feat: local customizations - refusal detection, kimi routing, usage pricing, auth providers

agent/auxiliary_client.py

@@ -63,6 +63,11 @@ _API_KEY_PROVIDER_AUX_MODELS: Dict[str, str] = {
     "opencode-zen": "gemini-3-flash",
     "opencode-go": "glm-5",
     "kilocode": "google/gemini-3-flash-preview",
+    # Uniwizard backends
+    "gemini": "gemini-2.5-flash",
+    "groq": "llama-3.3-70b-versatile",
+    "grok": "grok-3-mini-fast",
+    "openrouter": "openai/gpt-4.1-mini",
 }
 
 # OpenRouter app attribution headers

agent/usage_pricing.py

@@ -75,6 +75,22 @@ class CostResult:
     notes: tuple[str, ...] = ()
 
 
+@dataclass(frozen=True)
+class CostBreakdown:
+    input_usd: Optional[Decimal]
+    output_usd: Optional[Decimal]
+    cache_read_usd: Optional[Decimal]
+    cache_write_usd: Optional[Decimal]
+    request_usd: Optional[Decimal]
+    total_usd: Optional[Decimal]
+    status: CostStatus
+    source: CostSource
+    label: str
+    fetched_at: Optional[datetime] = None
+    pricing_version: Optional[str] = None
+    notes: tuple[str, ...] = ()
+
+
 _UTC_NOW = lambda: datetime.now(timezone.utc)
 
 
@@ -93,6 +109,25 @@ _OFFICIAL_DOCS_PRICING: Dict[tuple[str, str], PricingEntry] = {
         source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
         pricing_version="anthropic-prompt-caching-2026-03-16",
     ),
+    # Aliases for short model names (Anthropic API resolves these to dated versions)
+    ("anthropic", "claude-opus-4-6"): PricingEntry(
+        input_cost_per_million=Decimal("15.00"),
+        output_cost_per_million=Decimal("75.00"),
+        cache_read_cost_per_million=Decimal("1.50"),
+        cache_write_cost_per_million=Decimal("18.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
+    ("anthropic", "claude-opus-4.6"): PricingEntry(
+        input_cost_per_million=Decimal("15.00"),
+        output_cost_per_million=Decimal("75.00"),
+        cache_read_cost_per_million=Decimal("1.50"),
+        cache_write_cost_per_million=Decimal("18.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
     (
         "anthropic",
         "claude-sonnet-4-20250514",
@@ -105,6 +140,24 @@ _OFFICIAL_DOCS_PRICING: Dict[tuple[str, str], PricingEntry] = {
         source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
         pricing_version="anthropic-prompt-caching-2026-03-16",
     ),
+    ("anthropic", "claude-sonnet-4-5"): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        cache_read_cost_per_million=Decimal("0.30"),
+        cache_write_cost_per_million=Decimal("3.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
+    ("anthropic", "claude-sonnet-4.5"): PricingEntry(
+        input_cost_per_million=Decimal("3.00"),
+        output_cost_per_million=Decimal("15.00"),
+        cache_read_cost_per_million=Decimal("0.30"),
+        cache_write_cost_per_million=Decimal("3.75"),
+        source="official_docs_snapshot",
+        source_url="https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching",
+        pricing_version="anthropic-prompt-caching-2026-03-16",
+    ),
     # OpenAI
     (
         "openai",
@@ -654,3 +707,80 @@ def format_token_count_compact(value: int) -> str:
             return f"{sign}{text}{suffix}"
 
     return f"{value:,}"
+
+
+
+def estimate_usage_cost_breakdown(
+    model_name: str,
+    usage: CanonicalUsage,
+    *,
+    provider: Optional[str] = None,
+    base_url: Optional[str] = None,
+    api_key: Optional[str] = None,
+) -> CostBreakdown:
+    """Estimate per-bucket cost breakdown for a usage record.
+
+    Returns the same status/source semantics as estimate_usage_cost(), but splits
+    the total into input/cache/output/request components when pricing data is
+    available. For subscription-included routes (e.g. openai-codex), all
+    components are reported as zero-cost instead of unknown.
+    """
+    cost_result = estimate_usage_cost(
+        model_name,
+        usage,
+        provider=provider,
+        base_url=base_url,
+        api_key=api_key,
+    )
+    route = resolve_billing_route(model_name, provider=provider, base_url=base_url)
+    entry = get_pricing_entry(model_name, provider=provider, base_url=base_url, api_key=api_key)
+    if not entry and route.billing_mode == "subscription_included":
+        entry = PricingEntry(
+            input_cost_per_million=_ZERO,
+            output_cost_per_million=_ZERO,
+            cache_read_cost_per_million=_ZERO,
+            cache_write_cost_per_million=_ZERO,
+            request_cost=_ZERO,
+            source="none",
+            pricing_version="included-route",
+        )
+
+    if not entry:
+        return CostBreakdown(
+            input_usd=None,
+            output_usd=None,
+            cache_read_usd=None,
+            cache_write_usd=None,
+            request_usd=None,
+            total_usd=cost_result.amount_usd,
+            status=cost_result.status,
+            source=cost_result.source,
+            label=cost_result.label,
+            fetched_at=cost_result.fetched_at,
+            pricing_version=cost_result.pricing_version,
+            notes=cost_result.notes,
+        )
+
+    def _component(tokens: int, rate: Optional[Decimal]) -> Optional[Decimal]:
+        if rate is None:
+            return None
+        return (Decimal(tokens or 0) * rate) / _ONE_MILLION
+
+    request_usd = None
+    if entry.request_cost is not None:
+        request_usd = Decimal(usage.request_count or 0) * entry.request_cost
+
+    return CostBreakdown(
+        input_usd=_component(usage.input_tokens, entry.input_cost_per_million),
+        output_usd=_component(usage.output_tokens, entry.output_cost_per_million),
+        cache_read_usd=_component(usage.cache_read_tokens, entry.cache_read_cost_per_million),
+        cache_write_usd=_component(usage.cache_write_tokens, entry.cache_write_cost_per_million),
+        request_usd=request_usd,
+        total_usd=cost_result.amount_usd,
+        status=cost_result.status,
+        source=cost_result.source,
+        label=cost_result.label,
+        fetched_at=cost_result.fetched_at,
+        pricing_version=cost_result.pricing_version,
+        notes=cost_result.notes,
+    )

cli.py

@@ -4563,7 +4563,30 @@ class HermesCLI:
             print("(._.) No API calls made yet in this session.")
             return
 
-        # Current context window state
+        def _fmt_money(amount):
+            return "n/a" if amount is None else f"${float(amount):.4f}"
+
+        def _fmt_limit(remaining, limit):
+            if remaining is None and limit is None:
+                return "n/a"
+            if remaining is None:
+                return f"? / {limit:,}"
+            if limit is None:
+                return f"{remaining:,} / ?"
+            return f"{remaining:,} / {limit:,}"
+
+        def _fmt_reset(seconds):
+            if seconds is None:
+                return "n/a"
+            seconds = int(seconds)
+            if seconds < 60:
+                return f"{seconds}s"
+            minutes, secs = divmod(seconds, 60)
+            if minutes < 60:
+                return f"{minutes}m {secs}s"
+            hours, minutes = divmod(minutes, 60)
+            return f"{hours}h {minutes}m"
+
         compressor = agent.context_compressor
         last_prompt = compressor.last_prompt_tokens
         ctx_len = compressor.context_length
@@ -4571,14 +4594,21 @@ class HermesCLI:
         compressions = compressor.compression_count
 
         msg_count = len(self.conversation_history)
+        usage = CanonicalUsage(
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+            cache_read_tokens=cache_read_tokens,
+            cache_write_tokens=cache_write_tokens,
+        )
         cost_result = estimate_usage_cost(
             agent.model,
-            CanonicalUsage(
+            usage,
-                input_tokens=input_tokens,
+            provider=getattr(agent, "provider", None),
-                output_tokens=output_tokens,
+            base_url=getattr(agent, "base_url", None),
-                cache_read_tokens=cache_read_tokens,
+        )
-                cache_write_tokens=cache_write_tokens,
+        cost_breakdown = estimate_usage_cost_breakdown(
-            ),
+            agent.model,
+            usage,
             provider=getattr(agent, "provider", None),
             base_url=getattr(agent, "base_url", None),
         )
@@ -4605,6 +4635,38 @@ class HermesCLI:
             print(f"  Total cost:              {'included':>10}")
         else:
             print(f"  Total cost:              {'n/a':>10}")
+        print(f"  Cost input:              {_fmt_money(cost_breakdown.input_usd):>10}")
+        print(f"  Cost cache read:         {_fmt_money(cost_breakdown.cache_read_usd):>10}")
+        print(f"  Cost cache write:        {_fmt_money(cost_breakdown.cache_write_usd):>10}")
+        print(f"  Cost output:             {_fmt_money(cost_breakdown.output_usd):>10}")
+        if cost_breakdown.request_usd is not None:
+            print(f"  Cost requests:           {_fmt_money(cost_breakdown.request_usd):>10}")
+
+        rate_limits = getattr(agent, "session_openai_rate_limits", None) or {}
+        last_request_id = getattr(agent, "session_last_request_id", None)
+        rate_limit_events = getattr(agent, "session_rate_limit_events", 0) or 0
+        if last_request_id:
+            print(f"  Last request id:         {last_request_id:>10}")
+        if rate_limits:
+            status_code = rate_limits.get("status_code")
+            if status_code is not None:
+                print(f"  Last HTTP status:        {status_code:>10}")
+            req_remaining = rate_limits.get("remaining_requests")
+            req_limit = rate_limits.get("limit_requests")
+            req_reset = rate_limits.get("reset_requests_seconds")
+            if req_remaining is not None or req_limit is not None:
+                print(f"  Req limit:             {_fmt_limit(req_remaining, req_limit):>14}  reset {_fmt_reset(req_reset)}")
+            tok_remaining = rate_limits.get("remaining_tokens")
+            tok_limit = rate_limits.get("limit_tokens")
+            tok_reset = rate_limits.get("reset_tokens_seconds")
+            if tok_remaining is not None or tok_limit is not None:
+                print(f"  Token limit:           {_fmt_limit(tok_remaining, tok_limit):>14}  reset {_fmt_reset(tok_reset)}")
+            retry_after = rate_limits.get("retry_after_seconds")
+            if retry_after is not None:
+                print(f"  Retry after:             {_fmt_reset(retry_after):>10}")
+        if rate_limit_events:
+            print(f"  Rate limit hits:         {rate_limit_events:>10,}")
+
         print(f"  {'─' * 40}")
         print(f"  Current context:  {last_prompt:,} / {ctx_len:,} ({pct:.0f}%)")
         print(f"  Messages:         {msg_count}")

hermes_cli/auth.py

@@ -220,6 +220,39 @@ PROVIDER_REGISTRY: Dict[str, ProviderConfig] = {
         api_key_env_vars=("HF_TOKEN",),
         base_url_env_var="HF_BASE_URL",
     ),
+    # ── Uniwizard backends (added 2026-03-30) ─────────────────────────
+    "gemini": ProviderConfig(
+        id="gemini",
+        name="Google Gemini",
+        auth_type="api_key",
+        inference_base_url="https://generativelanguage.googleapis.com/v1beta/openai",
+        api_key_env_vars=("GEMINI_API_KEY",),
+        base_url_env_var="GEMINI_BASE_URL",
+    ),
+    "groq": ProviderConfig(
+        id="groq",
+        name="Groq",
+        auth_type="api_key",
+        inference_base_url="https://api.groq.com/openai/v1",
+        api_key_env_vars=("GROQ_API_KEY",),
+        base_url_env_var="GROQ_BASE_URL",
+    ),
+    "grok": ProviderConfig(
+        id="grok",
+        name="xAI Grok",
+        auth_type="api_key",
+        inference_base_url="https://api.x.ai/v1",
+        api_key_env_vars=("XAI_API_KEY", "GROK_API_KEY"),
+        base_url_env_var="XAI_BASE_URL",
+    ),
+    "openrouter": ProviderConfig(
+        id="openrouter",
+        name="OpenRouter",
+        auth_type="api_key",
+        inference_base_url="https://openrouter.ai/api/v1",
+        api_key_env_vars=("OPENROUTER_API_KEY",),
+        base_url_env_var="OPENROUTER_BASE_URL",
+    ),
 }
 
 

pyproject.toml

@@ -13,7 +13,8 @@ license = { text = "MIT" }
 dependencies = [
   # Core — pinned to known-good ranges to limit supply chain attack surface
   "openai>=2.21.0,<3",
-  "anthropic>=0.39.0,<1",\n  "google-genai>=1.2.0,<2",
+  "anthropic>=0.39.0,<1",
+  "google-genai>=1.2.0,<2",
   "python-dotenv>=1.2.1,<2",
   "fire>=0.7.1,<1",
   "httpx>=0.28.1,<1",

run_agent.py

@@ -3472,6 +3472,79 @@ class AIAgent:
         http_client = getattr(client, "_client", None)
         return bool(getattr(http_client, "is_closed", False))
 
+    def _coerce_rate_limit_int(self, value: Any) -> Optional[int]:
+        try:
+            if value is None or value == "":
+                return None
+            return int(float(str(value).strip()))
+        except Exception:
+            return None
+
+    def _parse_rate_limit_reset_seconds(self, value: Any) -> Optional[int]:
+        if value is None:
+            return None
+        text = str(value).strip().lower()
+        if not text:
+            return None
+        try:
+            return int(round(float(text)))
+        except Exception:
+            pass
+        total = 0.0
+        matches = re.findall(r"(\d+(?:\.\d+)?)(ms|s|m|h)", text)
+        if not matches:
+            return None
+        for number, unit in matches:
+            value_f = float(number)
+            if unit == "ms":
+                total += value_f / 1000.0
+            elif unit == "s":
+                total += value_f
+            elif unit == "m":
+                total += value_f * 60.0
+            elif unit == "h":
+                total += value_f * 3600.0
+        return int(round(total))
+
+    def _capture_openai_http_response(self, response: Any) -> None:
+        if self.api_mode == "anthropic_messages":
+            return
+        headers = getattr(response, "headers", None)
+        if not headers:
+            return
+        lowered = {str(k).lower(): str(v) for k, v in headers.items()}
+        telemetry = dict(getattr(self, "session_openai_rate_limits", {}) or {})
+
+        def _put(key: str, value: Any) -> None:
+            if value is not None:
+                telemetry[key] = value
+
+        _put("status_code", getattr(response, "status_code", None))
+        _put("limit_requests", self._coerce_rate_limit_int(lowered.get("x-ratelimit-limit-requests")))
+        _put("remaining_requests", self._coerce_rate_limit_int(lowered.get("x-ratelimit-remaining-requests")))
+        _put("limit_tokens", self._coerce_rate_limit_int(lowered.get("x-ratelimit-limit-tokens")))
+        _put("remaining_tokens", self._coerce_rate_limit_int(lowered.get("x-ratelimit-remaining-tokens")))
+        _put("reset_requests_seconds", self._parse_rate_limit_reset_seconds(lowered.get("x-ratelimit-reset-requests")))
+        _put("reset_tokens_seconds", self._parse_rate_limit_reset_seconds(lowered.get("x-ratelimit-reset-tokens")))
+
+        retry_after_seconds = None
+        retry_after_ms = self._coerce_rate_limit_int(lowered.get("retry-after-ms"))
+        if retry_after_ms is not None:
+            retry_after_seconds = max(0, int(round(retry_after_ms / 1000.0)))
+        if retry_after_seconds is None:
+            retry_after_seconds = self._parse_rate_limit_reset_seconds(lowered.get("retry-after"))
+        _put("retry_after_seconds", retry_after_seconds)
+        _put("observed_at", int(time.time()))
+
+        request_id = lowered.get("x-request-id") or lowered.get("openai-request-id")
+        if request_id:
+            self.session_last_request_id = request_id
+            _put("request_id", request_id)
+
+        self.session_openai_rate_limits = telemetry
+        if getattr(response, "status_code", None) == 429:
+            self.session_rate_limit_events = (getattr(self, "session_rate_limit_events", 0) or 0) + 1
+
     def _create_openai_client(self, client_kwargs: dict, *, reason: str, shared: bool) -> Any:
         if self.provider == "copilot-acp" or str(client_kwargs.get("base_url", "")).startswith("acp://copilot"):
             from agent.copilot_acp_client import CopilotACPClient
@@ -3485,6 +3558,23 @@ class AIAgent:
             )
             return client
         client = OpenAI(**client_kwargs)
+        http_client = getattr(client, "_client", None)
+        if http_client is not None and not getattr(http_client, "_hermes_response_telemetry_installed", False):
+            original_send = http_client.send
+
+            def _send_with_telemetry(request, *args, **kwargs):
+                response = original_send(request, *args, **kwargs)
+                try:
+                    self._capture_openai_http_response(response)
+                except Exception as exc:
+                    logger.debug("OpenAI response telemetry capture failed: %s", exc)
+                return response
+
+            http_client.send = _send_with_telemetry
+            try:
+                setattr(http_client, "_hermes_response_telemetry_installed", True)
+            except Exception:
+                pass
         logger.info(
             "OpenAI client created (%s, shared=%s) %s",
             reason,
@@ -7466,6 +7556,53 @@ class AIAgent:
                 if hasattr(self, '_incomplete_scratchpad_retries'):
                     self._incomplete_scratchpad_retries = 0
 
+                # ── Uniwizard: Semantic refusal detection ──────────────────
+                # Catches 200 OK responses where the model REFUSED the request.
+                # No existing LLM gateway does this. This is novel.
+                if (assistant_message.content
+                        and not assistant_message.tool_calls
+                        and self._fallback_index < len(self._fallback_chain)):
+                    _refusal_text = (assistant_message.content or "").strip()
+                    _REFUSAL_PATTERNS = (
+                        "I can't help with",
+                        "I cannot help with",
+                        "I'm not able to",
+                        "I am not able to",
+                        "I must decline",
+                        "I'm unable to",
+                        "I am unable to",
+                        "against my guidelines",
+                        "against my policy",
+                        "I can't assist with",
+                        "I cannot assist with",
+                        "I apologize, but I can't",
+                        "I'm sorry, but I can't",
+                        "I'm sorry, but I cannot",
+                        "not something I can help",
+                        "I don't think I should",
+                        "I can't fulfill that",
+                        "I cannot fulfill that",
+                        "I'm not comfortable",
+                        "I can't provide",
+                        "I cannot provide",
+                    )
+                    _refusal_lower = _refusal_text.lower()
+                    _is_refusal = any(p.lower() in _refusal_lower for p in _REFUSAL_PATTERNS)
+                    if _is_refusal:
+                        _fb_target = self._fallback_chain[self._fallback_index]
+                        self._emit_status(
+                            f"🚫 Semantic refusal detected from {self.provider}/{self.model}. "
+                            f"Rerouting to {_fb_target.get('model', '?')} via {_fb_target.get('provider', '?')}..."
+                        )
+                        logging.warning(
+                            "Refusal detected from %s/%s: %.80s...",
+                            self.provider, self.model, _refusal_text,
+                        )
+                        if self._try_activate_fallback():
+                            retry_count = 0
+                            continue
+                # ── End refusal detection ──────────────────────────────────
+
                 if self.api_mode == "codex_responses" and finish_reason == "incomplete":
                     if not hasattr(self, "_codex_incomplete_retries"):
                         self._codex_incomplete_retries = 0

tests/test_cli_status_bar.py

@@ -144,6 +144,42 @@ class TestCLIUsageReport:
         assert "0.064" in output
         assert "Session duration:" in output
         assert "Compressions:" in output
+        assert "Cost input:" in output
+        assert "Cost output:" in output
+
+    def test_show_usage_displays_rate_limit_telemetry(self, capsys):
+        cli_obj = _attach_agent(
+            _make_cli(model="openai/gpt-5.4"),
+            prompt_tokens=10_000,
+            completion_tokens=500,
+            total_tokens=10_500,
+            api_calls=3,
+            context_tokens=10_500,
+            context_length=200_000,
+        )
+        cli_obj.agent.provider = "openai-codex"
+        cli_obj.agent.session_openai_rate_limits = {
+            "status_code": 200,
+            "limit_requests": 60,
+            "remaining_requests": 48,
+            "reset_requests_seconds": 33,
+            "limit_tokens": 2000000,
+            "remaining_tokens": 1750000,
+            "reset_tokens_seconds": 90,
+            "retry_after_seconds": 5,
+        }
+        cli_obj.agent.session_last_request_id = "req_123"
+        cli_obj.agent.session_rate_limit_events = 2
+        cli_obj.verbose = False
+
+        cli_obj._show_usage()
+        output = capsys.readouterr().out
+
+        assert "Last request id:" in output
+        assert "Req limit:" in output
+        assert "Token limit:" in output
+        assert "Retry after:" in output
+        assert "Rate limit hits:" in output
 
     def test_show_usage_marks_unknown_pricing(self, capsys):
         cli_obj = _attach_agent(