sentrystr stuff

feat: prune dead zero-balance API keys with background janitor

.env.example

@@ -34,6 +34,14 @@ UPSTREAM_API_KEY=your-upstream-api-key
 # LOG_LEVEL=INFO
 # ENABLE_CONSOLE_LOGGING=true
 
+# SentryStr: mirror logs to Nostr relays (requires `pip install sentrystr`).
+# Announcements run on a background worker thread so they never block requests.
+# ENABLE_SENTRYSTR=false
+# SENTRYSTR_LOG_LEVEL=ERROR            # defaults to ERROR; keep >= LOG_LEVEL
+# SENTRYSTR_RELAYS="wss://relay.damus.io,wss://nos.lol"  # defaults to RELAYS
+# SENTRYSTR_NSEC=nsec1...              # defaults to NSEC; ephemeral key if unset
+# SENTRYSTR_RECIPIENT_NPUB=npub1...    # if set, log events are NIP-44 encrypted
+
 # Custom Model Management
 # BASE_URL=https://openrouter.ai/api/v1
 # MODELS_PATH=models.json

pyproject.toml

@@ -23,6 +23,11 @@ dependencies = [
     "litellm>=1.55.0",
 ]
 
+[project.optional-dependencies]
+# Decentralized log mirroring to Nostr relays. Enable at runtime with
+# ENABLE_SENTRYSTR=true (see routstr/core/logging.py).
+sentrystr = ["sentrystr>=0.2.0"]
+
 [dependency-groups]
 dev = [
     "mypy>=1.15.0",

routstr/auth.py

@@ -1021,32 +1021,37 @@ async def adjust_payment_for_tokens(
             # actual cost exceeded discounted reservation (due to tolerance_percentage)
             if cost_difference > 0:
                 # Always release the reservation and charge min(actual_cost, balance).
-                # Using a CASE expression makes this a single atomic UPDATE — no
-                # multi-level fallback needed and balance can never go negative.
+                # CASE expressions keep this atomic and safe even when the
+                # stale-reservation sweeper has already released the reservation.
                 chargeable = case(
                     (col(ApiKey.balance) >= total_cost_msats, total_cost_msats),
                     else_=col(ApiKey.balance),
                 )
+                overrun_safe_reserved = case(
+                    (
+                        col(ApiKey.reserved_balance) >= deducted_max_cost,
+                        col(ApiKey.reserved_balance) - deducted_max_cost,
+                    ),
+                    else_=0,
+                )
 
                 finalize_stmt = (
                     update(ApiKey)
                     .where(col(ApiKey.hashed_key) == billing_key.hashed_key)
-                    .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                     .values(
-                        reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                        reserved_balance=overrun_safe_reserved,
                         balance=col(ApiKey.balance) - chargeable,
                         total_spent=col(ApiKey.total_spent) + chargeable,
                     )
                 )
-                result = await session.exec(finalize_stmt)  # type: ignore[call-overload]
+                await session.exec(finalize_stmt)  # type: ignore[call-overload]
 
                 if billing_key.hashed_key != key.hashed_key:
                     child_stmt = (
                         update(ApiKey)
                         .where(col(ApiKey.hashed_key) == key.hashed_key)
-                        .where(col(ApiKey.reserved_balance) >= deducted_max_cost)
                         .values(
-                            reserved_balance=col(ApiKey.reserved_balance) - deducted_max_cost,
+                            reserved_balance=overrun_safe_reserved,
                             total_spent=col(ApiKey.total_spent) + min(billing_key.balance, total_cost_msats),
                         )
                     )
@@ -1054,51 +1059,38 @@ async def adjust_payment_for_tokens(
 
                 await session.commit()
 
-                if result.rowcount:
-                    await session.refresh(billing_key)
-                    if billing_key.hashed_key != key.hashed_key:
-                        await session.refresh(key)
-                    cost.total_msats = total_cost_msats
-                    logger.info(
-                        "Finalized payment with additional charge",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "charged_amount": total_cost_msats,
-                            "new_balance": billing_key.balance,
-                            "model": model,
-                        },
-                    )
-                    await _accumulate_fee(total_cost_msats)
-                    payments_logger.info(
-                        "FINALIZE",
-                        extra={
-                            "event": "finalize",
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "model": model,
-                            "cost_reserved": deducted_max_cost,
-                            "cost_charged": total_cost_msats,
-                            "input_tokens": cost.input_tokens,
-                            "output_tokens": cost.output_tokens,
-                            "balance": billing_key.balance,
-                            "reserved_balance": billing_key.reserved_balance,
-                            "total_spent": billing_key.total_spent,
-                            "finalize_type": "overrun",
-                        },
-                    )
-                else:
-                    # Guard fired: reservation was already released by a concurrent
-                    # finalization for this key. Nothing left to do.
-                    logger.warning(
-                        "Finalization skipped - reservation already released",
-                        extra={
-                            "key_hash": key.hashed_key[:8] + "...",
-                            "billing_key_hash": billing_key.hashed_key[:8] + "...",
-                            "attempted_charge": total_cost_msats,
-                            "model": model,
-                        },
-                    )
+                await session.refresh(billing_key)
+                if billing_key.hashed_key != key.hashed_key:
+                    await session.refresh(key)
+                cost.total_msats = total_cost_msats
+                logger.info(
+                    "Finalized payment with additional charge",
+                    extra={
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "charged_amount": total_cost_msats,
+                        "new_balance": billing_key.balance,
+                        "model": model,
+                    },
+                )
+                await _accumulate_fee(total_cost_msats)
+                payments_logger.info(
+                    "FINALIZE",
+                    extra={
+                        "event": "finalize",
+                        "key_hash": key.hashed_key[:8] + "...",
+                        "billing_key_hash": billing_key.hashed_key[:8] + "...",
+                        "model": model,
+                        "cost_reserved": deducted_max_cost,
+                        "cost_charged": total_cost_msats,
+                        "input_tokens": cost.input_tokens,
+                        "output_tokens": cost.output_tokens,
+                        "balance": billing_key.balance,
+                        "reserved_balance": billing_key.reserved_balance,
+                        "total_spent": billing_key.total_spent,
+                        "finalize_type": "overrun",
+                    },
+                )
             else:
                 # Refund some of the base cost
                 refund = abs(cost_difference)

routstr/core/exceptions.py

@@ -7,11 +7,26 @@ logger = get_logger(__name__)
 
 
 class UpstreamError(Exception):
-    """Exception raised when an upstream provider fails."""
+    """Exception raised when an upstream provider fails.
 
-    def __init__(self, message: str, status_code: int = 502):
+    ``code`` carries a stable, machine-readable classification (e.g.
+    ``UPSTREAM_RATE_LIMIT``) so callers can distinguish failure kinds without
+    string-matching the message. ``details`` holds optional structured,
+    redaction-safe context. Both default to ``None`` for backwards
+    compatibility.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        status_code: int = 502,
+        code: str | None = None,
+        details: dict[str, object] | None = None,
+    ):
         self.message = message
         self.status_code = status_code
+        self.code = code
+        self.details = details
         super().__init__(message)
 
 

routstr/core/logging.py

@@ -51,6 +51,8 @@ from pythonjsonlogger import jsonlogger
 from rich.console import Console
 from rich.logging import RichHandler
 
+from .redaction import redact_obj, redact_org_ids
+
 # Only use RichHandler when stdout is a real TTY. In non-TTY contexts
 # (docker logs, pipes, CI) Rich pads every line to width and wraps long
 # records, producing visually-empty trailing whitespace and split records.
@@ -180,6 +182,37 @@ class RequestIdFilter(logging.Filter):
         return True
 
 
+# Standard ``LogRecord`` attributes that are never user-supplied ``extra``
+# fields; skipped when redacting structured extras (``msg``/``message`` are
+# handled separately above).
+_NON_EXTRA_RECORD_ATTRS = frozenset(
+    {
+        "name",
+        "msg",
+        "args",
+        "levelname",
+        "levelno",
+        "pathname",
+        "filename",
+        "module",
+        "exc_info",
+        "exc_text",
+        "stack_info",
+        "lineno",
+        "funcName",
+        "created",
+        "msecs",
+        "relativeCreated",
+        "thread",
+        "threadName",
+        "processName",
+        "process",
+        "taskName",
+        "message",
+    }
+)
+
+
 class SecurityFilter(logging.Filter):
     """Filter to remove sensitive information from logs."""
 
@@ -199,31 +232,57 @@ class SecurityFilter(logging.Filter):
         "refund_address",
     }
 
+    def _scrub(self, message: str) -> str:
+        """Redact secrets from a single string."""
+        standalone_patterns = [
+            r"Bearer\s+([a-zA-Z0-9_\-\.]{10,})",  # Bearer token (must be 10 characters or more to reduce false-positives)
+            r"cashu[A-Z]+([a-zA-Z0-9_\-\.=/+]+)",  # Cashu tokens
+            r"nsec[a-z0-9]+",  # Nostr Public / Private Key
+        ]
+        for pattern in standalone_patterns:
+            message = re.sub(pattern, "[REDACTED]", message, flags=re.IGNORECASE)
+
+        for key in self.SENSITIVE_KEYS:
+            if key in message.lower():
+                key_patterns = [
+                    rf"{key}\s*[:=]\s*([a-zA-Z0-9_\-\.=/+]+)",  # key:value or key=value (including any variant with spaces)
+                    rf'{key}\s*[:=]\s*["\']([^"\']+)["\']',  # key:"value" or key='value' (including any variant with spaces)
+                ]
+                for pattern in key_patterns:
+                    message = re.sub(
+                        pattern, f"{key}: [REDACTED]", message, flags=re.IGNORECASE
+                    )
+        return message
+
     def filter(self, record: logging.LogRecord) -> bool:
         """Filter out sensitive information from log records."""
         try:
-            message = record.getMessage()
-            standalone_patterns = [
-                r"Bearer\s+([a-zA-Z0-9_\-\.]{10,})",  # Bearer token (must be 10 characters or more to reduce false-positives)
-                r"cashu[A-Z]+([a-zA-Z0-9_\-\.=/+]+)",  # Cashu tokens
-                r"nsec[a-z0-9]+",  # Nostr Public / Private Key
-            ]
-            for pattern in standalone_patterns:
-                message = re.sub(pattern, "[REDACTED]", message, flags=re.IGNORECASE)
-
-            for key in self.SENSITIVE_KEYS:
-                if key in message.lower():
-                    key_patterns = [
-                        rf"{key}\s*[:=]\s*([a-zA-Z0-9_\-\.=/+]+)",  # key:value or key=value (including any variant with spaces)
-                        rf'{key}\s*[:=]\s*["\']([^"\']+)["\']',  # key:"value" or key='value' (including any variant with spaces)
-                    ]
-                    for pattern in key_patterns:
-                        message = re.sub(
-                            pattern, f"{key}: [REDACTED]", message, flags=re.IGNORECASE
-                        )
-            record.msg = message
+            message = redact_org_ids(record.getMessage())
+            record.msg = self._scrub(message)
             record.args = ()
 
+            # The exception traceback is appended by the *formatter* (e.g. when a
+            # QueueHandler flattens the record before it is announced to the
+            # public, permanent SentryStr relays), so the message scrub above
+            # never sees it. Pre-render and redact it here and cache the result
+            # in `exc_text` so every formatter reuses the scrubbed version.
+            if record.exc_info and record.exc_info[0] is not None and not record.exc_text:
+                record.exc_text = logging.Formatter().formatException(record.exc_info)
+            if record.exc_text:
+                record.exc_text = self._scrub(redact_org_ids(record.exc_text))
+            if record.stack_info:
+                record.stack_info = self._scrub(redact_org_ids(record.stack_info))
+
+            # Structured `extra={...}` fields are emitted by the JSON formatter
+            # straight from the record dict and never pass through the message
+            # formatting above. Redact organization IDs from any string-valued
+            # extra so they cannot leak via structured logs.
+            for attr, value in list(record.__dict__.items()):
+                if attr in _NON_EXTRA_RECORD_ATTRS:
+                    continue
+                if isinstance(value, (str, dict, list, tuple)):
+                    record.__dict__[attr] = redact_obj(value)
+
         except Exception:
             pass
 
@@ -408,6 +467,120 @@ def setup_logging() -> None:
 
     logging.config.dictConfig(LOGGING_CONFIG)
 
+    _install_sentrystr_logging(LOGGING_CONFIG)
+
+
+# Handle for the background SentryStr logging worker (None when disabled).
+_sentrystr_handle: Any = None
+
+# Bound on the in-memory queue feeding the background SentryStr worker. Records
+# beyond this are dropped rather than blocking the caller or growing memory
+# without limit when a relay is slow or unreachable.
+_SENTRYSTR_QUEUE_SIZE = 10_000
+
+# Max seconds to wait for the queued backlog to flush on shutdown before
+# abandoning it, so shutdown cannot hang on a slow or unreachable relay.
+_SENTRYSTR_SHUTDOWN_TIMEOUT = 10.0
+
+
+def _install_sentrystr_logging(logging_config: dict[str, Any]) -> None:
+    """Mirror routstr application logs to Nostr relays via SentryStr.
+
+    Announcing to relays blocks, so SentryStr's installer runs the publish on a
+    dedicated background worker thread fed by an in-memory queue; the request /
+    event-loop threads only pay the cost of enqueueing a record. Disabled unless
+    `ENABLE_SENTRYSTR` is set. Any failure here is swallowed so optional logging
+    can never take down startup.
+    """
+    global _sentrystr_handle
+
+    if _sentrystr_handle is not None:
+        return
+
+    try:
+        from .settings import settings
+    except Exception:
+        return
+
+    if not getattr(settings, "enable_sentrystr", False):
+        return
+
+    routstr_logger = logging.getLogger("routstr")
+
+    relays = list(settings.sentrystr_relays or settings.relays or [])
+    if not relays:
+        routstr_logger.warning(
+            "SentryStr logging is enabled but no relays are configured; set "
+            "SENTRYSTR_RELAYS or RELAYS. Skipping SentryStr logging."
+        )
+        return
+
+    try:
+        from sentrystr import install_sentrystr_logging
+    except Exception:
+        routstr_logger.warning(
+            "SentryStr logging is enabled but the `sentrystr` package is not "
+            "installed. Install it with `pip install sentrystr`. Skipping "
+            "SentryStr logging."
+        )
+        return
+
+    # Default to ERROR: these events are announced to public, permanent relays,
+    # so only genuine problems should be mirrored unless the operator opts into
+    # something more verbose via SENTRYSTR_LOG_LEVEL. (Keep this at or above
+    # LOG_LEVEL; a lower value is gated out by the routstr loggers' own level.)
+    level = (settings.sentrystr_log_level or "ERROR").upper()
+
+    # routstr application loggers set propagate=False, so the handler must be
+    # attached to each of them rather than relying on propagation to the root.
+    routstr_loggers = [
+        name
+        for name in logging_config.get("loggers", {})
+        if name == "routstr" or name.startswith("routstr.")
+    ] or ["routstr"]
+
+    try:
+        _sentrystr_handle = install_sentrystr_logging(
+            relays=relays,
+            private_key=(settings.sentrystr_nsec or settings.nsec or None),
+            level=level,
+            loggers=routstr_loggers,
+            recipient_pubkey=(settings.sentrystr_recipient_npub or None),
+            platform="routstr",
+            formatter=logging.Formatter("%(message)s"),
+            # Run the same scrubbing/enrichment filters used by the other
+            # handlers, on the calling thread, before records are handed to the
+            # background worker. RequestIdFilter must run here because it reads a
+            # ContextVar that is only set on the request thread; SecurityFilter
+            # must run here so secrets are redacted before anything is announced
+            # to the (public, permanent) relays. Passing them to the installer
+            # attaches them before the handler goes live, so no record can slip
+            # through unfiltered during startup.
+            filters=[RequestIdFilter(), SecurityFilter()],
+            # Bound the queue so a slow/unreachable relay can't grow memory.
+            queue_size=_SENTRYSTR_QUEUE_SIZE,
+        )
+    except Exception as exc:
+        routstr_logger.warning("Failed to initialize SentryStr logging: %s", exc)
+        return
+
+    routstr_logger.info(
+        "SentryStr logging enabled",
+        extra={"sentrystr_level": level, "sentrystr_relay_count": len(relays)},
+    )
+
+
+def shutdown_sentrystr_logging() -> None:
+    """Flush and stop the background SentryStr logging worker, if running."""
+    global _sentrystr_handle
+    handle = _sentrystr_handle
+    _sentrystr_handle = None
+    if handle is not None:
+        try:
+            handle.close(timeout=_SENTRYSTR_SHUTDOWN_TIMEOUT)
+        except Exception:
+            pass
+
 
 def get_logger(name: str) -> logging.Logger:
     """Get a logger instance for the given module name."""

routstr/core/main.py

@@ -230,6 +230,13 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
                 extra={"error": str(e), "error_type": type(e).__name__},
             )
 
+        # Flush and stop the background SentryStr logging worker (no-op when
+        # SentryStr logging is disabled). Run off the event loop: the drain is
+        # blocking (though bounded) and must not stall the loop during shutdown.
+        from .logging import shutdown_sentrystr_logging
+
+        await asyncio.to_thread(shutdown_sentrystr_logging)
+
 
 class _ImmutableStaticFiles(StaticFiles):
     """Static files with long Cache-Control for content-hashed Next.js assets.
@@ -379,7 +386,10 @@ if UI_DIST_PATH.exists() and UI_DIST_PATH.is_dir():
 
 else:
     logger.warning(
-        f"UI dist directory not found at {UI_DIST_PATH}, skipping static file serving"
+        "UI dist directory not found at %s; serving API only. Run `make ui-build` "
+        "to build the static UI served from here, or `make ui-dev` for the Next.js "
+        "dev server with hot reload on :3000 (it targets this backend on :8000).",
+        UI_DIST_PATH,
     )
 
     @app.get("/", include_in_schema=False)

routstr/core/redaction.py

@@ -0,0 +1,51 @@
+"""Redaction helpers for sensitive provider identifiers.
+
+Single source of truth for stripping account-scoped identifiers (e.g. OpenAI
+organization IDs) from any text before it is logged, returned to a caller, or
+written to an audit entry.
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Any
+
+# OpenAI-style organization identifiers look like ``org-<base62>``. Require at
+# least 6 trailing chars so the already-redacted literal ``org-[REDACTED]`` is
+# never re-matched (``[`` is not in the character class).
+_ORG_ID_PATTERN = re.compile(r"\borg-[A-Za-z0-9]{6,}\b")
+
+ORG_ID_PLACEHOLDER = "org-[REDACTED]"
+
+
+def redact_org_ids(text: str) -> str:
+    """Replace OpenAI-style organization IDs with ``org-[REDACTED]``.
+
+    Args:
+        text: Arbitrary text that may embed an ``org-*`` identifier.
+
+    Returns:
+        The text with every organization ID replaced. Non-string input is
+        returned unchanged after coercion to ``str``.
+    """
+    if not text:
+        return text
+    return _ORG_ID_PATTERN.sub(ORG_ID_PLACEHOLDER, text)
+
+
+def redact_obj(obj: Any) -> Any:
+    """Recursively redact organization IDs in arbitrary nested structures.
+
+    Strings are redacted in place; dicts and lists/tuples are walked so that
+    identifiers nested inside structured payloads (e.g. log ``extra`` fields or
+    error ``details``) are also stripped. Other types are returned unchanged.
+    """
+    if isinstance(obj, str):
+        return redact_org_ids(obj)
+    if isinstance(obj, dict):
+        return {key: redact_obj(value) for key, value in obj.items()}
+    if isinstance(obj, list):
+        return [redact_obj(value) for value in obj]
+    if isinstance(obj, tuple):
+        return tuple(redact_obj(value) for value in obj)
+    return obj

routstr/core/settings.py

@@ -16,7 +16,12 @@ class Settings(BaseSettings):
 
         @classmethod
         def parse_env_var(cls, field_name: str, raw_value: str) -> Any:  # type: ignore[override]
-            if field_name in {"cashu_mints", "cors_origins", "relays"}:
+            if field_name in {
+                "cashu_mints",
+                "cors_origins",
+                "relays",
+                "sentrystr_relays",
+            }:
                 v = str(raw_value).strip()
                 if v == "":
                     return []
@@ -103,6 +108,21 @@ class Settings(BaseSettings):
     log_level: str = Field(default="INFO", env="LOG_LEVEL")
     enable_console_logging: bool = Field(default=True, env="ENABLE_CONSOLE_LOGGING")
 
+    # SentryStr (decentralized logging to Nostr relays). When enabled, routstr
+    # application logs at `sentrystr_log_level` and above are mirrored to Nostr
+    # relays on a background worker thread (see routstr/core/logging.py).
+    enable_sentrystr: bool = Field(default=False, env="ENABLE_SENTRYSTR")
+    # Minimum level forwarded to relays. Empty -> defaults to ERROR. Should be
+    # at or above `log_level`, since the routstr loggers gate lower records.
+    sentrystr_log_level: str = Field(default="", env="SENTRYSTR_LOG_LEVEL")
+    # Relays to announce to. Empty -> fall back to the discovery `relays`.
+    sentrystr_relays: list[str] = Field(default_factory=list, env="SENTRYSTR_RELAYS")
+    # Secret key (nsec or hex) used to sign log events. Empty -> fall back to
+    # `nsec`; if that is also empty an ephemeral key is generated per process.
+    sentrystr_nsec: str = Field(default="", env="SENTRYSTR_NSEC")
+    # Optional recipient npub. When set, log events are NIP-44 encrypted to it.
+    sentrystr_recipient_npub: str = Field(default="", env="SENTRYSTR_RECIPIENT_NPUB")
+
     # Other
     chat_completions_api_version: str = Field(
         default="", env="CHAT_COMPLETIONS_API_VERSION"

routstr/payment/helpers.py

@@ -11,6 +11,8 @@ from PIL import Image
 from sqlmodel.ext.asyncio.session import AsyncSession
 
 from ..core import get_logger
+from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..core.settings import settings
 from ..wallet import deserialize_token_from_string
 
@@ -418,16 +420,27 @@ def create_error_response(
     status_code: int,
     request: Request,
     token: str | None = None,
+    code: str | int | None = None,
+    details: dict[str, object] | None = None,
 ) -> Response:
-    """Create a standardized error response."""
+    """Create a standardized error response.
+
+    ``code`` is a stable, machine-readable classification (e.g.
+    ``UPSTREAM_RATE_LIMIT``); when omitted it defaults to the HTTP status code
+    for backwards compatibility. ``details`` carries optional structured,
+    redaction-safe context.
+    """
+    error_obj: dict[str, object] = {
+        "message": redact_org_ids(message),
+        "type": error_type,
+        "code": code if code is not None else status_code,
+    }
+    if details is not None:
+        error_obj["details"] = details
     return Response(
         content=json.dumps(
             {
-                "error": {
-                    "message": message,
-                    "type": error_type,
-                    "code": status_code,
-                },
+                "error": error_obj,
                 "request_id": getattr(request.state, "request_id", "unknown"),
             }
         ),
@@ -435,3 +448,20 @@ def create_error_response(
         media_type="application/json",
         headers={"X-Cashu": token} if token else {},
     )
+
+
+def create_upstream_error_response(
+    error: UpstreamError,
+    request: Request,
+    fallback_status: int = 502,
+) -> Response:
+    """Build an error response from an :class:`UpstreamError`, preserving its
+    structured ``code``, ``details``, and original ``status_code``."""
+    return create_error_response(
+        "upstream_error",
+        str(error),
+        error.status_code or fallback_status,
+        request=request,
+        code=getattr(error, "code", None),
+        details=getattr(error, "details", None),
+    )

routstr/proxy.py

@@ -24,6 +24,7 @@ from .payment.helpers import (
     calculate_discounted_max_cost,
     check_token_balance,
     create_error_response,
+    create_upstream_error_response,
     get_max_cost_for_model,
 )
 from .payment.models import Model
@@ -211,9 +212,7 @@ async def proxy(
                     e,
                 )
                 if i == len(all_upstreams) - 1:
-                    last_error_response = create_error_response(
-                        "upstream_error", str(e), 502, request=request
-                    )
+                    last_error_response = create_upstream_error_response(e, request)
                 continue
         return last_error_response or create_error_response(
             "upstream_error", "All upstreams failed", 502, request=request
@@ -283,11 +282,10 @@ async def proxy(
                     last_error = e
                 continue
 
+        if last_error is not None:
+            return create_upstream_error_response(last_error, request)
         return create_error_response(
-            "upstream_error",
-            str(last_error) if last_error else "All upstreams failed",
-            502,
-            request=request,
+            "upstream_error", "All upstreams failed", 502, request=request
         )
 
     elif auth := headers.get("authorization", None):
@@ -343,9 +341,7 @@ async def proxy(
             except UpstreamError as e:
                 logger.warning(f"Upstream {upstream.provider_type} failed (GET): {e}")
                 if i == len(upstreams) - 1:
-                    last_error_response = create_error_response(
-                        "upstream_error", str(e), 502, request=request
-                    )
+                    last_error_response = create_upstream_error_response(e, request)
                 continue
         return last_error_response or create_error_response(
             "upstream_error", "All upstreams failed", 502, request=request
@@ -525,9 +521,7 @@ async def proxy(
             # If this was the last provider
             if i == len(upstreams) - 1:
                 await revert_pay_for_request(key, session, max_cost_for_model)
-                return create_error_response(
-                    "upstream_error", str(e), 502, request=request
-                )
+                return create_upstream_error_response(e, request)
 
             # Otherwise loop continues to next provider
             continue

routstr/upstream/base.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import asyncio
 import json
+import math
 import traceback
 import uuid
 from collections.abc import AsyncGenerator, AsyncIterator, Iterator
@@ -23,6 +24,7 @@ from ..core.db import (
     store_cashu_transaction,
 )
 from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..payment.cost_calculation import (
     CostData,
     CostDataError,
@@ -47,6 +49,7 @@ from .cache_breakpoints import (
 )
 from .count_tokens import count_tokens_locally
 from .litellm_routing import detect_litellm_prefix
+from .rate_limit import UPSTREAM_RATE_LIMIT, classify_rate_limit
 
 logger = get_logger(__name__)
 
@@ -582,7 +585,7 @@ class BaseUpstreamProvider:
             preview = body_bytes.decode("utf-8", errors="ignore").strip()
             if preview:
                 message = preview[:500]
-        return message, upstream_code
+        return redact_org_ids(message), upstream_code
 
     async def on_upstream_error_redirect(
         self, status_code: int, error_message: str
@@ -625,10 +628,23 @@ class BaseUpstreamProvider:
             body_bytes = b""
             body_read_error = f"{type(exc).__name__}: {exc}"
 
+        # ``message`` is already redacted by ``_extract_upstream_error_message``;
+        # the raw body preview is redacted here before it reaches logs or the
+        # forwarded envelope so provider account identifiers never leak.
         message, upstream_code = self._extract_upstream_error_message(body_bytes)
-        body_preview = body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+        body_preview = redact_org_ids(
+            body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+        )
         is_json_body = _is_json_content_type(content_type)
 
+        # Classify upstream rate-limit failures into a stable, structured error.
+        rate_limit = classify_rate_limit(status_code, message, headers)
+        error_code: str | int = upstream_code or status_code
+        error_details: dict[str, object] | None = None
+        if rate_limit is not None:
+            error_code = UPSTREAM_RATE_LIMIT
+            error_details = rate_limit.as_details()
+
         logger.warning(
             "Upstream %s returned %s for model=%s path=%s: %s",
             self.provider_type,
@@ -642,6 +658,7 @@ class BaseUpstreamProvider:
                 "model": model_id or "unknown",
                 "upstream_status": status_code,
                 "upstream_code": upstream_code,
+                "error_code": error_code,
                 "upstream_content_type": content_type,
                 "upstream_request_id": upstream_request_id,
                 "message_preview": message[:200],
@@ -676,13 +693,41 @@ class BaseUpstreamProvider:
         ):
             headers.pop(header_name, None)
 
+        # Propagate a usable retry hint to the caller when the upstream supplied
+        # one but did not echo a ``Retry-After`` header. RFC 7231 delta-seconds
+        # is an integer, so round sub-second hints up to a usable ``1``.
+        if (
+            rate_limit is not None
+            and rate_limit.retry_after_seconds is not None
+            and "retry-after" not in {k.lower() for k in headers}
+        ):
+            headers["Retry-After"] = str(max(1, math.ceil(rate_limit.retry_after_seconds)))
+
         if is_json_body:
             if not content_type:
                 headers.pop("content-type", None)
                 headers.pop("Content-Type", None)
             media_type = content_type or None
+            # Re-serialise the body with organization IDs stripped. The narrow
+            # ``org-*`` regex preserves the surrounding JSON structure.
+            redacted_text = redact_org_ids(body_bytes.decode("utf-8", errors="ignore"))
+            redacted_body = redacted_text.encode()
+            # Surface the stable rate-limit classification on the forwarded
+            # body so callers can switch on ``error.code`` without parsing the
+            # provider-specific message. Fall back to the redacted bytes if the
+            # body is not a JSON object with an ``error`` mapping.
+            if rate_limit is not None:
+                try:
+                    parsed = json.loads(redacted_text)
+                    err = parsed.get("error") if isinstance(parsed, dict) else None
+                    if isinstance(err, dict):
+                        err["code"] = UPSTREAM_RATE_LIMIT
+                        err["details"] = error_details
+                        redacted_body = json.dumps(parsed).encode()
+                except (ValueError, AttributeError):
+                    pass
             return Response(
-                content=body_bytes,
+                content=redacted_body,
                 status_code=status_code,
                 headers=headers,
                 media_type=media_type,
@@ -693,15 +738,18 @@ class BaseUpstreamProvider:
         for header_name in ("content-type", "Content-Type"):
             headers.pop(header_name, None)
 
+        error_obj: dict[str, object] = {
+            "message": message or "Upstream returned a non-JSON error response",
+            "type": "upstream_error",
+            "code": error_code,
+            "upstream_status": status_code,
+            "upstream_content_type": content_type or None,
+            "upstream_body_preview": body_preview or None,
+        }
+        if error_details is not None:
+            error_obj["details"] = error_details
         envelope = {
-            "error": {
-                "message": message or "Upstream returned a non-JSON error response",
-                "type": "upstream_error",
-                "code": upstream_code or status_code,
-                "upstream_status": status_code,
-                "upstream_content_type": content_type or None,
-                "upstream_body_preview": body_preview or None,
-            },
+            "error": error_obj,
             "request_id": getattr(request.state, "request_id", None),
         }
 
@@ -2521,9 +2569,16 @@ class BaseUpstreamProvider:
                         body_bytes = await response.aread()
                     except Exception:
                         body_bytes = b""
-                    body_preview = body_bytes.decode(
-                        "utf-8", errors="ignore"
-                    ).strip()[:500]
+                    # Redact provider account identifiers before the body text
+                    # reaches logs or the raised error.
+                    body_preview = redact_org_ids(
+                        body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+                    )
+                    rate_limit = classify_rate_limit(
+                        response.status_code,
+                        body_preview,
+                        dict(response.headers),
+                    )
                     logger.error(
                         "Upstream %s returned %s for model=%s path=%s: %s",
                         self.provider_type,
@@ -2535,6 +2590,7 @@ class BaseUpstreamProvider:
                             "provider": self.provider_type,
                             "model": original_model_id or "unknown",
                             "status_code": response.status_code,
+                            "error_code": rate_limit.code if rate_limit else None,
                             "reason_phrase": response.reason_phrase,
                             "path": path,
                             "body_preview": body_preview,
@@ -2547,6 +2603,8 @@ class BaseUpstreamProvider:
                         f"for model {original_model_id or 'unknown'}: "
                         f"{body_preview[:200] or '<empty>'}",
                         status_code=response.status_code,
+                        code=rate_limit.code if rate_limit else None,
+                        details=rate_limit.as_details() if rate_limit else None,
                     )
 
                 try:
@@ -2843,9 +2901,16 @@ class BaseUpstreamProvider:
                         body_bytes = await response.aread()
                     except Exception:
                         body_bytes = b""
-                    body_preview = body_bytes.decode(
-                        "utf-8", errors="ignore"
-                    ).strip()[:500]
+                    # Redact provider account identifiers before the body text
+                    # reaches logs or the raised error.
+                    body_preview = redact_org_ids(
+                        body_bytes.decode("utf-8", errors="ignore").strip()[:500]
+                    )
+                    rate_limit = classify_rate_limit(
+                        response.status_code,
+                        body_preview,
+                        dict(response.headers),
+                    )
                     logger.error(
                         "Upstream %s returned %s for model=%s path=%s: %s",
                         self.provider_type,
@@ -2857,6 +2922,7 @@ class BaseUpstreamProvider:
                             "provider": self.provider_type,
                             "model": original_model_id or "unknown",
                             "status_code": response.status_code,
+                            "error_code": rate_limit.code if rate_limit else None,
                             "path": path,
                             "body_preview": body_preview,
                         },
@@ -2868,6 +2934,8 @@ class BaseUpstreamProvider:
                         f"for model {original_model_id or 'unknown'}: "
                         f"{body_preview[:200] or '<empty>'}",
                         status_code=response.status_code,
+                        code=rate_limit.code if rate_limit else None,
+                        details=rate_limit.as_details() if rate_limit else None,
                     )
 
                 try:

routstr/upstream/messages_dispatch.py

@@ -29,7 +29,9 @@ import litellm
 
 from ..core import get_logger
 from ..core.exceptions import UpstreamError
+from ..core.redaction import redact_org_ids
 from ..payment.models import Model
+from .rate_limit import classify_rate_limit
 
 logger = get_logger(__name__)
 
@@ -505,23 +507,33 @@ async def dispatch_anthropic_messages(
     try:
         result = await litellm.anthropic.messages.acreate(**kwargs)
     except Exception as exc:
-        exc_message = getattr(exc, "message", None) or str(exc) or repr(exc)
+        raw_message = getattr(exc, "message", None) or str(exc) or repr(exc)
+        # Redact provider account identifiers before the message reaches logs
+        # or the surfaced error.
+        exc_message = redact_org_ids(raw_message)
         exc_status = getattr(exc, "status_code", None)
         exc_response = getattr(exc, "response", None)
         response_text = None
         if exc_response is not None:
             try:
-                response_text = getattr(exc_response, "text", str(exc_response))
+                response_text = redact_org_ids(
+                    getattr(exc_response, "text", str(exc_response))
+                )
             except Exception:
                 response_text = "<unreadable>"
+        status_for_classify = exc_status if isinstance(exc_status, int) else 502
+        rate_limit = classify_rate_limit(
+            status_for_classify, exc_message, getattr(exc, "headers", None)
+        )
         logger.error(
             "litellm dispatch failed",
             extra={
                 "error": exc_message,
                 "error_type": type(exc).__name__,
                 "status_code": exc_status,
+                "error_code": rate_limit.code if rate_limit else None,
                 "llm_provider": getattr(exc, "llm_provider", None),
-                "body": getattr(exc, "body", None),
+                "body": redact_org_ids(str(getattr(exc, "body", "") or "")) or None,
                 "response_text": response_text,
                 "model": litellm_model,
                 "api_base": base_url,
@@ -529,7 +541,9 @@ async def dispatch_anthropic_messages(
         )
         raise UpstreamError(
             f"Upstream error via litellm: {exc_message}",
-            status_code=exc_status if isinstance(exc_status, int) else 502,
+            status_code=status_for_classify,
+            code=rate_limit.code if rate_limit else None,
+            details=rate_limit.as_details() if rate_limit else None,
         ) from exc
 
     if not client_stream and hasattr(result, "__aiter__"):

routstr/upstream/openrouter.py

@@ -1,3 +1,4 @@
+import json
 from typing import TYPE_CHECKING
 
 import httpx
@@ -18,6 +19,38 @@ class OpenRouterUpstreamProvider(BaseUpstreamProvider):
     supports_anthropic_messages = True
     litellm_provider_prefix = "openrouter/"
 
+    def prepare_request_body(
+        self, body: bytes | None, model_obj: Model
+    ) -> bytes | None:
+        """Set provider.require_parameters on tool-use requests.
+
+        Without it OpenRouter can route a tool call to an endpoint that doesn't
+        support function calling and 404 with "No endpoints found that support
+        tool use". We leave a client-supplied value untouched.
+        """
+        body = super().prepare_request_body(body, model_obj)
+        if not body:
+            return body
+
+        try:
+            data = json.loads(body)
+        except json.JSONDecodeError:
+            return body
+
+        if not isinstance(data, dict) or not data.get("tools"):
+            return body
+
+        provider = data.get("provider")
+        if not isinstance(provider, dict):
+            provider = {}
+
+        if "require_parameters" in provider:
+            return body
+
+        provider["require_parameters"] = True
+        data["provider"] = provider
+        return json.dumps(data).encode()
+
     def _apply_provider_field(self, response_json: object) -> None:
         """Stamp the ``provider`` field for OpenRouter responses.
 

routstr/upstream/rate_limit.py

@@ -0,0 +1,136 @@
+"""Detection and parsing of upstream provider rate-limit errors.
+
+Upstream OpenAI-compatible providers signal rate limits via HTTP 429 and/or a
+human-readable message such as::
+
+    Rate limit reached for gpt-5.5-2026-04-23 (for limit gpt-5.5) in organization
+    org-XXXX on tokens per min (TPM): Limit 180000000, Used 180000000,
+    Requested 8929. Please try again in 2ms.
+
+This module classifies those failures into a stable :data:`UPSTREAM_RATE_LIMIT`
+code and extracts useful debugging fields. All retained text is redacted of
+organization IDs first.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import asdict, dataclass
+
+from ..core.redaction import redact_org_ids
+
+# Stable error code callers can switch on to distinguish upstream rate limits
+# from generic request failures. The literal value matches the identifier named
+# in issue #555 ("UPSTREAM_RATE_LIMIT") so the public API contract is exact.
+UPSTREAM_RATE_LIMIT = "UPSTREAM_RATE_LIMIT"
+
+# Message fragments that indicate a rate-limit even when the status code is not
+# 429 (some providers wrap it in a 400/500 envelope).
+_RATE_LIMIT_MARKERS = (
+    "rate limit reached",
+    "rate_limit_exceeded",
+    "rate limit exceeded",
+    "too many requests",
+)
+
+_MODEL_RE = re.compile(r"Rate limit reached for ([^\s(]+)", re.IGNORECASE)
+_LIMIT_NAME_RE = re.compile(r"\(for limit ([^)]+)\)", re.IGNORECASE)
+_METRIC_RE = re.compile(r"on ([a-z ]+\((?:TPM|RPM|TPD|RPD|IPM)\))", re.IGNORECASE)
+_LIMIT_RE = re.compile(r"Limit (\d+)", re.IGNORECASE)
+_USED_RE = re.compile(r"Used (\d+)", re.IGNORECASE)
+_REQUESTED_RE = re.compile(r"Requested (\d+)", re.IGNORECASE)
+_RETRY_RE = re.compile(r"try again in ([\d.]+)\s*(ms|s)", re.IGNORECASE)
+
+
+@dataclass
+class RateLimitInfo:
+    """Structured, redaction-safe view of an upstream rate-limit error."""
+
+    code: str
+    message: str
+    model: str | None = None
+    limit_name: str | None = None
+    metric: str | None = None
+    limit: int | None = None
+    used: int | None = None
+    requested: int | None = None
+    retry_after_seconds: float | None = None
+
+    def as_details(self) -> dict[str, object]:
+        """Return a JSON-serialisable dict for embedding in an error envelope."""
+        return {k: v for k, v in asdict(self).items() if v is not None}
+
+
+def _looks_like_rate_limit(status_code: int, message: str) -> bool:
+    if status_code == 429:
+        return True
+    lowered = message.lower()
+    return any(marker in lowered for marker in _RATE_LIMIT_MARKERS)
+
+
+def _parse_retry_after_header(headers: dict[str, str] | None) -> float | None:
+    """Parse a ``Retry-After`` header (delta-seconds form) into seconds."""
+    if not headers:
+        return None
+    raw = headers.get("retry-after") or headers.get("Retry-After")
+    if raw is None:
+        return None
+    try:
+        return float(str(raw).strip())
+    except (TypeError, ValueError):
+        return None
+
+
+def _int_or_none(match: re.Match[str] | None) -> int | None:
+    if match is None:
+        return None
+    try:
+        return int(match.group(1))
+    except (TypeError, ValueError):
+        return None
+
+
+def classify_rate_limit(
+    status_code: int,
+    message: str,
+    headers: dict[str, str] | None = None,
+) -> RateLimitInfo | None:
+    """Classify an upstream error as a rate-limit and extract its fields.
+
+    Args:
+        status_code: HTTP status code from the upstream response.
+        message: Upstream error message (may contain sensitive identifiers).
+        headers: Optional upstream response headers, used for ``Retry-After``.
+
+    Returns:
+        A :class:`RateLimitInfo` when the error is a rate-limit, else ``None``.
+    """
+    message = message or ""
+    if not _looks_like_rate_limit(status_code, message):
+        return None
+
+    redacted = redact_org_ids(message)
+
+    model_match = _MODEL_RE.search(redacted)
+    metric_match = _METRIC_RE.search(redacted)
+
+    retry_after = _parse_retry_after_header(headers)
+    if retry_after is None:
+        retry_match = _RETRY_RE.search(redacted)
+        if retry_match is not None:
+            value = float(retry_match.group(1))
+            retry_after = value / 1000.0 if retry_match.group(2).lower() == "ms" else value
+
+    limit_name_match = _LIMIT_NAME_RE.search(redacted)
+
+    return RateLimitInfo(
+        code=UPSTREAM_RATE_LIMIT,
+        message=redacted,
+        model=model_match.group(1) if model_match else None,
+        limit_name=limit_name_match.group(1).strip() if limit_name_match else None,
+        metric=metric_match.group(1).strip() if metric_match else None,
+        limit=_int_or_none(_LIMIT_RE.search(redacted)),
+        used=_int_or_none(_USED_RE.search(redacted)),
+        requested=_int_or_none(_REQUESTED_RE.search(redacted)),
+        retry_after_seconds=retry_after,
+    )

routstr/wallet.py

@@ -1,4 +1,5 @@
 import asyncio
+import re
 import time
 import typing
 from typing import TypedDict
@@ -129,6 +130,72 @@ async def send_token(amount: int, unit: str, mint_url: str | None = None) -> str
     return token
 
 
+# A foreign mint's fee_reserve is a non-binding estimate (NUT-05): the mint may
+# demand more when re-quoting or at melt execution. Instead of padding the
+# estimate with a safety buffer (which strands the margin at the foreign mint
+# on every swap), the swap retries with the amount recomputed from the fees the
+# mint actually demands, up to this many attempts.
+_MAX_SWAP_ATTEMPTS = 3
+
+_MINT_ERROR_CODE_RE = re.compile(r"\(Code: (\d+)\)")
+_MELT_SHORTFALL_RE = re.compile(r"Provided: (\d+), needed: (\d+)")
+
+# Insufficient-melt-inputs failures differ across mint implementations. 11005 is
+# the registered "Transaction is not balanced" code (cdk), specific enough to
+# trust on the code alone. 11000 is nutshell's generic, unregistered
+# TransactionError covering many unrelated failures, so it only counts as a fee
+# shortfall alongside the "not enough inputs" detail text. With no code suffix at
+# all, that same text is the only signal.
+
+
+def _net_minted_amount(amount_msat: int, token_unit: str, fees: int) -> int:
+    """
+    Convert the token value minus fees (given in the token unit) into an
+    amount in the primary mint's unit.
+    """
+    fee_msat = fees * 1000 if token_unit == "sat" else fees
+    remaining_msat = amount_msat - fee_msat
+    if settings.primary_mint_unit == "sat":
+        return int(remaining_msat // 1000)
+    return int(remaining_msat)
+
+
+def _melt_insufficient_shortfall(error: Exception) -> int | None:
+    """
+    Classify a melt failure: return the observed shortfall (in the token unit)
+    when the mint rejected the inputs as insufficient, or None when the failure
+    is unrelated to fees and must not be retried (e.g. a Lightning payment
+    failure, where a smaller invoice would not help).
+
+    Cashu errors carry no structured amounts (NUT-00 defines only detail/code,
+    flattened to "Mint Error: <detail> (Code: <code>)" by cashu-py), so the
+    classification uses the code and the shortfall must be inferred: the
+    "Provided: X, needed: Y" amounts are nutshell-specific free text and only
+    refine the shortfall when present; otherwise shrink one unit at a time.
+    """
+    message = str(error)
+    code_match = _MINT_ERROR_CODE_RE.search(message)
+    code = code_match.group(1) if code_match is not None else None
+    has_shortfall_text = "not enough inputs" in message.lower()
+
+    match code:
+        case "11005":  # registered TransactionUnbalanced: trust the code
+            pass
+        case "11000" if has_shortfall_text:  # generic nutshell error: needs the text
+            pass
+        case None if has_shortfall_text:  # no code suffix: text is the only signal
+            pass
+        case _:  # other codes, a bare 11000, or no signal: must not retry
+            return None
+
+    amounts = _MELT_SHORTFALL_RE.search(message)
+    if amounts is not None:
+        provided, needed = int(amounts.group(1)), int(amounts.group(2))
+        if needed > provided:
+            return needed - provided
+    return 1
+
+
 async def _calculate_swap_amount(
     amount_msat: int,
     token_unit: str,
@@ -167,26 +234,18 @@ async def _calculate_swap_amount(
 
         fee_reserve = dummy_melt_quote.fee_reserve
         input_fees = token_wallet.get_fees_for_proofs(proofs)
-        if token_unit == "sat":
-            fee_msat = (fee_reserve + input_fees) * 1000
-        else:
-            fee_msat = fee_reserve + input_fees
-
-        amount_msat_after_fee = amount_msat - fee_msat
-
-        if settings.primary_mint_unit == "sat":
-            minted_amount = int(amount_msat_after_fee // 1000)
-        else:
-            minted_amount = int(amount_msat_after_fee)
+        total_fees = fee_reserve + input_fees
+        minted_amount = _net_minted_amount(amount_msat, token_unit, total_fees)
 
         if minted_amount <= 0:
-            raise ValueError(f"Fees ({fee_reserve + input_fees} {token_unit}) exceed token amount")
+            raise ValueError(f"Fees ({total_fees} {token_unit}) exceed token amount")
 
         logger.info(
             "swap_to_primary_mint: fee estimation result",
             extra={
                 "token_amount_sat": amount_msat // 1000,
-                "estimated_fee_sat": fee_msat // 1000,
+                "estimated_fee": total_fees,
+                "estimated_fee_unit": token_unit,
                 "input_fees": input_fees,
                 "minted_amount": minted_amount,
                 "minted_unit": settings.primary_mint_unit,
@@ -251,67 +310,116 @@ async def swap_to_primary_mint(
         token_obj.proofs,
     )
 
-    mint_quote = await primary_wallet.request_mint(minted_amount)
-    logger.info(
-        "swap_to_primary_mint: mint quote received",
-        extra={"mint_quote_id": mint_quote.quote},
-    )
+    # The estimate above is non-binding: the mint may demand a higher fee on the
+    # real quote or reject the melt outright. Retry the quote/melt cycle with the
+    # amount recomputed from the fees the mint actually demands.
+    observed_extra_fee = 0
+    attempt = 0
+    while True:
+        attempt += 1
+        mint_quote = await primary_wallet.request_mint(minted_amount)
+        logger.info(
+            "swap_to_primary_mint: mint quote received",
+            extra={"mint_quote_id": mint_quote.quote, "attempt": attempt},
+        )
 
-    melt_quote = await token_wallet.melt_quote(mint_quote.request)
-    input_fees = token_wallet.get_fees_for_proofs(token_obj.proofs)
-    total_needed = melt_quote.amount + melt_quote.fee_reserve + input_fees
-    logger.info(
-        "swap_to_primary_mint: melt quote received",
-        extra={
-            "melt_quote_id": melt_quote.quote,
-            "melt_amount": melt_quote.amount,
-            "melt_fee_reserve": melt_quote.fee_reserve,
-            "input_fees": input_fees,
-            "total_needed": total_needed,
-            "token_amount": token_amount,
-        },
-    )
-
-    if total_needed > token_amount:
-        logger.warning(
-            "swap_to_primary_mint: insufficient token amount for melt fees",
+        melt_quote = await token_wallet.melt_quote(mint_quote.request)
+        input_fees = token_wallet.get_fees_for_proofs(token_obj.proofs)
+        total_needed = melt_quote.amount + melt_quote.fee_reserve + input_fees
+        logger.info(
+            "swap_to_primary_mint: melt quote received",
             extra={
-                "token_amount": token_amount,
+                "melt_quote_id": melt_quote.quote,
                 "melt_amount": melt_quote.amount,
                 "melt_fee_reserve": melt_quote.fee_reserve,
                 "input_fees": input_fees,
                 "total_needed": total_needed,
-                "shortfall": total_needed - token_amount,
+                "token_amount": token_amount,
+                "attempt": attempt,
             },
         )
-        raise ValueError(
-            f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
-            f"melt fees. Needed: {total_needed} {token_obj.unit} "
-            f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve} + input_fees: {input_fees})"
-        )
 
-    try:
-        _ = await token_wallet.melt(
-            proofs=token_obj.proofs,
-            invoice=mint_quote.request,
-            fee_reserve_sat=melt_quote.fee_reserve,
-            quote_id=melt_quote.quote,
-        )
-    except Exception as e:
-        logger.error(
-            "swap_to_primary_mint: melt failed",
-            extra={
-                "error": str(e),
-                "error_type": type(e).__name__,
-                "foreign_mint": token_obj.mint,
-                "token_amount": token_amount,
-                "melt_quote_id": melt_quote.quote,
-                "total_needed": total_needed,
-            },
-        )
-        raise ValueError(
-            f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
-        ) from e
+        if total_needed > token_amount:
+            recomputed = _net_minted_amount(
+                amount_msat,
+                token_obj.unit,
+                melt_quote.fee_reserve + input_fees + observed_extra_fee,
+            )
+            if attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.warning(
+                    "swap_to_primary_mint: insufficient token amount for melt fees",
+                    extra={
+                        "token_amount": token_amount,
+                        "melt_amount": melt_quote.amount,
+                        "melt_fee_reserve": melt_quote.fee_reserve,
+                        "input_fees": input_fees,
+                        "total_needed": total_needed,
+                        "shortfall": total_needed - token_amount,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Token amount ({token_amount} {token_obj.unit}) is insufficient to cover "
+                    f"melt fees. Needed: {total_needed} {token_obj.unit} "
+                    f"(amount: {melt_quote.amount} + fee: {melt_quote.fee_reserve} + input_fees: {input_fees})"
+                )
+            logger.warning(
+                "swap_to_primary_mint: melt quote exceeds token amount, retrying",
+                extra={
+                    "total_needed": total_needed,
+                    "token_amount": token_amount,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        try:
+            _ = await token_wallet.melt(
+                proofs=token_obj.proofs,
+                invoice=mint_quote.request,
+                fee_reserve_sat=melt_quote.fee_reserve,
+                quote_id=melt_quote.quote,
+            )
+        except Exception as e:
+            shortfall = _melt_insufficient_shortfall(e)
+            recomputed = 0
+            if shortfall is not None:
+                observed_extra_fee += shortfall
+                recomputed = _net_minted_amount(
+                    amount_msat,
+                    token_obj.unit,
+                    melt_quote.fee_reserve + input_fees + observed_extra_fee,
+                )
+            if shortfall is None or attempt >= _MAX_SWAP_ATTEMPTS or recomputed <= 0:
+                logger.error(
+                    "swap_to_primary_mint: melt failed",
+                    extra={
+                        "error": str(e),
+                        "error_type": type(e).__name__,
+                        "foreign_mint": token_obj.mint,
+                        "token_amount": token_amount,
+                        "melt_quote_id": melt_quote.quote,
+                        "total_needed": total_needed,
+                        "attempts": attempt,
+                    },
+                )
+                raise ValueError(
+                    f"Failed to melt token from foreign mint {token_obj.mint}: {e}"
+                ) from e
+            logger.warning(
+                "swap_to_primary_mint: mint demanded more than quoted at melt, retrying",
+                extra={
+                    "shortfall": shortfall,
+                    "retry_minted_amount": recomputed,
+                    "attempt": attempt,
+                },
+            )
+            minted_amount = recomputed
+            continue
+
+        break
 
     logger.info(
         "swap_to_primary_mint: melt succeeded, minting on primary",

tests/integration/test_free_response_stale_reservation.py

@@ -0,0 +1,146 @@
+"""Regression tests for charging after stale reservation cleanup."""
+
+import time
+import uuid
+from unittest.mock import patch
+
+import pytest
+from sqlmodel.ext.asyncio.session import AsyncSession
+
+from routstr.core.db import ApiKey
+from routstr.payment.cost_calculation import CostData
+
+
+def _make_key(balance: int, reserved: int) -> ApiKey:
+    return ApiKey(
+        hashed_key=f"test_{uuid.uuid4().hex}",
+        balance=balance,
+        reserved_balance=reserved,
+        total_spent=0,
+        total_requests=1,
+    )
+
+
+def _cost_data(total_msats: int) -> CostData:
+    return CostData(
+        base_msats=0,
+        input_msats=total_msats // 2,
+        output_msats=total_msats - total_msats // 2,
+        total_msats=total_msats,
+        total_usd=0.0,
+        input_tokens=100,
+        output_tokens=100,
+    )
+
+
+@pytest.mark.asyncio
+async def test_overrun_charges_after_reservation_swept(
+    integration_session: AsyncSession,
+) -> None:
+    """Overrun finalize must charge even when the reservation was already released."""
+    from routstr.auth import adjust_payment_for_tokens
+
+    deducted_max_cost = 990  # discounted reservation
+    actual_token_cost = 1000  # actual cost overruns the reservation
+
+    # Sweeper has zeroed reserved_balance but left balance untouched.
+    key = _make_key(balance=1000, reserved=0)
+    integration_session.add(key)
+    await integration_session.commit()
+
+    response_data = {
+        "model": "test-model",
+        "usage": {"prompt_tokens": 100, "completion_tokens": 100},
+    }
+
+    with patch(
+        "routstr.auth.calculate_cost",
+        return_value=_cost_data(actual_token_cost),
+    ):
+        await adjust_payment_for_tokens(
+            key, response_data, integration_session, deducted_max_cost
+        )
+
+    await integration_session.refresh(key)
+
+    assert key.total_spent == actual_token_cost, (
+        f"Request was not billed (total_spent={key.total_spent}) — free response bug"
+    )
+    assert key.balance == 1000 - actual_token_cost, (
+        f"Balance not charged: {key.balance}"
+    )
+    assert key.balance >= 0
+    assert key.reserved_balance == 0
+
+
+@pytest.mark.asyncio
+async def test_free_response_path_closed_end_to_end(
+    integration_session: AsyncSession,
+    patched_db_engine: None,
+) -> None:
+    """A reservation released by the real sweeper must not yield a free response."""
+    from routstr.auth import adjust_payment_for_tokens, pay_for_request
+    from routstr.core.db import create_session, release_stale_reservations
+
+    deducted_max_cost = 990
+    actual_token_cost = 1000
+    key_hash = f"test_sweep_{uuid.uuid4().hex}"
+
+    async with create_session() as session:
+        session.add(
+            ApiKey(
+                hashed_key=key_hash,
+                balance=1000,
+                reserved_balance=0,
+                total_spent=0,
+                total_requests=0,
+            )
+        )
+        await session.commit()
+
+    # Reserve the request, then backdate reserved_at so the sweeper treats it as
+    # stale (simulates a stream that outlived stale_reservation_timeout_seconds).
+    async with create_session() as session:
+        key = await session.get(ApiKey, key_hash)
+        assert key is not None
+        await pay_for_request(key, deducted_max_cost, session)
+        await session.refresh(key)
+        assert key.reserved_balance == deducted_max_cost
+        key.reserved_at = int(time.time()) - 10_000
+        session.add(key)
+        await session.commit()
+
+    # Sweeper releases the stale reservation without charging.
+    async with create_session() as session:
+        released = await release_stale_reservations(session, max_age_seconds=300)
+        assert released == 1
+
+    async with create_session() as session:
+        key = await session.get(ApiKey, key_hash)
+        assert key is not None
+        assert key.reserved_balance == 0, "Precondition: sweeper zeroed the reservation"
+
+        response_data = {
+            "model": "test-model",
+            "usage": {"prompt_tokens": 100, "completion_tokens": 100},
+        }
+        with patch(
+            "routstr.auth.calculate_cost",
+            return_value=_cost_data(actual_token_cost),
+        ):
+            await adjust_payment_for_tokens(
+                key, response_data, session, deducted_max_cost
+            )
+
+    async with create_session() as session:
+        final = await session.get(ApiKey, key_hash)
+        assert final is not None
+
+    assert final.total_spent == actual_token_cost, (
+        f"Free response: total_spent={final.total_spent}, expected {actual_token_cost}"
+    )
+    assert final.balance == 1000 - actual_token_cost, (
+        f"Balance not charged after sweep: {final.balance}"
+    )
+    assert final.balance >= 0
+    assert final.reserved_balance == 0

tests/integration/test_swap_fee_retry.py

@@ -0,0 +1,194 @@
+"""
+Integration tests for reactive swap fee retries via the wallet topup endpoint.
+
+Foreign-mint tokens are swapped to the primary mint using the foreign mint's
+melt quote, whose fee_reserve is a non-binding estimate (NUT-05): the mint may
+demand more when re-quoting or at melt execution. These tests cover the
+endpoint behaviour in those cases:
+
+1. The mint demands one sat more at melt time than every quote reported
+   (the mint.cubabitcoin.org incident): the swap retries with a smaller
+   invoice and the topup succeeds, crediting the recomputed amount.
+2. The real melt quote reports a higher fee_reserve than the estimate: the
+   swap re-quotes from the observed fee and the topup succeeds.
+3. The mint escalates its fee demands on every attempt: the retry budget is
+   exhausted and the endpoint returns 400 with a clear error (never 500),
+   without ever executing a melt.
+"""
+
+from collections.abc import Callable
+from unittest.mock import AsyncMock, Mock, patch
+
+import pytest
+from httpx import AsyncClient, Response
+
+from routstr.core.settings import settings
+
+# Captured at collection time, before the integration_app fixture replaces it
+# with the testmint stub that bypasses swapping (see conftest.py).
+from routstr.wallet import recieve_token as _real_recieve_token
+
+PRIMARY_MINT = "http://primary:3338"
+
+
+def _make_swap_mocks(
+    token_amount: int,
+    fee_reserves: list[int],
+    input_fees: int = 0,
+    mint_url: str = "http://foreign-mint:3338",
+) -> tuple[Mock, Mock, Mock]:
+    """Return (token, token_wallet, primary_wallet) mocks that act like a mint.
+
+    Mint quotes pass the requested amount through their ``request`` field and
+    melt quotes echo that amount back, so the mocks stay consistent for
+    whatever amounts the implementation requests. ``fee_reserves`` supplies the
+    fee_reserve of each successive melt quote (the first serves the estimation
+    pass); requesting more quotes than provided fails the test.
+    """
+    mock_token = Mock()
+    mock_token.mint = mint_url
+    mock_token.unit = "sat"
+    mock_token.amount = token_amount
+    mock_token.keysets = ["keyset1"]
+    mock_token.proofs = [Mock(amount=token_amount)]
+
+    mock_token_wallet = Mock()
+    mock_token_wallet.load_mint = AsyncMock()
+    mock_token_wallet.load_proofs = AsyncMock()
+    mock_token_wallet.get_fees_for_proofs = Mock(return_value=input_fees)
+
+    mock_primary_wallet = Mock()
+    mock_primary_wallet.load_mint = AsyncMock()
+    mock_primary_wallet.load_proofs = AsyncMock()
+    mock_primary_wallet.available_balance = Mock(amount=0)
+    mock_primary_wallet.mint = AsyncMock(return_value=Mock())
+
+    fees = iter(fee_reserves)
+
+    def _next_fee() -> int:
+        try:
+            return next(fees)
+        except StopIteration:
+            raise AssertionError(
+                "more melt quotes requested than fee_reserves provided"
+            ) from None
+
+    mock_primary_wallet.request_mint = AsyncMock(
+        side_effect=lambda amount: Mock(quote=f"mint_quote_{amount}", request=amount)
+    )
+    mock_token_wallet.melt_quote = AsyncMock(
+        side_effect=lambda invoice: Mock(
+            quote=f"melt_quote_{invoice}", amount=invoice, fee_reserve=_next_fee()
+        )
+    )
+    mock_token_wallet.melt = AsyncMock(return_value=Mock())
+
+    return mock_token, mock_token_wallet, mock_primary_wallet
+
+
+def _wallet_router(primary_wallet: Mock, token_wallet: Mock) -> Callable[..., Mock]:
+    """Route get_wallet calls to the primary or foreign wallet mock by URL."""
+
+    def fake_get_wallet(mint_url: str, unit: str = "sat", load: bool = True) -> Mock:
+        return primary_wallet if mint_url == PRIMARY_MINT else token_wallet
+
+    return fake_get_wallet
+
+
+async def _post_topup(
+    client: AsyncClient,
+    mock_token: Mock,
+    token_wallet: Mock,
+    primary_wallet: Mock,
+) -> Response:
+    """POST /v1/wallet/topup with the swap layer mocked at the mint boundary.
+
+    The conftest's testmint stub for recieve_token is swapped back for the
+    real implementation so the request exercises the actual swap path.
+    """
+    with patch("routstr.wallet.recieve_token", _real_recieve_token):
+        with patch(
+            "routstr.wallet.deserialize_token_from_string", return_value=mock_token
+        ):
+            with patch(
+                "routstr.wallet.get_wallet",
+                side_effect=_wallet_router(primary_wallet, token_wallet),
+            ):
+                with patch.object(settings, "primary_mint", PRIMARY_MINT):
+                    with patch.object(settings, "primary_mint_unit", "sat"):
+                        with patch.object(settings, "cashu_mints", [PRIMARY_MINT]):
+                            return await client.post(
+                                "/v1/wallet/topup",
+                                params={"cashu_token": "cashuAtest_foreign_token"},
+                            )
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_melt_demands_more_than_quoted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 179-sat token where every quote reports fee_reserve=1 but the mint
+    rejects the first melt demanding 180. The retry shrinks the invoice to 177
+    and the topup credits 177 sats (177_000 msats)."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[1, 1, 1], mint_url="http://mint.cubabitcoin.org"
+    )
+    token_wallet.melt.side_effect = [
+        Exception(
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 179, needed: 180 (Code: 11000)"
+        ),
+        Mock(),
+    ]
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 177_000
+    assert token_wallet.melt.call_count == 2
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_retries_when_quote_fee_exceeds_estimate(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A 1000-sat token estimated at fee 20, but the real quote demands 23.
+    The retry recomputes 1000 - 23 = 977, which fits, and the topup credits
+    977 sats (977_000 msats) with a single melt."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[20, 23, 23]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 200
+    assert response.json()["msats"] == 977_000
+    assert token_wallet.melt.call_count == 1
+
+
+@pytest.mark.integration
+@pytest.mark.asyncio
+async def test_topup_returns_400_when_retries_exhausted(
+    authenticated_client: AsyncClient,
+) -> None:
+    """A mint that escalates fee_reserve on every re-quote (1 → 10 → 25 → 50)
+    exhausts the retry budget: clean 400 with an actionable message, melt never
+    executed."""
+    mock_token, token_wallet, primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[1, 10, 25, 50]
+    )
+
+    response = await _post_topup(
+        authenticated_client, mock_token, token_wallet, primary_wallet
+    )
+
+    assert response.status_code == 400
+    assert "too small to cover swap fees" in response.json()["detail"]
+    assert token_wallet.melt_quote.call_count == 4  # estimation + 3 attempts
+    token_wallet.melt.assert_not_called()

tests/unit/test_require_parameters.py

@@ -0,0 +1,95 @@
+import json
+import os
+
+os.environ.setdefault("UPSTREAM_BASE_URL", "http://test")
+os.environ.setdefault("UPSTREAM_API_KEY", "test")
+os.environ.setdefault("LIGHTNING_ADDRESS", "test@stm.to")
+
+from routstr.upstream import GenericUpstreamProvider
+from routstr.upstream.openrouter import OpenRouterUpstreamProvider
+
+
+def _model(model_id: str = "openai/gpt-4o"):  # type: ignore[no-untyped-def]
+    from routstr.payment.models import Architecture, Model, Pricing
+
+    return Model(
+        id=model_id,
+        name=model_id,
+        created=0,
+        description="",
+        context_length=128000,
+        architecture=Architecture(
+            modality="text->text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="GPT",
+            instruct_type=None,
+        ),
+        pricing=Pricing(prompt=0.0, completion=0.0),
+    )
+
+
+def _tool_body() -> dict:
+    return {
+        "model": "openai/gpt-4o",
+        "messages": [{"role": "user", "content": "What's the weather?"}],
+        "tools": [
+            {
+                "type": "function",
+                "function": {"name": "get_weather", "parameters": {}},
+            }
+        ],
+    }
+
+
+def _prepare(provider, body: dict) -> dict:  # type: ignore[no-untyped-def]
+    out = provider.prepare_request_body(json.dumps(body).encode(), _model())
+    assert out is not None
+    return json.loads(out)
+
+
+def test_injects_require_parameters_for_tool_request() -> None:
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), _tool_body())
+    assert data["provider"]["require_parameters"] is True
+
+
+def test_generic_provider_on_openrouter_url_is_left_alone() -> None:
+    # Only OpenRouterUpstreamProvider injects; a generic provider pointed at the
+    # same base URL doesn't.
+    provider = GenericUpstreamProvider(base_url="https://openrouter.ai/api/v1")
+    data = _prepare(provider, _tool_body())
+    assert "provider" not in data
+
+
+def test_no_injection_without_tools() -> None:
+    body = {"model": "openai/gpt-4o", "messages": [{"role": "user", "content": "hi"}]}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert "provider" not in data
+
+
+def test_empty_tools_list_does_not_inject() -> None:
+    body = _tool_body()
+    body["tools"] = []
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert "provider" not in data
+
+
+def test_direct_provider_does_not_inject() -> None:
+    provider = GenericUpstreamProvider(base_url="https://api.openai.com/v1")
+    data = _prepare(provider, _tool_body())
+    assert "provider" not in data
+
+
+def test_keeps_client_set_require_parameters() -> None:
+    body = _tool_body()
+    body["provider"] = {"require_parameters": False}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert data["provider"]["require_parameters"] is False
+
+
+def test_preserves_other_provider_fields() -> None:
+    body = _tool_body()
+    body["provider"] = {"order": ["openai", "azure"]}
+    data = _prepare(OpenRouterUpstreamProvider(api_key="test"), body)
+    assert data["provider"]["order"] == ["openai", "azure"]
+    assert data["provider"]["require_parameters"] is True

tests/unit/test_upstream_rate_limit.py

@@ -0,0 +1,396 @@
+"""Tests for upstream rate-limit detection, classification, and org-ID redaction.
+
+Covers issue #555: upstream OpenAI-compatible providers return rate-limit
+errors that embed a sensitive organization ID. The proxy must classify these
+distinctly (``UPSTREAM_RATE_LIMIT``), preserve useful debugging fields, and
+never emit a raw ``org-*`` identifier in logs, errors, or returned bodies.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock, Mock, patch
+
+import httpx
+import pytest
+
+from routstr.core.redaction import redact_org_ids
+from routstr.upstream.base import BaseUpstreamProvider
+from routstr.upstream.rate_limit import (
+    UPSTREAM_RATE_LIMIT,
+    RateLimitInfo,
+    classify_rate_limit,
+)
+
+# The exact scenario from the issue, with a realistic (fake) org identifier.
+RAW_ORG_ID = "org-abc123XYZ456def"
+RATE_LIMIT_MESSAGE = (
+    f"Rate limit reached for gpt-5.5-2026-04-23 (for limit gpt-5.5) in "
+    f"organization {RAW_ORG_ID} on tokens per min (TPM): Limit 180000000, "
+    f"Used 180000000, Requested 8929. Please try again in 2ms. Visit "
+    f"https://platform.openai.com/account/rate-limits to learn more."
+)
+
+
+def _make_request(request_id: str = "req-123") -> Mock:
+    request = Mock(spec=["method", "state"])
+    request.method = "POST"
+    request.state = Mock()
+    request.state.request_id = request_id
+    return request
+
+
+def _make_upstream_response(
+    *,
+    body: bytes,
+    status_code: int = 429,
+    content_type: str | None = "application/json",
+    extra_headers: dict[str, str] | None = None,
+) -> httpx.Response:
+    headers: dict[str, str] = {}
+    if content_type is not None:
+        headers["content-type"] = content_type
+    if extra_headers:
+        headers.update(extra_headers)
+    return httpx.Response(status_code=status_code, headers=headers, content=body)
+
+
+@pytest.fixture
+def provider() -> BaseUpstreamProvider:
+    return BaseUpstreamProvider(
+        base_url="https://privateprovider.xyz", api_key="k", provider_fee=1.0
+    )
+
+
+# --------------------------------------------------------------------------- #
+# Redaction
+# --------------------------------------------------------------------------- #
+
+
+def test_redact_org_ids_replaces_identifier() -> None:
+    assert RAW_ORG_ID not in redact_org_ids(RATE_LIMIT_MESSAGE)
+    assert "org-[REDACTED]" in redact_org_ids(RATE_LIMIT_MESSAGE)
+
+
+def test_redact_org_ids_is_idempotent() -> None:
+    once = redact_org_ids(RATE_LIMIT_MESSAGE)
+    assert redact_org_ids(once) == once
+
+
+def test_redact_org_ids_leaves_unrelated_text() -> None:
+    assert redact_org_ids("organize the org-chart") == "organize the org-chart"
+    assert redact_org_ids("") == ""
+
+
+# --------------------------------------------------------------------------- #
+# Classification
+# --------------------------------------------------------------------------- #
+
+
+def test_classify_exact_scenario() -> None:
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert isinstance(info, RateLimitInfo)
+    assert info.code == UPSTREAM_RATE_LIMIT
+    assert info.model == "gpt-5.5-2026-04-23"
+    assert info.limit_name == "gpt-5.5"
+    assert info.metric == "tokens per min (TPM)"
+    assert info.limit == 180000000
+    assert info.used == 180000000
+    assert info.requested == 8929
+    assert info.retry_after_seconds == pytest.approx(0.002)
+    # Redaction-safe: no raw org id survives into the structured view.
+    assert RAW_ORG_ID not in info.message
+    assert RAW_ORG_ID not in json.dumps(info.as_details())
+
+
+def test_classify_by_status_code_without_marker() -> None:
+    info = classify_rate_limit(429, "slow down")
+    assert info is not None
+    assert info.code == UPSTREAM_RATE_LIMIT
+
+
+def test_classify_by_message_marker_without_429() -> None:
+    info = classify_rate_limit(400, "rate_limit_exceeded for this key")
+    assert info is not None
+
+
+def test_retry_after_header_takes_precedence() -> None:
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE, {"Retry-After": "12"})
+    assert info is not None
+    assert info.retry_after_seconds == pytest.approx(12.0)
+
+
+def test_non_rate_limit_error_is_not_classified() -> None:
+    assert classify_rate_limit(400, "invalid request: missing field 'model'") is None
+    assert classify_rate_limit(500, "internal server error") is None
+
+
+# --------------------------------------------------------------------------- #
+# forward_upstream_error_response integration
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_json_rate_limit_body_is_redacted_and_forwarded(
+    provider: BaseUpstreamProvider,
+) -> None:
+    body = json.dumps(
+        {"error": {"message": RATE_LIMIT_MESSAGE, "type": "rate_limit_exceeded"}}
+    ).encode()
+    upstream = _make_upstream_response(body=body, status_code=429)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 429
+    raw = bytes(response.body).decode()
+    # No raw organization id may survive in the forwarded body.
+    assert RAW_ORG_ID not in raw
+    assert "org-[REDACTED]" in raw
+    # Body remains valid JSON; the original type is preserved while a stable
+    # rate-limit code is injected so callers can switch on it.
+    payload: dict[str, Any] = json.loads(raw)
+    assert payload["error"]["type"] == "rate_limit_exceeded"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["model"] == "gpt-5.5-2026-04-23"
+    # A retry hint extracted from the message is surfaced as a header.
+    assert "retry-after" in {k.lower() for k in response.headers}
+
+
+@pytest.mark.asyncio
+async def test_non_json_rate_limit_envelope_uses_stable_code(
+    provider: BaseUpstreamProvider,
+) -> None:
+    upstream = _make_upstream_response(
+        body=RATE_LIMIT_MESSAGE.encode(),
+        status_code=429,
+        content_type="text/plain",
+    )
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["model"] == "gpt-5.5-2026-04-23"
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+@pytest.mark.asyncio
+async def test_non_rate_limit_json_error_unchanged(
+    provider: BaseUpstreamProvider,
+) -> None:
+    body = json.dumps(
+        {"error": {"message": "missing field 'model'", "type": "invalid_request"}}
+    ).encode()
+    upstream = _make_upstream_response(body=body, status_code=400)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 400
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "invalid_request"
+    assert "retry-after" not in {k.lower() for k in response.headers}
+
+
+# --------------------------------------------------------------------------- #
+# UpstreamError -> proxy response (preserves code/details/status)
+# --------------------------------------------------------------------------- #
+
+
+def test_create_upstream_error_response_preserves_structure() -> None:
+    from routstr.core.exceptions import UpstreamError
+    from routstr.payment.helpers import create_upstream_error_response
+
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert info is not None
+    err = UpstreamError(
+        f"Upstream error via litellm: {RATE_LIMIT_MESSAGE}",
+        status_code=429,
+        code=info.code,
+        details=info.as_details(),
+    )
+
+    response = create_upstream_error_response(err, _make_request())
+
+    # Original upstream status is preserved (not flattened to 502).
+    assert response.status_code == 429
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["requested"] == 8929
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+def test_generic_upstream_error_still_defaults_to_502() -> None:
+    from routstr.core.exceptions import UpstreamError
+    from routstr.payment.helpers import create_upstream_error_response
+
+    err = UpstreamError("connection refused")  # status_code defaults to 502
+
+    response = create_upstream_error_response(err, _make_request())
+
+    assert response.status_code == 502
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == 502
+    assert "details" not in payload["error"]
+
+
+# --------------------------------------------------------------------------- #
+# Structured log-extra redaction
+# --------------------------------------------------------------------------- #
+
+
+def test_security_filter_redacts_org_id_in_extra() -> None:
+    import logging
+
+    from routstr.core.logging import SecurityFilter
+
+    record = logging.LogRecord(
+        name="test",
+        level=logging.ERROR,
+        pathname=__file__,
+        lineno=1,
+        msg="upstream failed",
+        args=(),
+        exc_info=None,
+    )
+    # Simulate an ``extra={"body_preview": ...}`` field carrying an org id.
+    setattr(record, "body_preview", RATE_LIMIT_MESSAGE)
+
+    assert SecurityFilter().filter(record) is True
+    redacted: str = getattr(record, "body_preview")
+    assert RAW_ORG_ID not in redacted
+    assert "org-[REDACTED]" in redacted
+
+
+def test_security_filter_redacts_org_id_in_nested_extra() -> None:
+    import logging
+
+    from routstr.core.logging import SecurityFilter
+
+    record = logging.LogRecord(
+        name="test",
+        level=logging.ERROR,
+        pathname=__file__,
+        lineno=1,
+        msg="upstream failed",
+        args=(),
+        exc_info=None,
+    )
+    # Nested structures: dict containing a list containing the org id.
+    setattr(record, "body", {"error": {"messages": [RATE_LIMIT_MESSAGE]}})
+
+    assert SecurityFilter().filter(record) is True
+    serialized = json.dumps(getattr(record, "body"))
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+# --------------------------------------------------------------------------- #
+# 5xx-wrapped rate limit through forward_upstream_error_response
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_5xx_wrapped_rate_limit_is_classified(
+    provider: BaseUpstreamProvider,
+) -> None:
+    # Some providers wrap a rate-limit in a 5xx envelope; classification must
+    # key off the message marker, not only the 429 status.
+    body = json.dumps({"error": {"message": RATE_LIMIT_MESSAGE}}).encode()
+    upstream = _make_upstream_response(body=body, status_code=500)
+
+    response = await provider.forward_upstream_error_response(
+        _make_request(), "v1/chat/completions", upstream
+    )
+
+    assert response.status_code == 500
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+
+
+# --------------------------------------------------------------------------- #
+# Real proxy loop: structured error surfaced + reservation reverted once
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio
+async def test_proxy_loop_surfaces_rate_limit_and_reverts_once() -> None:
+    from routstr import proxy as proxy_module
+    from routstr.core.db import ApiKey
+    from routstr.core.exceptions import UpstreamError
+
+    info = classify_rate_limit(429, RATE_LIMIT_MESSAGE)
+    assert info is not None
+
+    key = ApiKey(hashed_key="rlkey", balance=10_000)
+
+    request = MagicMock()
+    request.method = "POST"
+    request.headers = {"authorization": "Bearer sk-rlkey"}
+    request.body = AsyncMock(return_value=b'{"model": "test-model"}')
+    request.state = MagicMock()
+    request.state.request_id = "req-rl"
+
+    upstream = MagicMock()
+    upstream.provider_type = "test"
+    upstream.prepare_headers = MagicMock(side_effect=lambda h: h)
+    upstream.forward_request = AsyncMock(
+        side_effect=UpstreamError(
+            f"Upstream error via litellm: {RATE_LIMIT_MESSAGE}",
+            status_code=429,
+            code=info.code,
+            details=info.as_details(),
+        )
+    )
+
+    session = MagicMock()
+    revert_mock = AsyncMock(return_value=True)
+
+    with (
+        patch.object(proxy_module, "get_model_instance", return_value=MagicMock()),
+        patch.object(proxy_module, "get_provider_for_model", return_value=[upstream]),
+        patch.object(
+            proxy_module, "get_max_cost_for_model", AsyncMock(return_value=1_000)
+        ),
+        patch.object(
+            proxy_module,
+            "calculate_discounted_max_cost",
+            AsyncMock(return_value=1_000),
+        ),
+        patch.object(proxy_module, "check_token_balance", MagicMock()),
+        patch.object(
+            proxy_module, "get_bearer_token_key", AsyncMock(return_value=key)
+        ),
+        patch.object(proxy_module, "pay_for_request", AsyncMock(return_value=1_000)),
+        patch.object(proxy_module, "revert_pay_for_request", revert_mock),
+    ):
+        response = await proxy_module.proxy(
+            request, "v1/chat/completions", session=session
+        )
+
+    # Original 429 status and the stable code/details survive to the client.
+    assert response.status_code == 429
+    payload: dict[str, Any] = json.loads(bytes(response.body))
+    assert payload["error"]["type"] == "upstream_error"
+    assert payload["error"]["code"] == UPSTREAM_RATE_LIMIT
+    assert payload["error"]["details"]["requested"] == 8929
+    serialized = json.dumps(payload)
+    assert RAW_ORG_ID not in serialized
+    assert "org-[REDACTED]" in serialized
+    # Single upstream failed -> reservation reverted exactly once (no double-charge).
+    revert_mock.assert_awaited_once_with(key, session, 1_000)

tests/unit/test_wallet.py

@@ -241,7 +241,9 @@ async def test_credit_balance_rejects_missing_key() -> None:
 
 @pytest.mark.asyncio
 async def test_swap_to_primary_mint_insufficient_for_fees() -> None:
-    """Token amount is less than melt_quote.amount + melt_quote.fee_reserve."""
+    """Token amount is less than melt_quote.amount + melt_quote.fee_reserve.
+    The quote mocks are static, so every retry observes the same shortfall —
+    the swap must still give up and raise."""
     from routstr.wallet import swap_to_primary_mint
 
     mock_token = Mock()
@@ -283,50 +285,6 @@ async def test_swap_to_primary_mint_insufficient_for_fees() -> None:
     mock_token_wallet.melt.assert_not_called()
 
 
-@pytest.mark.asyncio
-async def test_swap_to_primary_mint_melt_error_wrapped() -> None:
-    """Melt failure from cashu lib is wrapped as ValueError."""
-    from routstr.wallet import swap_to_primary_mint
-
-    mock_token = Mock()
-    mock_token.mint = "http://foreign:3338"
-    mock_token.unit = "sat"
-    mock_token.amount = 5000
-    mock_token.keysets = ["keyset1"]
-    mock_token.proofs = [{"amount": 5000}]
-
-    mock_token_wallet = Mock()
-    mock_token_wallet.load_mint = AsyncMock()
-    mock_token_wallet.load_proofs = AsyncMock()
-    mock_token_wallet.get_fees_for_proofs = Mock(return_value=0)
-
-    mock_primary_wallet = Mock()
-    mock_primary_wallet.load_mint = AsyncMock()
-    mock_primary_wallet.load_proofs = AsyncMock()
-
-    mock_mint_quote = Mock()
-    mock_mint_quote.quote = "mint_quote_456"
-    mock_mint_quote.request = "lnbc1..."
-    mock_primary_wallet.request_mint = AsyncMock(return_value=mock_mint_quote)
-
-    mock_melt_quote = Mock()
-    mock_melt_quote.quote = "melt_quote_456"
-    mock_melt_quote.amount = 4940
-    mock_melt_quote.fee_reserve = 50  # total 4990 < 5000, passes fee check
-    mock_token_wallet.melt_quote = AsyncMock(return_value=mock_melt_quote)
-    mock_token_wallet.melt = AsyncMock(
-        side_effect=Exception("Provided: 5000, needed: 5100 (Code: 11000)")
-    )
-
-    from routstr.core.settings import settings
-
-    with patch.object(settings, "primary_mint", "http://primary:3338"):
-        with patch.object(settings, "primary_mint_unit", "sat"):
-            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
-                with pytest.raises(ValueError, match="Failed to melt token"):
-                    await swap_to_primary_mint(mock_token, mock_token_wallet)
-
-
 @pytest.mark.asyncio
 async def test_recieve_token_untrusted_mint() -> None:
     mock_wallet = Mock()
@@ -394,54 +352,80 @@ async def test_swap_to_primary_mint_already_on_primary() -> None:
     mock_token_wallet.melt_quote.assert_not_called()
 
 
-async def test_swap_to_primary_mint_success() -> None:
-    """Test successful swap with dynamic fee calculation."""
-    from routstr.wallet import swap_to_primary_mint
+# ---------------------------------------------------------------------------
+# Swap fee estimation and reactive retry
+#
+# Spec: the estimation pass subtracts only observed fees (no safety buffer).
+# swap_to_primary_mint then runs the mint-quote/melt-quote/melt cycle and, when
+# the foreign mint demands more than estimated (at quote or at melt time),
+# retries with the amount recomputed from the observed fee — at most 3 attempts.
+# Melt failures unrelated to fees are not retried.
+# ---------------------------------------------------------------------------
 
+
+def _make_swap_mocks(
+    token_amount: int,
+    fee_reserves: list[int],
+    input_fees: int = 0,
+    mint_url: str = "http://foreign-mint:3338",
+) -> tuple[Mock, Mock, Mock]:
+    """Return (token, token_wallet, primary_wallet) mocks that act like a mint.
+
+    Mint quotes pass the requested amount through their ``request`` field and
+    melt quotes echo that amount back, so the mocks stay consistent for
+    whatever amounts the implementation requests. ``fee_reserves`` supplies the
+    fee_reserve of each successive melt quote (the first serves the estimation
+    pass); requesting more quotes than provided fails the test.
+    """
     mock_token = Mock()
-    mock_token.mint = "http://foreign:3338"
+    mock_token.mint = mint_url
     mock_token.unit = "sat"
-    mock_token.amount = 1000
+    mock_token.amount = token_amount
     mock_token.keysets = ["keyset1"]
-    mock_token.proofs = [{"amount": 1000}]
+    mock_token.proofs = [Mock(amount=token_amount)]
 
     mock_token_wallet = Mock()
     mock_token_wallet.load_mint = AsyncMock()
     mock_token_wallet.load_proofs = AsyncMock()
-    mock_token_wallet.get_fees_for_proofs = Mock(return_value=0)
+    mock_token_wallet.get_fees_for_proofs = Mock(return_value=input_fees)
 
     mock_primary_wallet = Mock()
     mock_primary_wallet.load_mint = AsyncMock()
     mock_primary_wallet.load_proofs = AsyncMock()
+    mock_primary_wallet.available_balance = Mock(amount=0)
+    mock_primary_wallet.mint = AsyncMock(return_value=Mock())
 
-    # Mocks for the estimation phase
-    # 1. request_mint(dummy_amount=1000) -> invoice_dummy
-    # 2. melt_quote(invoice_dummy) -> fee=10
+    fees = iter(fee_reserves)
 
-    # Mocks for the execution phase
-    # 3. request_mint(minted_amount=990) -> invoice_real
-    # 4. melt_quote(invoice_real) -> amount=990, fee=10
-    # 5. melt() -> success
-    # 6. mint() -> success
+    def _next_fee() -> int:
+        try:
+            return next(fees)
+        except StopIteration:
+            raise AssertionError(
+                "more melt quotes requested than fee_reserves provided"
+            ) from None
 
-    mock_mint_quote_dummy = Mock(quote="dummy_quote", request="lnbc_dummy")
-    mock_mint_quote_real = Mock(quote="real_quote", request="lnbc_real")
-
-    # side_effect for request_mint to return dummy then real
     mock_primary_wallet.request_mint = AsyncMock(
-        side_effect=[mock_mint_quote_dummy, mock_mint_quote_real]
+        side_effect=lambda amount: Mock(quote=f"mint_quote_{amount}", request=amount)
     )
-
-    mock_melt_quote_dummy = Mock(amount=1000, fee_reserve=10)
-    mock_melt_quote_real = Mock(amount=990, fee_reserve=10)
-
-    # side_effect for melt_quote
     mock_token_wallet.melt_quote = AsyncMock(
-        side_effect=[mock_melt_quote_dummy, mock_melt_quote_real]
+        side_effect=lambda invoice: Mock(
+            quote=f"melt_quote_{invoice}", amount=invoice, fee_reserve=_next_fee()
+        )
     )
+    mock_token_wallet.melt = AsyncMock(return_value=Mock())
 
-    mock_token_wallet.melt = AsyncMock(return_value="melted_proofs")
-    mock_primary_wallet.mint = AsyncMock(return_value="minted_proofs")
+    return mock_token, mock_token_wallet, mock_primary_wallet
+
+
+@pytest.mark.asyncio
+async def test_swap_to_primary_mint_success() -> None:
+    """No retry needed: real quote matches the estimate, full net amount minted."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
 
     from routstr.core.settings import settings
 
@@ -452,17 +436,659 @@ async def test_swap_to_primary_mint_success() -> None:
                     mock_token, mock_token_wallet
                 )
 
-                assert amount == 990  # 1000 - 10
-                assert unit == "sat"
-                assert mint == "http://primary:3338"
+    assert amount == 990  # 1000 - fee_reserve(10), no buffer subtracted
+    assert unit == "sat"
+    assert mint == "http://primary:3338"
+    assert mock_primary_wallet.request_mint.call_count == 2
+    mock_primary_wallet.request_mint.assert_any_call(1000)
+    mock_primary_wallet.request_mint.assert_any_call(990)
+    assert mock_token_wallet.melt_quote.call_count == 2
+    assert mock_token_wallet.melt.call_count == 1
+    assert mock_primary_wallet.mint.called
 
-                # Verify call order/counts
-                assert mock_primary_wallet.request_mint.call_count == 2
-                # First call with full amount for estimation
-                mock_primary_wallet.request_mint.assert_any_call(1000)
-                # Second call with calculated amount
-                mock_primary_wallet.request_mint.assert_any_call(990)
 
-                assert mock_token_wallet.melt_quote.call_count == 2
-                assert mock_token_wallet.melt.called
-                assert mock_primary_wallet.mint.called
+@pytest.mark.asyncio
+@pytest.mark.parametrize("fee_reserve", [1, 10, 100])
+async def test_calculate_swap_amount_subtracts_only_observed_fees(
+    fee_reserve: int,
+) -> None:
+    """Estimation: minted_amount = token - fee_reserve, with no safety buffer."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[fee_reserve]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            result = await _calculate_swap_amount(
+                amount_msat=1_000_000,
+                token_unit="sat",
+                token_mint_url="http://foreign-mint:3338",
+                token_wallet=mock_token_wallet,
+                primary_wallet=mock_primary_wallet,
+                proofs=[],
+            )
+
+    assert result == 1000 - fee_reserve
+
+
+@pytest.mark.asyncio
+async def test_calculate_swap_amount_includes_input_fees() -> None:
+    """Estimation subtracts NUT-02 input fees alongside the melt fee_reserve."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        500, fee_reserves=[10], input_fees=3
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            result = await _calculate_swap_amount(
+                amount_msat=500_000,
+                token_unit="sat",
+                token_mint_url="http://foreign-mint:3338",
+                token_wallet=mock_token_wallet,
+                primary_wallet=mock_primary_wallet,
+                proofs=[],
+            )
+
+    assert result == 487  # 500 - 10 - 3
+
+
+@pytest.mark.asyncio
+async def test_swap_retries_when_real_quote_exceeds_estimate() -> None:
+    """The real melt quote demands a higher fee than the estimate (20 → 23).
+    Instead of failing, the swap recomputes the amount from the observed fee
+    and re-quotes: 1000 - 23 = 977, which fits (977 + 23 <= 1000)."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[20, 23, 23]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert amount == 977
+    assert unit == "sat"
+    mock_primary_wallet.request_mint.assert_any_call(980)
+    mock_primary_wallet.request_mint.assert_any_call(977)
+    assert mock_token_wallet.melt_quote.call_count == 3  # estimation + 2 attempts
+    assert mock_token_wallet.melt.call_count == 1
+
+
+@pytest.mark.asyncio
+async def test_swap_retries_when_melt_demands_more_than_quoted() -> None:
+    """The mint.cubabitcoin.org incident: every quote reports fee_reserve=1,
+    but the mint demands 2 sats at melt time ("Provided: 179, needed: 180").
+    The swap must retry with a smaller invoice (177) so the second melt fits,
+    instead of failing the topup."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[1, 1, 1], mint_url="http://mint.cubabitcoin.org"
+    )
+    mock_token_wallet.melt.side_effect = [
+        Exception(
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 179, needed: 180 (Code: 11000)"
+        ),
+        Mock(),
+    ]
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert amount == 177  # 179 - 1 (estimate) - 1 (observed melt shortfall)
+    assert mock_token_wallet.melt.call_count == 2
+    mock_primary_wallet.request_mint.assert_any_call(178)
+    mock_primary_wallet.request_mint.assert_any_call(177)
+
+
+@pytest.mark.asyncio
+async def test_swap_retries_on_cdk_unbalanced_error() -> None:
+    """cdk-based mints report insufficient melt inputs as the registered code
+    11005 (TransactionUnbalanced) with their own message wording — no
+    Provided/needed amounts to parse. The retry must classify it by code and
+    fall back to shrinking by 1."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[1, 1, 1]
+    )
+    mock_token_wallet.melt.side_effect = [
+        Exception("Mint Error: Transaction unbalanced: 179, 178, 2 (Code: 11005)"),
+        Mock(),
+    ]
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert amount == 177
+    assert mock_token_wallet.melt.call_count == 2
+
+
+@pytest.mark.asyncio
+async def test_swap_quote_retries_exhausted() -> None:
+    """A mint that escalates fee_reserve on every re-quote exhausts the retry
+    budget (3 attempts) and fails cleanly; melt is never executed."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[1, 10, 25, 50]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="insufficient to cover melt fees"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+    assert mock_token_wallet.melt_quote.call_count == 4  # estimation + 3 attempts
+    mock_token_wallet.melt.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_swap_melt_retries_exhausted() -> None:
+    """A mint that always demands more at melt time than it quoted exhausts
+    the retry budget; the last melt failure is wrapped as ValueError."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        5000, fee_reserves=[50, 50, 50, 50]
+    )
+    mock_token_wallet.melt = AsyncMock(
+        side_effect=Exception(
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 5000, needed: 5200 (Code: 11000)"
+        )
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="Failed to melt token"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+    assert mock_token_wallet.melt.call_count == 3
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "primary_unit,token_unit,amount_msat,fees,expected",
+    [
+        ("sat", "sat", 179_000, 2, 177),
+        ("msat", "sat", 179_000, 2, 177_000),
+        ("sat", "msat", 179_000, 2_000, 177),
+    ],
+)
+async def test_net_minted_amount_unit_conversions(
+    primary_unit: str, token_unit: str, amount_msat: int, fees: int, expected: int
+) -> None:
+    """Fee subtraction converts correctly between sat and msat on either side."""
+    from routstr.core.settings import settings
+    from routstr.wallet import _net_minted_amount
+
+    with patch.object(settings, "primary_mint_unit", primary_unit):
+        assert _net_minted_amount(amount_msat, token_unit, fees) == expected
+
+
+@pytest.mark.parametrize(
+    "message,expected",
+    [
+        # nutshell: retryable with exact shortfall from the detail text
+        (
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 179, needed: 182 (Code: 11000)",
+            3,
+        ),
+        # verbatim production error from issue #468, including cashu-py's
+        # "could not pay invoice" wrapper around the mint detail
+        (
+            "could not pay invoice: Mint Error: not enough inputs provided "
+            "for melt. Provided: 179, needed: 180 (Code: 11000)",
+            1,
+        ),
+        # cdk: registered TransactionUnbalanced code, no parsable amounts
+        ("Mint Error: Transaction unbalanced: 179, 178, 2 (Code: 11005)", 1),
+        # nutshell wording without a code suffix
+        ("not enough inputs provided for melt", 1),
+        # nonsensical amounts (needed <= provided) fall back to the minimal step
+        (
+            "Mint Error: not enough inputs provided for melt. "
+            "Provided: 180, needed: 179 (Code: 11000)",
+            1,
+        ),
+        # a generic 11000 without the shortfall text is not a fee shortfall:
+        # 11000 is nutshell's catch-all TransactionError, so retrying (shrinking
+        # the invoice) would never help and only masks the real error
+        ("Mint Error: Duplicate inputs provided. (Code: 11000)", None),
+        # spent proofs must never be retried: the funds are gone
+        ("Mint Error: Token already spent. (Code: 11001)", None),
+        # Lightning failures must never be retried: a smaller invoice won't help
+        ("Mint Error: Lightning payment failed. (Code: 20004)", None),
+        # unrecognizable errors (timeouts, bugs) must never be retried
+        ("Connection timeout", None),
+    ],
+)
+def test_melt_shortfall_classifier(message: str, expected: int | None) -> None:
+    """Retry classification across mint implementations and failure classes."""
+    from routstr.wallet import _melt_insufficient_shortfall
+
+    assert _melt_insufficient_shortfall(Exception(message)) == expected
+
+
+@pytest.mark.asyncio
+async def test_calculate_swap_amount_same_mint_short_circuit() -> None:
+    """When the token is already on the primary mint no fees apply and no
+    quotes are requested."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            result = await _calculate_swap_amount(
+                amount_msat=1_000_000,
+                token_unit="sat",
+                token_mint_url="http://primary:3338",
+                token_wallet=mock_token_wallet,
+                primary_wallet=mock_primary_wallet,
+                proofs=[],
+            )
+
+    assert result == 1000
+    mock_primary_wallet.request_mint.assert_not_called()
+    mock_token_wallet.melt_quote.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_calculate_swap_amount_msat_primary_unit() -> None:
+    """With an msat primary mint the dummy quote and result stay in msats."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[2]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "msat"):
+            result = await _calculate_swap_amount(
+                amount_msat=179_000,
+                token_unit="sat",
+                token_mint_url="http://foreign-mint:3338",
+                token_wallet=mock_token_wallet,
+                primary_wallet=mock_primary_wallet,
+                proofs=[],
+            )
+
+    assert result == 177_000  # 179_000 msat - 2 sat fee
+    mock_primary_wallet.request_mint.assert_called_once_with(179_000)
+
+
+@pytest.mark.asyncio
+async def test_calculate_swap_amount_fees_exceed_token() -> None:
+    """Fees larger than the token itself fail fast, before any melt."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[200]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with pytest.raises(ValueError, match="exceed token amount"):
+                await _calculate_swap_amount(
+                    amount_msat=179_000,
+                    token_unit="sat",
+                    token_mint_url="http://foreign-mint:3338",
+                    token_wallet=mock_token_wallet,
+                    primary_wallet=mock_primary_wallet,
+                    proofs=[],
+                )
+
+
+@pytest.mark.asyncio
+async def test_calculate_swap_amount_wraps_estimation_failure() -> None:
+    """Estimation infrastructure failures surface as a single clear ValueError."""
+    from routstr.wallet import _calculate_swap_amount
+
+    _, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        179, fee_reserves=[]
+    )
+    mock_primary_wallet.request_mint = AsyncMock(
+        side_effect=Exception("mint offline")
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with pytest.raises(ValueError, match="Failed to estimate fees"):
+                await _calculate_swap_amount(
+                    amount_msat=179_000,
+                    token_unit="sat",
+                    token_mint_url="http://foreign-mint:3338",
+                    token_wallet=mock_token_wallet,
+                    primary_wallet=mock_primary_wallet,
+                    proofs=[],
+                )
+
+
+@pytest.mark.asyncio
+async def test_swap_coerces_non_integer_amount() -> None:
+    """Token amounts arriving as floats are coerced before any arithmetic."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    mock_token.amount = 1000.0
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert amount == 990
+    assert isinstance(amount, int)
+
+
+@pytest.mark.asyncio
+async def test_swap_rejects_unknown_unit() -> None:
+    """Units other than sat/msat are rejected before any quote is requested."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[]
+    )
+    mock_token.unit = "usd"
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="Invalid unit"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+    mock_primary_wallet.request_mint.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_swap_msat_token_already_on_primary() -> None:
+    """msat-denominated tokens on the primary mint short-circuit unchanged."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, _ = _make_swap_mocks(
+        179_000, fee_reserves=[], mint_url="http://primary:3338"
+    )
+    mock_token.unit = "msat"
+    mock_token_wallet.split = AsyncMock()
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_token_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert (amount, unit, mint) == (179_000, "msat", "http://primary:3338")
+
+
+# ---------------------------------------------------------------------------
+# Mint-on-primary failure handling after a successful melt
+#
+# At this point the foreign proofs are already spent: failures here mean funds
+# are in limbo, so errors must propagate (never be swallowed) and recovery must
+# never credit proofs the wallet does not actually hold.
+# ---------------------------------------------------------------------------
+
+
+def _with_recovery_mocks(
+    mock_primary_wallet: Mock, mint_error: str, balances: list[int]
+) -> None:
+    """Make primary mint() fail and stage available_balance per load_proofs call."""
+    mock_primary_wallet.mint = AsyncMock(side_effect=Exception(mint_error))
+    mock_primary_wallet.keysets = ["keyset_primary"]
+    balance_iter = iter(balances)
+
+    def advance_balance(reload: bool = False) -> None:
+        mock_primary_wallet.available_balance = Mock(amount=next(balance_iter))
+
+    mock_primary_wallet.load_proofs = AsyncMock(side_effect=advance_balance)
+    mock_primary_wallet.restore_tokens_for_keyset = AsyncMock()
+
+
+@pytest.mark.asyncio
+async def test_swap_mint_failure_propagates_unwrapped() -> None:
+    """A non-recoverable mint failure after melt propagates as-is (a 500, not a
+    client error): the melt already spent the foreign proofs."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    _with_recovery_mocks(
+        mock_primary_wallet, "Mint Error: Quote is expired (Code: 20007)", [0]
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(Exception, match="Quote is expired") as exc_info:
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+    assert type(exc_info.value) is Exception  # original error, not wrapped
+    assert mock_token_wallet.melt.call_count == 1
+    mock_primary_wallet.restore_tokens_for_keyset.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_swap_recovers_orphaned_proofs_on_outputs_already_signed() -> None:
+    """11003 (outputs already signed): a recovery scan that restores the full
+    minted amount lets the swap complete normally."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    _with_recovery_mocks(
+        mock_primary_wallet,
+        "Mint Error: outputs already signed (Code: 11003)",
+        [0, 990],  # pre-mint balance, post-recovery balance
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                amount, unit, mint = await swap_to_primary_mint(
+                    mock_token, mock_token_wallet
+                )
+
+    assert amount == 990
+    mock_primary_wallet.restore_tokens_for_keyset.assert_awaited_once_with(
+        "keyset_primary", to=1, batch=25
+    )
+
+
+@pytest.mark.asyncio
+async def test_swap_recovery_shortfall_refuses_credit() -> None:
+    """When the recovery scan restores less than the minted amount, the swap
+    must fail rather than credit proofs the wallet does not hold."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    _with_recovery_mocks(
+        mock_primary_wallet,
+        "Mint Error: outputs already signed (Code: 11003)",
+        [0, 100],  # recovery restores only 100 of the expected 990
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="Swap recovery failed"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+
+@pytest.mark.asyncio
+async def test_swap_recovery_failure_wrapped() -> None:
+    """When the recovery scan itself fails, the error is wrapped and raised —
+    never swallowed."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    _with_recovery_mocks(
+        mock_primary_wallet,
+        "Mint Error: outputs already signed (Code: 11003)",
+        [0],
+    )
+    mock_primary_wallet.restore_tokens_for_keyset = AsyncMock(
+        side_effect=Exception("wallet db locked")
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="recovery unsuccessful"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+
+@pytest.mark.asyncio
+async def test_recieve_token_rejects_multiple_keysets() -> None:
+    """Multi-keyset tokens are rejected before touching any wallet."""
+    with patch("routstr.wallet.deserialize_token_from_string") as mock_deserialize:
+        mock_token = Mock()
+        mock_token.keysets = ["keyset1", "keyset2"]
+        mock_deserialize.return_value = mock_token
+
+        with pytest.raises(ValueError, match="Multiple keysets"):
+            await recieve_token("cashuAmultikeyset")
+
+
+@pytest.mark.asyncio
+async def test_credit_balance_msat_unit_not_converted() -> None:
+    """msat-denominated redemptions are credited as-is, without a 1000x."""
+    mock_key = Mock()
+    mock_key.balance = 0
+    mock_key.hashed_key = "test_hash"
+    mock_session = AsyncMock()
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "cashu_mints", ["http://mint:3338"]):
+        with patch(
+            "routstr.wallet.recieve_token",
+            return_value=(1_000_000, "msat", "http://mint:3338"),
+        ):
+            amount = await credit_balance("cashuAtest", mock_key, mock_session)
+
+    assert amount == 1_000_000
+    assert mock_session.commit.called
+
+
+@pytest.mark.asyncio
+async def test_credit_balance_survives_audit_store_failure() -> None:
+    """A failure writing the CashuTransaction history record must not undo the
+    already-committed balance credit. (The silent swallow is a known
+    audit-trail gap slated for its own fix — this test pins the financial
+    invariant that the user keeps their credit, not the swallow itself.)"""
+    mock_key = Mock()
+    mock_key.balance = 0
+    mock_key.hashed_key = "test_hash"
+    mock_session = AsyncMock()
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "cashu_mints", ["http://mint:3338"]):
+        with patch(
+            "routstr.wallet.recieve_token",
+            return_value=(1000, "sat", "http://mint:3338"),
+        ):
+            with patch(
+                "routstr.wallet.store_cashu_transaction",
+                side_effect=Exception("history table locked"),
+            ):
+                amount = await credit_balance("cashuAtest", mock_key, mock_session)
+
+    assert amount == 1_000_000
+    assert mock_session.commit.called
+
+
+@pytest.mark.asyncio
+async def test_swap_does_not_retry_on_payment_failure() -> None:
+    """Melt failures unrelated to fees (e.g. routing failure) are not retried:
+    a smaller invoice would not help, and the error must surface immediately."""
+    from routstr.wallet import swap_to_primary_mint
+
+    mock_token, mock_token_wallet, mock_primary_wallet = _make_swap_mocks(
+        1000, fee_reserves=[10, 10]
+    )
+    mock_token_wallet.melt = AsyncMock(
+        side_effect=Exception("Mint Error: Lightning payment failed. (Code: 20004)")
+    )
+
+    from routstr.core.settings import settings
+
+    with patch.object(settings, "primary_mint", "http://primary:3338"):
+        with patch.object(settings, "primary_mint_unit", "sat"):
+            with patch("routstr.wallet.get_wallet", return_value=mock_primary_wallet):
+                with pytest.raises(ValueError, match="Failed to melt token"):
+                    await swap_to_primary_mint(mock_token, mock_token_wallet)
+
+    assert mock_token_wallet.melt.call_count == 1
+    assert mock_primary_wallet.request_mint.call_count == 2

ui/README.md

@@ -1,36 +1,44 @@
-This is a [Next.js](https://nextjs.org) project bootstrapped with [`create-next-app`](https://nextjs.org/docs/app/api-reference/cli/create-next-app).
+# Routstr node admin UI
 
-## Getting Started
+A [Next.js](https://nextjs.org) app (App Router, **static export**) that provides the
+admin dashboard for a `routstr-core` node: login, settings, providers, balances,
+transactions, usage, and logs.
 
-First, run the development server:
+There is no separate web server in production. `next build` produces a fully static
+export (`next.config.ts` sets `output: 'export'`), and the FastAPI backend serves it
+directly from `../ui_out/` (see `routstr/core/main.py`). So the UI and the API are
+served from the **same origin** in production.
 
-```bash
-npm run dev
-# or
-yarn dev
-# or
-pnpm dev
-# or
-bun dev
-```
+## Developing the UI (hot reload)
 
-Open [http://localhost:3000](http://localhost:3000) with your browser to see the result.
+The everyday loop runs two processes side by side — you do **not** rebuild the static
+export while developing:
 
-You can start editing the page by modifying `app/page.tsx`. The page auto-updates as you edit the file.
+1. Start the backend on `:8000` — from the repo root: `make docker-up` (or
+   `uvicorn routstr.core.main:app --reload`).
+2. Start the Next.js dev server on `:3000` — from the repo root: `make ui-dev`
+   (or `cd ui && pnpm dev`). Edits hot-reload instantly.
 
-This project uses [`next/font`](https://nextjs.org/docs/app/building-your-application/optimizing/fonts) to automatically optimize and load [Geist](https://vercel.com/font), a new font family for Vercel.
+Open http://localhost:3000. With no `NEXT_PUBLIC_API_URL` set, the UI falls back to
+`http://127.0.0.1:8000` in development (see `lib/api/services/configuration.ts`), so it
+talks to the local backend out of the box.
 
-## Learn More
+Because dev is cross-origin (`:3000` → `:8000`), it relies on the backend's CORS
+allowing the UI origin. The default `cors_origins` is `["*"]`; if you tighten CORS,
+keep `http://localhost:3000` allowed for development.
 
-To learn more about Next.js, take a look at the following resources:
+## Building the integrated/static UI (what production serves)
 
-- [Next.js Documentation](https://nextjs.org/docs) - learn about Next.js features and API.
-- [Learn Next.js](https://nextjs.org/learn) - an interactive Next.js tutorial.
+To produce the bundle that FastAPI serves from `../ui_out/`:
 
-You can check out [the Next.js GitHub repository](https://github.com/vercel/next.js) - your feedback and contributions are welcome!
+- `make ui-build` — builds with local Node/pnpm (`scripts/build-ui.sh`), then moves
+  `ui/out/*` to `../ui_out/`.
+- `make ui-build-docker` — same, but inside Docker (no local Node needed).
 
-## Deploy on Vercel
+`NEXT_PUBLIC_*` variables are read from the repo-root `.env` at build time and baked in.
+For a same-origin deployment leave `NEXT_PUBLIC_API_URL` empty (relative paths); the UI
+uses `window.location.origin` at runtime. After building, start the backend and open
+http://localhost:8000 — the dashboard is served at `/` and `/admin`.
 
-The easiest way to deploy your Next.js app is to use the [Vercel Platform](https://vercel.com/new?utm_medium=default-template&filter=next.js&utm_source=create-next-app&utm_campaign=create-next-app-readme) from the creators of Next.js.
-
-Check out our [Next.js deployment documentation](https://nextjs.org/docs/app/building-your-application/deploying) for more details.
+If `../ui_out/` does not exist, the backend logs a warning at startup and serves the API
+only (hitting a UI route returns a small JSON fallback instead of the dashboard).