
+                access_mode="read_write",
+                env=AgentDatabaseEnv(url="DATABASE_URL"),
+                migrations=AgentDatabaseMigrations(path="db/migrations"),
+            ),
+        )
     )
-    tools_used = ("deepagents", "langchain")
 
-    @a2a.tool(description="Ask the starter DeepAgent to answer with tool calls when useful")
-    async def ask(self, ctx: RunContext[{{ auth_type }}], prompt: str) -> str:
-        creds = ctx.llm
-        await ctx.emit_progress(f"llm: {creds.model} via {creds.source}")
-        if not creds.api_key:
-            return (
-                "LLM key required. Add an LLM credential in Settings > LLM "
-                "credentials before running this agent; for local --invoke "
-                "runs set AGENT_LLM_KEY."
+    @a2a.tool(
+        description="Compare two to ten vendor quotes with weighted price, delivery, and warranty scoring; persist the result and a production receipt.",
+        timeout_seconds=30,
+        idempotent=True,
+        cost_class="deterministic",
     )
-        graph = self._build_deep_agent(ctx=ctx, creds=creds)
-        state = await graph.ainvoke(
-            {"messages": [{"role": "user", "content": prompt}]},
-            config={"recursion_limit": DEEPAGENTS_RECURSION_LIMIT},
-        )
-        await ctx.emit_progress("deepagent finished")
-        return _last_message_text(state)
-
-    def _build_deep_agent(
+    async def compare_quotes(
         self,
-        *,
-        ctx: RunContext[{{ auth_type }}],
-        creds: LLMCreds,
-    ) -> Any:
-        # Lazy imports keep `a2a card` usable before local dependencies are
-        # installed. `a2a deploy` installs requirements.txt during the build.
-        from a2a_pack.deepagents import create_a2a_deep_agent
-        from langchain.agents.middleware import wrap_model_call
-        from langchain_core.tools import tool
+        ctx: RunContext[PlatformUserAuth],
+        comparison_id: str,
+        quotes: list[QuoteInput],
+        weights: QuoteWeights,
+    ) -> dict[str, Any]:
+        tenant = _tenant_key(ctx)
+        validation = _validate_comparison_request(comparison_id, quotes, weights)
+        if validation is not None:
+            return validation
 
-        @tool
-        def text_stats(text: str) -> str:
-            """Return exact word, character, and line counts for text."""
-            words = [part for part in text.split() if part.strip()]
-            return json.dumps(
+        result = _score_quotes(comparison_id, quotes, weights)
+        receipt = _build_receipt(
+            tenant_key=tenant,
+            skill_name="compare_quotes",
+            comparison_id=comparison_id,
+            inputs={
+                "comparison_id": comparison_id,
+                "quotes": [quote.model_dump(mode="json") for quote in quotes],
+                "weights": weights.model_dump(mode="json"),
+            },
+            result=result,
+        )
+        persisted = _persist_comparison(tenant, comparison_id, result, receipt)
+        result["receipt"] = {
+            "receipt_id": receipt["receipt_id"],
+            "persisted": persisted["ok"],
+            "created_at": receipt["created_at"],
+        }
+        if persisted.get("warning"):
+            result.setdefault("warnings", []).append(persisted["warning"])
+        await ctx.emit_progress(f"saved comparison {comparison_id}")
+        return result
+
+    @a2a.tool(
+        description="Reopen a saved quote comparison for the authenticated user.",
+        timeout_seconds=20,
+        idempotent=True,
+        cost_class="deterministic",
+    )
+    async def get_comparison(
+        self,
+        ctx: RunContext[PlatformUserAuth],
+        comparison_id: str,
+    ) -> dict[str, Any]:
+        tenant = _tenant_key(ctx)
+        if not _valid_comparison_id(comparison_id):
+            return {"ok": False, "code": "invalid_comparison_id", "comparison_id": comparison_id}
+        record = _load_comparison(tenant, comparison_id)
+        if record is None:
+            return {"ok": False, "code": "comparison_not_found", "comparison_id": comparison_id}
+        result = dict(record["result"])
+        result["comparison_id"] = comparison_id
+        result["ok"] = True
+        result["reloaded"] = True
+        result["saved_at"] = record.get("updated_at")
+        if record.get("latest_receipt_id"):
+            result["latest_receipt_id"] = record["latest_receipt_id"]
+        return result
+
+    @a2a.tool(
+        description="Browser-safe upload bridge: accept bounded base64 JSON/CSV/text quote documents, parse them, compare, persist, and receipt the result.",
+        timeout_seconds=30,
+        idempotent=True,
+        cost_class="deterministic",
+    )
+    async def compare_quotes_from_browser_upload(
+        self,
+        ctx: RunContext[PlatformUserAuth],
+        comparison_id: str,
+        documents: list[BrowserDocument],
+        weights: QuoteWeights,
+    ) -> dict[str, Any]:
+        parsed = _decode_browser_documents(documents)
+        if not parsed["ok"]:
+            return {"ok": False, "comparison_id": comparison_id, **parsed}
+        quotes = parsed["quotes"]
+        result = await self.compare_quotes(ctx, comparison_id, quotes, weights)
+        result["source"] = "browser_upload"
+        result["parsed_documents"] = parsed["documents"]
+        return result
+
+    @a2a.tool(
+        description="External Agent API upload path: accept a typed FileUpload JSON/CSV/text document, parse quotes, compare, persist, and receipt the result.",
+        timeout_seconds=30,
+        idempotent=True,
+        cost_class="deterministic",
+    )
+    async def compare_quotes_from_file(
+        self,
+        ctx: RunContext[PlatformUserAuth],
+        comparison_id: str,
+        document: Annotated[
+            UploadedFile,
+            FileUpload(
+                accept=sorted(ALLOWED_UPLOAD_MEDIA_TYPES),
+                max_bytes=MAX_UPLOAD_BYTES,
+                description="JSON or CSV quote file with vendor, quantity, unit_price, delivery_days, warranty_months.",
+            ),
+        ],
+        weights: QuoteWeights,
+    ) -> dict[str, Any]:
+        if document.size_bytes > MAX_UPLOAD_BYTES:
+            return {"ok": False, "code": "file_too_large", "comparison_id": comparison_id}
+        media_type = (document.media_type or "").split(";", 1)[0].strip().lower()
+        if media_type not in ALLOWED_UPLOAD_MEDIA_TYPES:
+            return {"ok": False, "code": "unsupported_media_type", "comparison_id": comparison_id}
+        try:
+            reader = getattr(ctx.workspace, "read_bytes", None)
+            if reader is None:
+                return {"ok": False, "code": "workspace_read_unavailable", "comparison_id": comparison_id}
+            data = reader(document.path)
+        except Exception:
+            return {"ok": False, "code": "file_read_failed", "comparison_id": comparison_id}
+        parsed = _parse_quote_document(data, filename=document.filename, media_type=media_type)
+        if not parsed["ok"]:
+            return {"ok": False, "comparison_id": comparison_id, **parsed}
+        result = await self.compare_quotes(ctx, comparison_id, parsed["quotes"], weights)
+        result["source"] = "file_upload"
+        result["parsed_documents"] = [{"filename": document.filename, "quotes": len(parsed["quotes"])}]
+        return result
+
+
+def _tenant_key(ctx: RunContext[PlatformUserAuth]) -> str:
+    auth = ctx.auth
+    stable = auth.user_id if auth.user_id is not None else (auth.sub or auth.email)
+    return f"user:{stable}"
+
+
+def _valid_comparison_id(value: str) -> bool:
+    return bool(COMPARISON_ID_RE.fullmatch(str(value or "").strip()))
+
+
+def _validate_comparison_request(
+    comparison_id: str,
+    quotes: list[QuoteInput],
+    weights: QuoteWeights,
+) -> dict[str, Any] | None:
+    if not _valid_comparison_id(comparison_id):
+        return {"ok": False, "code": "invalid_comparison_id", "comparison_id": comparison_id}
+    if len(quotes) < 2:
+        return {"ok": False, "code": "at_least_two_quotes_required", "comparison_id": comparison_id}
+    if len(quotes) > MAX_QUOTES:
+        return {"ok": False, "code": "too_many_quotes", "comparison_id": comparison_id}
+    vendor_names = [quote.vendor.casefold() for quote in quotes]
+    if len(set(vendor_names)) != len(vendor_names):
+        return {"ok": False, "code": "duplicate_vendor", "comparison_id": comparison_id}
+    if weights.price + weights.delivery + weights.warranty <= 0:
+        return {"ok": False, "code": "weights_sum_to_zero", "comparison_id": comparison_id}
+    return None
+
+
+def _score_quotes(comparison_id: str, quotes: list[QuoteInput], weights: QuoteWeights) -> dict[str, Any]:
+    total_weight = Decimal(str(weights.price + weights.delivery + weights.warranty))
+    prices = [_money(q.unit_price) for q in quotes]
+    totals = [_money(q.unit_price * q.quantity) for q in quotes]
+    deliveries = [Decimal(q.delivery_days) for q in quotes]
+    warranties = [Decimal(q.warranty_months) for q in quotes]
+
+    rows: list[dict[str, Any]] = []
+    for index, quote in enumerate(quotes):
+        price_score = _lower_is_better(totals[index], totals)
+        delivery_score = _lower_is_better(deliveries[index], deliveries)
+        warranty_score = _higher_is_better(warranties[index], warranties)
+        weighted_score = (
+            price_score * Decimal(str(weights.price))
+            + delivery_score * Decimal(str(weights.delivery))
+            + warranty_score * Decimal(str(weights.warranty))
+        ) / total_weight
+        rows.append(
             {
-                    "characters": len(text),
-                    "words": len(words),
-                    "lines": len(text.splitlines()) or 1,
+                "vendor": quote.vendor,
+                "quantity": quote.quantity,
+                "unit_price": _float2(prices[index]),
+                "total_price": _float2(totals[index]),
+                "delivery_days": quote.delivery_days,
+                "warranty_months": quote.warranty_months,
+                "score": _float2(weighted_score),
+                "score_breakdown": {
+                    "price": _float2(price_score),
+                    "delivery": _float2(delivery_score),
+                    "warranty": _float2(warranty_score),
+                },
             }
         )
+    rows.sort(key=lambda item: (-item["score"], item["total_price"], item["delivery_days"], -item["warranty_months"], item["vendor"]))
+    for rank, row in enumerate(rows, start=1):
+        row["rank"] = rank
+    recommendation = rows[0]
+    return {
+        "ok": True,
+        "comparison_id": comparison_id,
+        "recommendation": {
+            "vendor": recommendation["vendor"],
+            "score": recommendation["score"],
+            "total_price": recommendation["total_price"],
+            "delivery_days": recommendation["delivery_days"],
+            "warranty_months": recommendation["warranty_months"],
+            "rationale": (
+                f"{recommendation['vendor']} has the best weighted score using "
+                f"price={weights.price:g}, delivery={weights.delivery:g}, warranty={weights.warranty:g}."
+            ),
+        },
+        "rankings": rows,
+        "weights": weights.model_dump(mode="json"),
+        "quote_count": len(quotes),
+    }
 
-        @wrap_model_call
-        async def log_model_call(request: Any, handler: Any) -> Any:
-            messages = request.state.get("messages", [])
-            print(
-                "[middleware] model_call "
-                f"model={creds.model} source={creds.source} messages={len(messages)}"
-            )
-            return await handler(request)
 
-        backend = ctx.workspace_backend()
-        skill_sources = _seed_runtime_skills(backend, ctx)
-        # create_a2a_deep_agent resolves provider:model strings with
-        # langchain.init_chat_model from ctx.llm, preserving LiteLLM routing,
-        # provider-specific extra body, and runtime model overrides.
-        return create_a2a_deep_agent(
-            ctx,
-            creds=creds,
-            backend=backend,
-            skills=skill_sources or None,
-            tools=[text_stats],
-            middleware=[log_model_call],
-            system_prompt=SYSTEM_PROMPT,
+def _money(value: float | int | Decimal) -> Decimal:
+    return Decimal(str(value)).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
+
+
+def _float2(value: Decimal) -> float:
+    return float(value.quantize(Decimal("0.01"), rounding=ROUND_HALF_UP))
+
+
+def _lower_is_better(value: Decimal, values: list[Decimal]) -> Decimal:
+    low, high = min(values), max(values)
+    if low == high:
+        return Decimal("100")
+    return ((high - value) / (high - low) * Decimal("100")).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
+
+
+def _higher_is_better(value: Decimal, values: list[Decimal]) -> Decimal:
+    low, high = min(values), max(values)
+    if low == high:
+        return Decimal("100")
+    return ((value - low) / (high - low) * Decimal("100")).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
+
+
+def _decode_browser_documents(documents: list[BrowserDocument]) -> dict[str, Any]:
+    if not documents:
+        return {"ok": False, "code": "no_documents_provided"}
+    if len(documents) > MAX_BROWSER_DOCUMENTS:
+        return {"ok": False, "code": "too_many_documents"}
+    all_quotes: list[QuoteInput] = []
+    summaries: list[dict[str, Any]] = []
+    for doc in documents:
+        try:
+            data = base64.b64decode(doc.data_base64.encode("ascii"), validate=True)
+        except Exception:
+            return {"ok": False, "code": "invalid_base64", "filename": doc.filename}
+        if len(data) > MAX_UPLOAD_BYTES:
+            return {"ok": False, "code": "file_too_large", "filename": doc.filename}
+        parsed = _parse_quote_document(data, filename=doc.filename, media_type=doc.media_type)
+        if not parsed["ok"]:
+            return {"ok": False, "filename": doc.filename, **parsed}
+        all_quotes.extend(parsed["quotes"])
+        summaries.append({"filename": doc.filename, "quotes": len(parsed["quotes"])})
+    return {"ok": True, "quotes": all_quotes, "documents": summaries}
+
+
+def _parse_quote_document(data: bytes, *, filename: str, media_type: str) -> dict[str, Any]:
+    if len(data) > MAX_UPLOAD_BYTES:
+        return {"ok": False, "code": "file_too_large"}
+    text = data.decode("utf-8-sig", errors="replace").strip()
+    if not text:
+        return {"ok": False, "code": "empty_document"}
+    try:
+        if media_type in {"application/json", "text/json"} or filename.lower().endswith(".json"):
+            raw = json.loads(text)
+            items = raw.get("quotes", raw) if isinstance(raw, dict) else raw
+            if not isinstance(items, list):
+                return {"ok": False, "code": "quotes_array_required"}
+            return {"ok": True, "quotes": [QuoteInput.model_validate(item) for item in items]}
+        rows = list(csv.DictReader(io.StringIO(text)))
+        if rows and rows[0]:
+            return {"ok": True, "quotes": [_quote_from_row(row) for row in rows]}
+        return {"ok": False, "code": "unsupported_document_format"}
+    except Exception:
+        return {"ok": False, "code": "document_parse_failed"}
+
+
+def _quote_from_row(row: dict[str, Any]) -> QuoteInput:
+    normalized = {str(key).strip().lower(): value for key, value in row.items()}
+    return QuoteInput(
+        vendor=str(normalized.get("vendor") or normalized.get("supplier") or ""),
+        quantity=int(float(normalized.get("quantity") or normalized.get("qty") or 0)),
+        unit_price=float(normalized.get("unit_price") or normalized.get("price") or 0),
+        delivery_days=int(float(normalized.get("delivery_days") or normalized.get("delivery") or 0)),
+        warranty_months=int(float(normalized.get("warranty_months") or normalized.get("warranty") or 0)),
     )
 
 
-def _runtime_skills_root(ctx: RunContext[Any]) -> str:
-    workspace = getattr(ctx, "_workspace", None)
-    prefixes = tuple(getattr(workspace, "write_prefixes", ()) or ())
-    if not prefixes:
-        outputs_prefix = getattr(workspace, "outputs_prefix", None)
-        prefixes = (outputs_prefix or "outputs/",)
-    prefix = str(prefixes[0]).strip("/")
-    return f"/{prefix}/{RUNTIME_SKILLS_DIR}" if prefix else f"/{RUNTIME_SKILLS_DIR}"
+def _database_url() -> str | None:
+    return os.environ.get("DATABASE_URL")
 
 
-def _seed_runtime_skills(backend: Any, ctx: RunContext[Any]) -> list[str]:
-    """Copy packaged DeepAgents skills into the invocation workspace.
+def _connect() -> Any:
+    if psycopg is None:
+        raise RuntimeError("psycopg is not installed")
+    url = _database_url()
+    if not url:
+        raise RuntimeError("DATABASE_URL is not configured")
+    return psycopg.connect(url, row_factory=dict_row)
 
-    DeepAgents loads skills from its backend, while source-controlled
-    ``skills/`` folders live in the image. This bridge lets generated agents
-    ship reusable SKILL.md bundles without giving up durable A2A workspace
-    files.
+
+def _persist_comparison(
+    tenant_key: str,
+    comparison_id: str,
+    result: dict[str, Any],
+    receipt: dict[str, Any],
+) -> dict[str, Any]:
+    if not _database_url() or psycopg is None:
+        _LOCAL_COMPARISONS[(tenant_key, comparison_id)] = {
+            "result": json.loads(json.dumps(result)),
+            "updated_at": receipt["created_at"],
+            "latest_receipt_id": receipt["receipt_id"],
+        }
+        _LOCAL_RECEIPTS.append(receipt)
+        return {"ok": True, "warning": "local_ephemeral_persistence"}
+    try:
+        with _connect() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
                     """
-    root = Path(__file__).parent / "skills"
-    if not root.exists():
-        return []
-    runtime_skills_root = _runtime_skills_root(ctx)
-    uploads: list[tuple[str, bytes]] = []
-    for path in root.rglob("*"):
-        if path.is_file():
-            rel = path.relative_to(root).as_posix()
-            uploads.append((runtime_skills_root + rel, path.read_bytes()))
-    if uploads:
-        backend.upload_files(uploads)
-        return [runtime_skills_root]
-    return []
+                    INSERT INTO quote_judge_comparisons
+                        (tenant_key, comparison_id, input_json, result_json, recommendation_vendor, latest_receipt_id, updated_at)
+                    VALUES (%s, %s, %s, %s, %s, %s, NOW())
+                    ON CONFLICT (tenant_key, comparison_id) DO UPDATE SET
+                        input_json = EXCLUDED.input_json,
+                        result_json = EXCLUDED.result_json,
+                        recommendation_vendor = EXCLUDED.recommendation_vendor,
+                        latest_receipt_id = EXCLUDED.latest_receipt_id,
+                        updated_at = NOW()
+                    """,
+                    (
+                        tenant_key,
+                        comparison_id,
+                        Jsonb({"weights": result["weights"], "quote_count": result["quote_count"]}),
+                        Jsonb(result),
+                        result["recommendation"]["vendor"],
+                        receipt["receipt_id"],
+                    ),
+                )
+                cur.execute(
+                    """
+                    INSERT INTO quote_judge_receipts
+                        (receipt_id, tenant_key, comparison_id, skill_name, tool_name, status, input_hash, result_hash, receipt_json, payload)
+                    VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s, %s)
+                    """,
+                    (
+                        receipt["receipt_id"],
+                        tenant_key,
+                        comparison_id,
+                        receipt["skill_name"],
+                        receipt["skill_name"],
+                        receipt["status"],
+                        receipt["input_hash"],
+                        receipt["result_hash"],
+                        Jsonb(receipt),
+                        Jsonb(receipt),
+                    ),
+                )
+            conn.commit()
+        return {"ok": True}
+    except Exception:
+        return {"ok": False, "warning": "persistence_unavailable"}
 
 
-def _last_message_text(state: dict[str, Any]) -> str:
-    messages = state.get("messages") or []
-    if not messages:
-        return json.dumps(state, default=str)
+def _load_comparison(tenant_key: str, comparison_id: str) -> dict[str, Any] | None:
+    if not _database_url() or psycopg is None:
+        return _LOCAL_COMPARISONS.get((tenant_key, comparison_id))
+    try:
+        with _connect() as conn:
+            with conn.cursor() as cur:
+                cur.execute(
+                    """
+                    SELECT result_json, updated_at, latest_receipt_id
+                    FROM quote_judge_comparisons
+                    WHERE tenant_key = %s AND comparison_id = %s
+                    """,
+                    (tenant_key, comparison_id),
+                )
+                row = cur.fetchone()
+        if not row:
+            return None
+        updated_at = row["updated_at"]
+        return {
+            "result": row["result_json"],
+            "updated_at": updated_at.isoformat() if hasattr(updated_at, "isoformat") else str(updated_at),
+            "latest_receipt_id": row.get("latest_receipt_id"),
+        }
+    except Exception:
+        return None
 
-    content = getattr(messages[-1], "content", None)
-    if isinstance(content, str):
-        return content
-    if isinstance(content, list):
-        parts: list[str] = []
-        for item in content:
-            if isinstance(item, dict):
-                text = item.get("text") or item.get("content")
-                if text:
-                    parts.append(str(text))
-            elif item:
-                parts.append(str(item))
-        return "\n".join(parts) if parts else json.dumps(content, default=str)
-    return str(content or messages[-1])
+
+def _build_receipt(
+    *,
+    tenant_key: str,
+    skill_name: str,
+    comparison_id: str,
+    inputs: dict[str, Any],
+    result: dict[str, Any],
+) -> dict[str, Any]:
+    now = datetime.now(timezone.utc).isoformat()
+    tenant_hash = _hash_json({"tenant_key": tenant_key})
+    input_hash = _hash_json(inputs)
+    result_hash = _hash_json(result)
+    receipt_id = "qjr_" + _hash_json({
+        "tenant_hash": tenant_hash,
+        "skill_name": skill_name,
+        "comparison_id": comparison_id,
+        "input_hash": input_hash,
+        "result_hash": result_hash,
+    })[:24]
+    return {
+        "schema": "quotejudge.production_receipt.v1",
+        "receipt_id": receipt_id,
+        "agent_name": "quote-judge-studio-v1",
+        "agent_version": "0.1.0",
+        "skill_name": skill_name,
+        "comparison_id": comparison_id,
+        "status": "ok" if result.get("ok") else "error",
+        "tenant_hash": tenant_hash,
+        "input_hash": input_hash,
+        "result_hash": result_hash,
+        "created_at": now,
+    }
+
+
+def _hash_json(value: Any) -> str:
+    encoded = json.dumps(value, sort_keys=True, separators=(",", ":"), default=str).encode("utf-8")
+    return hashlib.sha256(encoded).hexdigest()