eggpool 0.7.0__tar.gz → 0.7.2__tar.gz

{eggpool-0.7.0 → eggpool-0.7.2}/CHANGELOG.md

@@ -244,6 +244,30 @@ Documentation in `architecture/README.md`, `docs/live-config-rehash.md`,
   command inventory updated with slow-writer, extended soak, and soak
   runner commands.
 
+### Fixed
+
+- **Provider-suffixed model ids no longer leak the ``/provider-id``
+  suffix to upstream.** ``/v1/models`` exposes provider-scoped entries in
+  the form ``<model-id>/<provider-id>``. When a client used one of those
+  suffixed ids against an endpoint whose body is transcoded or rewritten
+  -- ``/v1/chat/completions`` for an Anthropic model, or any cross-
+  protocol retry -- the suffix leaked into the JSON body the upstream
+  received, which rejected the request with ``Model <model-id>/<provider-id>
+  is not supported``. The proxy layer now normalizes the in-memory
+  payload's ``model`` field to the parsed base id at every point where a
+  body is rebuilt for upstream dispatch: the transcode preflight cache
+  (`handle_proxy_request` in `src/eggpool/api/proxy_request.py`), the
+  coordinator's re-translation reset
+  (`_apply_selected_provider_transcode` in
+  `src/eggpool/request/coordinator.py`), and the legacy provider-request
+  adapter (`RequestCoordinator._legacy_provider_request`). The
+  immutable ``client_payload`` contract is preserved -- only the
+  provider-bound snapshot loses the suffix. Regression coverage is
+  parameterized across multiple model ids in
+  `tests/unit/test_prepared_transcode.py` (`TestCoordinatorResetStripsProviderSuffix`)
+  so the fix is model-agnostic and future model names are covered by
+  default.
+
 ## [0.6.0] - 2026-07-09
 
 ### Fixed

{eggpool-0.7.0 → eggpool-0.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: eggpool
-Version: 0.7.0
+Version: 0.7.2
 Summary: A lightweight proxy that aggregates multiple LLM provider accounts behind an OpenAI Chat Completions-compatible endpoint
 Project-URL: Homepage, https://github.com/eggstack/eggpool
 Project-URL: Repository, https://github.com/eggstack/eggpool

{eggpool-0.7.0 → eggpool-0.7.2}/architecture/deep-dive-database.md

@@ -295,8 +295,11 @@ Error handling transitions to `FAILED_CLOSED` from `CONNECTING`, `READY`, or
 must close admission regardless of the current state.
 
 The `_invalidate_connection()` method detaches the connection, records the
-failure reason, and transitions to `FAILED_CLOSED`. It does not attempt
-reconnection — the deployment contract is a worker restart.
+failure reason, and transitions to `FAILED_CLOSED`. Transaction failures
+detach under the connection lock and close the old connection after the lock
+is released; direct callers perform the same handoff before awaiting close.
+It does not attempt reconnection — the deployment contract is a worker
+restart.
 
 ## Schema Baseline Decision (Plan 137)
 

{eggpool-0.7.0 → eggpool-0.7.2}/architecture/deep-dive-health.md

@@ -47,6 +47,8 @@ Per-account health tracking:
 ### `health/circuit_breaker.py`
 
 Circuit breaker implementation:
+- Synchronous state transitions use a short-held `threading.Lock` so the
+  breaker remains safe at thread/diagnostic boundaries.
 - **Closed** (healthy): requests flow normally
 - **Open** (unhealthy): requests blocked, cooldown timer running
 - **Half-open** (probe): single probe request allowed

{eggpool-0.7.0 → eggpool-0.7.2}/architecture/deep-dive-runtime.md

@@ -182,7 +182,8 @@ failed durable clear leaves the current in-memory suppression intact.
   guarantees exactly one asyncio loop regardless of `runtime_threads`, which
   only sizes the Rust-side I/O pool
 - All `asyncio.Lock` objects are loop-bound
-- `MetricsWriteCoalescer` is the only component using `threading.Lock`
+- `MetricsWriteCoalescer` and `CircuitBreaker` use short-held
+  `threading.Lock` guards for their synchronous cross-boundary methods
 - `fastcli` and `runtime_paths` are stdlib-only
 - PID file owned by supervisor
 - Generation swap never interrupts in-flight requests or accepted retained terminal jobs

{eggpool-0.7.0 → eggpool-0.7.2}/architecture/deep-dive-security.md

@@ -21,8 +21,12 @@ Strips sensitive headers from upstream responses before they reach clients:
 
 `[security].trusted_proxies` is an exact immediate-peer allowlist. The proxy
 uses `X-Forwarded-For` or `X-Real-IP` only when the ASGI peer address is in
-that list, and accepts only the first bounded, control-character-free value.
-An empty list ignores forwarded attribution headers and falls back to the
+that list. For `X-Forwarded-For`, it walks from the EggPool-facing side and
+uses the first bounded, control-character-free hop that is not itself trusted,
+so a leftmost client-supplied value cannot override a proxy-appended client
+identity. Configured proxies must overwrite or correctly append these headers;
+`X-Real-IP` is necessarily trusted as supplied by that configured peer. An
+empty list ignores forwarded attribution headers and falls back to the
 immediate peer.
 
 ### `auth.py` — Local API Key Authentication

{eggpool-0.7.0 → eggpool-0.7.2}/config.example.toml

@@ -109,6 +109,11 @@ catalog_withdrawal_policy = "preserve_until_health"
 loss_policy = "warn"      # "warn" or "reject" for loss-of-information
 prefer_native = true      # prefer native-protocol accounts in routing
 
+# Thinking/reasoning transcoding is enabled by default. Set this to false to
+# drop thinking controls and content instead of translating them.
+# [transcoder.features]
+# thinking = false
+
 # Global effort→budget token mapping for thinking/reasoning.
 # Values are used when the model's capability does not carry a
 # per-model effort_to_budget_tokens mapping.
@@ -240,7 +245,7 @@ ttl_seconds = 86400
 # metadata with display names, context limits, benchmarks, and pipeline
 # tags. Manual overrides and alias expansion are also supported.
 # [model_info]
-# enabled = false                  # optional metadata sidecar (default: false)
+# enabled = true                   # metadata sidecar (default: true)
 # include_in_models_endpoint = true  # compact fields in /v1/models (default: true)
 # refresh_interval_s = 3600        # global refresh cadence (seconds)
 # sparse_refresh_interval_s = 300  # accelerated refresh for sparse models (seconds)
@@ -296,7 +301,7 @@ ttl_seconds = 86400
 enabled = false
 
 [model_info]
-enabled = false
+enabled = true
 
 [update_checker]
 enabled = false
@@ -369,6 +374,8 @@ cleanup_max_rows_per_pass = 5000
 allowed_hosts = []
 cors_origins = []
 # Honor X-Forwarded-For/X-Real-IP only when the immediate peer is listed here.
+# Configure only proxies that overwrite or correctly append these headers;
+# X-Forwarded-For uses the nearest untrusted hop in the chain.
 # An empty list ignores forwarded client-attribution headers.
 #
 # List both IPv4 and IPv6 loopback addresses explicitly when the proxy is
@@ -447,7 +454,8 @@ enabled = false
 # header = "Authorization"
 # scheme = "Bearer"
 #
-# # OpenCode Go exposes low/medium/high thinking controls for mimo-v2.5.
+# # OpenCode Go exposes low/medium/high thinking controls for mimo-v2.5 and
+# # minimal/low/medium/high/xhigh for muse-spark-1.2-contributor.
 # [providers.opencode-go.model_capabilities."mimo-v2.5".thinking]
 # status = "supported"
 # source = "provider_catalog"

{eggpool-0.7.0 → eggpool-0.7.2}/docs/deployment.md

@@ -350,9 +350,11 @@ production installation is documented below.
 
 ## Performance Profiles
 
-EggPool ships with a lean default and explicit fuller profiles. The default
-install uses one SQLite worker, low-wear analytics, loopback binding, and
-disabled optional enrichment/diagnostics. All profiles use the supported
+EggPool ships with a lean SBC default and a fuller standard default. The
+standard install uses one SQLite worker, low-wear analytics, loopback binding,
+and the metadata-enrichment sidecar (model_info) enabled by default. The SBC
+profile keeps optional enrichment/diagnostics disabled to bound resource
+use on Raspberry Pi class hardware. All profiles use the supported
 single-event-loop default (`threads = 1`).
 
 `eggpool onboard` asks whether to bind the API/dashboard to the LAN. Answering
@@ -360,10 +362,11 @@ yes writes `server.host = "0.0.0.0"`; noninteractive onboarding writes the safe
 loopback value. Dashboard authentication remains enabled by default through
 the generated server API key (`dashboard.public = false`).
 
-Optional background features are explicit: set `[model_info].enabled = true`
-for metadata enrichment, `[backup].enabled = true` for in-process archives,
-or `[update_checker].enabled = true` for the periodic PyPI status probe. Host
-DNS behavior is provided by the operating system; EggPool does not add a
+Optional background features that remain explicit: set `[backup].enabled = true`
+for in-process archives or `[update_checker].enabled = true` for the periodic
+PyPI status probe. Set `[model_info].enabled = false` to disable the
+metadata-enrichment sidecar (SBC profile does this by default). Host DNS
+behavior is provided by the operating system; EggPool does not add a
 process-local resolver cache.
 
 For a short deployment comparison, use the same host, Python, config shape,

{eggpool-0.7.0 → eggpool-0.7.2}/docs/thinking.md

@@ -28,16 +28,19 @@ Status values (`CapabilityStatus`):
 
 Source: `src/eggpool/catalog/capabilities.py:36-44`
 
-## 2. Enabling Thinking Transcoding
+## 2. Thinking Transcoding
 
-Thinking transcoding is **off by default**. The feature gate lives in `[transcoder.features]`:
+Thinking transcoding is **enabled by default**. No configuration is required.
+To disable it, set the feature gate in `[transcoder.features]`:
 
 ```toml
 [transcoder.features]
-thinking = true
+thinking = false
 ```
 
-When `thinking = false` (the default), the transcoder **drops** thinking-related fields with a structured `reasoning_content_dropped` warning rather than forwarding them. Specifically:
+When `thinking = false`, the transcoder **drops** thinking-related fields with a
+structured `reasoning_content_dropped` warning rather than forwarding them.
+Specifically:
 
 - **OpenAI → Anthropic**: `reasoning_content` in assistant messages is dropped.
 - **Anthropic → OpenAI**: `thinking` content blocks are dropped.
@@ -63,7 +66,7 @@ Source: `src/eggpool/transcoder/policy.py:217-225`
 
 ```toml
 [transcoder.features]
-thinking = true          # Enable thinking/reasoning transcoding (default: false)
+thinking = true          # Enable thinking/reasoning transcoding (default: true)
 ```
 
 ### Budget Defaults
@@ -500,9 +503,10 @@ Source: `src/eggpool/metrics/thinking.py`, `src/eggpool/api/stats.py:452-463`, `
 
 **Symptoms:** Client sends `reasoning_effort` but the upstream receives no thinking controls.
 
-**Cause:** Thinking transcoding is disabled. The feature gate `[transcoder.features]` defaults to `thinking = false`.
+**Cause:** Thinking transcoding was explicitly disabled with
+`[transcoder.features].thinking = false`.
 
-**Fix:** Enable the feature:
+**Fix:** Remove the override or enable the feature:
 
 ```toml
 [transcoder.features]
@@ -562,7 +566,7 @@ native_protocols = ["anthropic"]
 
 **Cause:** Thinking transcoding is likely disabled for streaming. The feature gate affects both streaming and non-streaming paths.
 
-**Fix:** Ensure `[transcoder.features] thinking = true` is set. Check `[transcoder.openai_reasoning_fields]` to verify the streaming delta field name matches your client's expectations (default: `"reasoning"`).
+**Fix:** Remove the disabling override or ensure `[transcoder.features] thinking = true` is set. Check `[transcoder.openai_reasoning_fields]` to verify the streaming delta field name matches your client's expectations (default: `"reasoning"`).
 
 ---
 

{eggpool-0.7.0 → eggpool-0.7.2}/docs/transcoding.md

@@ -59,8 +59,8 @@ prefer_native = true
 tools = true
 # Image / document content parts (default: false)
 vision = false
-# Extended thinking blocks ↔ reasoning_content (default: false)
-thinking = false
+# Extended thinking blocks ↔ reasoning_content (default: true)
+thinking = true
 # OpenAI response_format / json_schema coercion (default: false)
 structured_outputs = false
 # Anthropic-only primitives explicit handling (default: false)

{eggpool-0.7.0 → eggpool-0.7.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "eggpool"
-version = "0.7.0"
+version = "0.7.2"
 description = "A lightweight proxy that aggregates multiple LLM provider accounts behind an OpenAI Chat Completions-compatible endpoint"
 readme = "README.md"
 requires-python = ">=3.11"

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/_share/config.example.toml

@@ -104,6 +104,11 @@ catalog_withdrawal_policy = "preserve_until_health"
 loss_policy = "warn"      # "warn" or "reject" for loss-of-information
 prefer_native = true      # prefer native-protocol accounts in routing
 
+# Thinking/reasoning transcoding is enabled by default. Set this to false to
+# drop thinking controls and content instead of translating them.
+# [transcoder.features]
+# thinking = false
+
 # Global effort→budget token mapping for thinking/reasoning.
 # Values are used when the model's capability does not carry a
 # per-model effort_to_budget_tokens mapping.
@@ -232,7 +237,7 @@ ttl_seconds = 86400
 # metadata with display names, context limits, benchmarks, and pipeline
 # tags. Manual overrides and alias expansion are also supported.
 # [model_info]
-# enabled = false                  # optional metadata sidecar (default: false)
+# enabled = true                   # metadata sidecar (default: true)
 # include_in_models_endpoint = true  # compact fields in /v1/models (default: true)
 # refresh_interval_s = 3600        # global refresh cadence (seconds)
 # sparse_refresh_interval_s = 300  # accelerated refresh for sparse models (seconds)
@@ -288,7 +293,7 @@ ttl_seconds = 86400
 enabled = false
 
 [model_info]
-enabled = false
+enabled = true
 
 [update_checker]
 enabled = false
@@ -361,6 +366,8 @@ cleanup_max_rows_per_pass = 5000
 allowed_hosts = []
 cors_origins = []
 # Honor X-Forwarded-For/X-Real-IP only when the immediate peer is listed here.
+# Configure only proxies that overwrite or correctly append these headers;
+# X-Forwarded-For uses the nearest untrusted hop in the chain.
 # An empty list ignores forwarded client-attribution headers.
 trusted_proxies = []
 redact_headers = ["authorization", "x-api-key"]
@@ -433,7 +440,8 @@ enabled = false
 # header = "Authorization"
 # scheme = "Bearer"
 #
-# # OpenCode Go exposes low/medium/high thinking controls for mimo-v2.5.
+# # OpenCode Go exposes low/medium/high thinking controls for mimo-v2.5 and
+# # minimal/low/medium/high/xhigh for muse-spark-1.2-contributor.
 # [providers.opencode-go.model_capabilities."mimo-v2.5".thinking]
 # status = "supported"
 # source = "provider_catalog"
@@ -912,5 +920,3 @@ protocol = "anthropic"
 
 [model_overrides."claude-sonnet-4"]
 protocol = "anthropic"
-
-

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/accounts/state.py

@@ -79,10 +79,12 @@ class AccountRuntimeState:
         if now is None:
             now = time.time()
         if self.health_state == "quota_exhausted":
-            if self.cooldown_until > 0 and now >= self.cooldown_until:
+            if self.cooldown_until == 0.0 or now >= self.cooldown_until:
                 self.health_state = "healthy"
                 self.cooldown_until = 0.0
                 self.consecutive_failures = 0
+            # Quota exhaustion has its own expiry semantics; do not fall
+            # through to the generic transient-state branch.
             return
         if self.health_state in ("rate_limited", "cooldown") and (
             self.cooldown_until == 0.0 or now >= self.cooldown_until
@@ -99,8 +101,6 @@ class AccountRuntimeState:
             return False
         if self.health_state in ("authentication_failed",):
             return False
-        if self.health_state == "quota_exhausted":
-            return self.cooldown_until > 0 and self.cooldown_until <= now
         return self.cooldown_until <= now
 
     def record_success(self) -> None:

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/api/proxy_request.py

@@ -257,13 +257,15 @@ def get_client_ip(
     if peer in trusted_proxies:
         forwarded_for = request.headers.get("x-forwarded-for")
         if forwarded_for is not None:
-            first_forwarded = next(
-                (part.strip() for part in forwarded_for.split(",") if part.strip()),
-                "",
-            )
-            forwarded_ip = _valid_forwarded_client_ip(first_forwarded)
-            if forwarded_ip is not None:
-                return forwarded_ip
+            # Walk from the EggPool-facing side of the chain.  The first
+            # valid hop that is not itself trusted is the nearest identity
+            # that the configured proxy chain can attest to; leftmost values
+            # remain client-controlled when a proxy appends rather than
+            # overwrites the header.
+            for part in reversed(forwarded_for.split(",")):
+                forwarded_ip = _valid_forwarded_client_ip(part)
+                if forwarded_ip is not None and forwarded_ip not in trusted_proxies:
+                    return forwarded_ip
 
         real_ip = _valid_forwarded_client_ip(request.headers.get("x-real-ip"))
         if real_ip is not None:
@@ -474,7 +476,7 @@ async def handle_proxy_request(
     except Exception as exc:
         request_state = getattr(request, "state", None)
         request_id_or_none = (
-            getattr(request_state, "request_id", None)
+            getattr(request_state, "proxy_request_id", None)
             if request_state is not None
             else None
         )
@@ -625,6 +627,21 @@ async def _handle_proxy_request_inner(
     with _span(span_recorder, SPAN_MODEL_PARSE):
         model_id, provider_id = parse_model_provider(model_value, known_providers)
 
+    # Normalize the payload's model field to the parsed model_id BEFORE the
+    # transcode preflight runs. The preflight caches the translated payload
+    # and a JSON-encoded body so the coordinator can skip a duplicate encode
+    # on the hot path. If the cached body still carries a provider-suffixed
+    # ``model-id/provider-id`` value (which is how ``/v1/models`` exposes
+    # provider-scoped entries), the upstream receives the suffix and rejects
+    # the request with ``Model <model-id>/<provider-id> is not supported``
+    # because the upstream does not understand eggpool's namespace. Rewriting
+    # the in-memory snapshot here keeps both the preflight encoded body and
+    # the provider-bound payload consistent.
+    preflight_payload = payload
+    if model_id != model_value:
+        preflight_payload = dict(payload)
+        preflight_payload["model"] = model_id
+
     # Preflight context limit check (guardrail, not primary enforcement).
     catalog = lease.runtime.catalog
     transcoder_policy = lease.runtime.transcoder_policy
@@ -662,7 +679,7 @@ async def _handle_proxy_request_inner(
                 model_id=model_id,
                 provider_id=provider_id,
                 client_protocol=endpoint.protocol,
-                payload=payload,
+                payload=preflight_payload,
                 transcoder_policy=transcoder_policy,
             )
             if preflight is not None:

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/catalog/capabilities.py

@@ -527,10 +527,11 @@ def aggregate_thinking_capabilities(
     ]
     budget_min = max(mins) if mins else None
     budget_max = min(maxes) if maxes else None
-    # Invariant: min <= max.  If violated, fall back to None.
+    # A collapsed model can span providers whose bounds do not overlap.  An
+    # aggregate minimum would then be unsafe for providers with a lower
+    # maximum, so retain the conservative upper bound and omit the minimum.
     if budget_min is not None and budget_max is not None and budget_min > budget_max:
         budget_min = None
-        budget_max = None
 
     # Merge effort metadata conservatively: retain the lowest budget for
     # each effort so the aggregate never promises more than one provider

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/catalog/models_dev.py

@@ -21,6 +21,7 @@ MODELS_DEV_BASE_URL = "https://models.dev"
 OPENCODE_GO_MODELS_DEV_PROVIDER_ID = "opencode-go"
 OPENCODE_COMPATIBLE_EFFORTS = ["low", "medium", "high"]
 _EFFORT_BUDGET_DEFAULTS = {
+    "minimal": 1024,
     "low": 1024,
     "medium": 4096,
     "high": 16384,
@@ -116,6 +117,33 @@ def derive_opencode_go_supported_efforts(
     """
     if metadata.get("reasoning") is not True:
         return []
+
+    # models.dev mirrors OpenCode's provider catalog and can carry the exact
+    # effort set for newer models. Prefer it when present so newly released
+    # levels (for example Muse Spark's ``xhigh``) are not silently dropped.
+    reasoning_options = metadata.get("reasoning_options")
+    if isinstance(reasoning_options, list):
+        efforts: list[str] = []
+        for option in cast("list[object]", reasoning_options):
+            if not isinstance(option, dict):
+                continue
+            option_dict = cast("dict[str, object]", option)
+            if option_dict.get("type") != "effort":
+                continue
+            values = option_dict.get("values")
+            if not isinstance(values, list):
+                continue
+            for value in cast("list[object]", values):
+                if not isinstance(value, str):
+                    continue
+                effort = value.strip().lower()
+                if effort == "med":
+                    effort = "medium"
+                if effort and effort not in efforts:
+                    efforts.append(effort)
+        if efforts:
+            return efforts
+
     lowered = model_id.lower()
     if "deepseek-v4" in lowered:
         return [*OPENCODE_COMPATIBLE_EFFORTS, "max"]

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/catalog/pricing.py

@@ -7,7 +7,7 @@ import math
 import re
 from dataclasses import dataclass
 from decimal import Decimal, InvalidOperation
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, cast
 
 from eggpool.constants import (
     DEFAULT_PROVIDER_ID,
@@ -250,11 +250,21 @@ _MAX_ESTIMATED_COST_PER_TOKEN_MICRODOLLARS = 100  # $0.10 / token
 MAX_RESERVATION_COST_MICRODOLLARS = 1_000_000  # $1.00 per reservation
 
 
+def usage_includes_cached_input(usage: object) -> bool:
+    """Return whether an OpenAI usage object reports cache as a subset."""
+    if not isinstance(usage, dict):
+        return False
+    usage_dict = cast("dict[str, Any]", usage)
+    prompt_details = usage_dict.get("prompt_tokens_details")
+    return isinstance(prompt_details, dict) and "cached_tokens" in prompt_details
+
+
 def total_billable_tokens(
     input_tokens: int,
     output_tokens: int,
     cache_read_tokens: int | None = None,
     cache_write_tokens: int | None = None,
+    reasoning_tokens: int | None = None,
 ) -> int:
     """Sum the billable token categories for a request.
 
@@ -265,7 +275,13 @@ def total_billable_tokens(
     """
     read = cache_read_tokens or 0
     write = cache_write_tokens or 0
-    return input_tokens + output_tokens + read + write
+    # Providers commonly report reasoning tokens as a breakdown of
+    # completion_tokens, but some compatible providers report the breakdown
+    # without including it in output_tokens.  Use the larger value so the
+    # reasoning-only shape is billable without double-counting the standard
+    # OpenAI shape.
+    output = max(output_tokens, reasoning_tokens or 0)
+    return input_tokens + output + read + write
 
 
 def is_plausible_request_cost(
@@ -275,6 +291,7 @@ def is_plausible_request_cost(
     output_tokens: int,
     cache_read_tokens: int | None = None,
     cache_write_tokens: int | None = None,
+    reasoning_tokens: int | None = None,
     max_cost_per_token: int = _MAX_TRUSTED_COST_PER_TOKEN_MICRODOLLARS,
 ) -> bool:
     """Whether a cost value is plausible for the billable tokens supplied.
@@ -291,6 +308,7 @@ def is_plausible_request_cost(
         output_tokens,
         cache_read_tokens,
         cache_write_tokens,
+        reasoning_tokens,
     )
     if total_tokens <= 0:
         return False
@@ -306,6 +324,7 @@ def choose_bounded_estimated_cost(
     output_tokens: int,
     cache_read_tokens: int | None = None,
     cache_write_tokens: int | None = None,
+    reasoning_tokens: int | None = None,
 ) -> tuple[int, str]:
     """Pick the most defensible bounded estimate for a request.
 
@@ -336,6 +355,7 @@ def choose_bounded_estimated_cost(
         output_tokens=output_tokens,
         cache_read_tokens=cache_read_tokens,
         cache_write_tokens=cache_write_tokens,
+        reasoning_tokens=reasoning_tokens,
         max_cost_per_token=_MAX_ESTIMATED_COST_PER_TOKEN_MICRODOLLARS,
     )
     reservation_bounded = (
@@ -355,6 +375,7 @@ def choose_bounded_estimated_cost(
             output_tokens=output_tokens,
             cache_read_tokens=cache_read_tokens,
             cache_write_tokens=cache_write_tokens,
+            reasoning_tokens=reasoning_tokens,
             max_cost_per_token=_MAX_ESTIMATED_COST_PER_TOKEN_MICRODOLLARS,
         )
     )
@@ -364,6 +385,7 @@ def choose_bounded_estimated_cost(
         output_tokens,
         cache_read_tokens,
         cache_write_tokens,
+        reasoning_tokens,
     )
     if total_tokens <= 0:
         return 0, "unknown"
@@ -622,6 +644,7 @@ class CostCalculator:
         provider_id: str | None = None,
         *,
         input_tokens_include_cache: bool = False,
+        reasoning_tokens: int = 0,
     ) -> tuple[int, str]:
         """Calculate cost in microdollars from token usage.
 
@@ -637,6 +660,11 @@ class CostCalculator:
                 token counts, so they are subtracted before pricing;
                 when ``False`` (Anthropic-protocol usage) all four
                 categories are disjoint and billed as given.
+            reasoning_tokens: Optional reasoning-token breakdown. Providers
+                usually include this in ``output_tokens``; when they do not,
+                the larger of the two values is priced to avoid dropping
+                reasoning-only usage without double-counting the standard
+                shape.
 
         Returns:
             Tuple of (cost_microdollars, exactness_level). Exactness is
@@ -651,6 +679,8 @@ class CostCalculator:
         output_tokens = _normalize_token_count(output_tokens)
         cache_read_tokens = _normalize_token_count(cache_read_tokens)
         cache_write_tokens = _normalize_token_count(cache_write_tokens)
+        reasoning_tokens = _normalize_token_count(reasoning_tokens)
+        output_tokens = max(output_tokens, reasoning_tokens)
 
         if input_tokens_include_cache:
             # OpenAI-protocol ``prompt_tokens`` includes the cached

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/catalog/service.py

@@ -561,26 +561,37 @@ class CatalogService:
                 )
                 live_task_accounts.append(state.name)
 
-            if static_tasks:
-                await asyncio.gather(*static_tasks, return_exceptions=True)
-            if live_tasks:
-                fetch_results = await asyncio.gather(
-                    *live_tasks, return_exceptions=True
-                )
-                for account_name, fetch_result in zip(
-                    live_task_accounts, fetch_results, strict=True
-                ):
-                    if isinstance(fetch_result, BaseException):
-                        logger.warning(
-                            "Catalog refresh for account %r preserved prior "
-                            "support after exception: %s",
-                            account_name,
-                            fetch_result,
-                        )
-                        outcomes_by_account[account_name] = AccountCatalogOutcome.FAILED
-                        continue
-                    outcome, _update = fetch_result
-                    outcomes_by_account[account_name] = outcome
+            try:
+                if static_tasks:
+                    await asyncio.gather(*static_tasks, return_exceptions=True)
+                if live_tasks:
+                    fetch_results = await asyncio.gather(
+                        *live_tasks, return_exceptions=True
+                    )
+                    for account_name, fetch_result in zip(
+                        live_task_accounts, fetch_results, strict=True
+                    ):
+                        if isinstance(fetch_result, BaseException):
+                            logger.warning(
+                                "Catalog refresh for account %r preserved prior "
+                                "support after exception: %s",
+                                account_name,
+                                fetch_result,
+                            )
+                            outcomes_by_account[account_name] = (
+                                AccountCatalogOutcome.FAILED
+                            )
+                            continue
+                        outcome, _update = fetch_result
+                        outcomes_by_account[account_name] = outcome
+            finally:
+                pending_tasks: list[asyncio.Task[Any]] = [
+                    task for task in (*static_tasks, *live_tasks) if not task.done()
+                ]
+                for task in pending_tasks:
+                    task.cancel()
+                if pending_tasks:
+                    await asyncio.gather(*pending_tasks, return_exceptions=True)
 
             # Keep one outcome per enabled account.  In particular, an
             # account skipped before task creation must not shift the result

{eggpool-0.7.0 → eggpool-0.7.2}/src/eggpool/config_reload_policy.py

@@ -29,6 +29,8 @@ from dataclasses import dataclass
 from enum import Enum
 from typing import TYPE_CHECKING, Final, cast
 
+from eggpool.config_validation import ConfigInternalError
+
 if TYPE_CHECKING:
     from pathlib import Path
 
@@ -578,8 +580,11 @@ def _to_dict(value: object) -> dict[str, object]:
         try:
             dumped: object = dump()
             return dict(cast("Iterable[tuple[str, object]]", dumped))
-        except Exception:
-            return {}
+        except Exception as exc:
+            raise ConfigInternalError(
+                "configuration diff failed while serializing "
+                f"{type(value).__qualname__}: {type(exc).__qualname__}"
+            ) from exc
     if hasattr(value, "__dict__"):
         return {
             k: v