coderouter-cli 2.7.2__tar.gz → 2.7.3__tar.gz

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/CHANGELOG.md

@@ -6,6 +6,46 @@ versioning follows [SemVer](https://semver.org/).
 
 ---
 
+## [v2.7.3] — 2026-07-05 (Per-request empty-response fallback + L3 bench evidence)
+
+Two PRs: **PR #45** (`empty_response_action` — the fallback leg of the
+three-tier tool-call reliability story) and **PR #46** (L3 benchmark
+refresh: model matrix, repairer-version A/B archives, results). Fully
+backward compatible: new knob defaults to `off`, no breaking config
+changes, no new dependencies.
+
+### Added
+
+- **`FallbackChain.empty_response_action: off | warn | fallback`**
+  (default `off`). With `fallback`, a 200 response that is *content-empty*
+  (no `tool_use`, no non-whitespace text; thinking-only counts as empty) is
+  re-dispatched in-flight to the next provider in the chain. Non-streaming
+  responses are judged on the finalized object; streaming buffers events
+  until the first real content token and discards an unobserved stream.
+  Emptiness is judged on content, not `usage.output_tokens` (unreliable on
+  some backends). Motivation: gemma4:26b returns blank 200s on
+  `no_tool_temptation`-type prompts (20/20 at temperature 0, direct AND via
+  router) — no text exists for the repair layer to fix, and the windowed
+  drift guard `empty_response_rate` cannot rescue a single blank turn
+  in-flight. 21 new tests. (PR #45)
+- **Benchmark refresh**: full L3 model matrix in
+  `benchmarks/tool-repair/providers.bench.yaml` (llama3.2/3.1, mistral,
+  qwen2.5-coder:1.5b, phi4-mini + originals, `bench-gemma-chain` enabled),
+  one-shot matrix runner `bench_l3_p1.sh`, latest live results plus
+  repairer-version A/B snapshots under `results/archive-v2.7.{0,1}/`.
+  (PR #46)
+
+### Results
+
+- Live chain validation (M3 Max, temperature 0, 100 requests, zero
+  errors): gemma4:26b → qwen3-coder:30b chain **80% → 100%** native —
+  all 20 blank turns rescued.
+- Three-tier story now fully measured: repair rescues broken-but-present
+  calls (qwen2.5×2 / mistral 0→100, phi4 +60pt), healthy models pass
+  through undegraded, and fallback closes what repair cannot touch.
+
+---
+
 ## [v2.7.2] — 2026-07-05 (R4 tool-call repair: nested-XML / JSON-envelope / call-syntax forms)
 
 Single feature PR **#43**, driven by the 2026-07-05 L3 live benchmark

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coderouter-cli
-Version: 2.7.2
+Version: 2.7.3
 Summary: Local-first, free-first, fallback-built-in LLM router. Claude Code / OpenAI compatible.
 Project-URL: Homepage, https://github.com/zephel01/CodeRouter
 Project-URL: Repository, https://github.com/zephel01/CodeRouter

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/coderouter/config/schemas.py

@@ -718,6 +718,57 @@ class FallbackChain(BaseModel):
         ),
     )
 
+    # --- ⑧ (empty-response): per-request empty-response fallback ------------
+    #
+    # Some local backends (observed: gemma4:26b on ``no_tool_temptation``
+    # prompts) return a 200 with a structurally-valid but *content-empty*
+    # Anthropic response — no tool_use and no non-whitespace text — for a
+    # fraction of requests. The drift guard's ``empty_response_rate`` is a
+    # windowed aggregate (default threshold 0.3, ``min_window_fill`` 6): it
+    # promotes/reloads a backend once the *rate* is bad, but cannot rescue a
+    # single blank turn in-flight. This knob adds a per-request in-flight
+    # fallback that re-dispatches the *same* request to the next provider in
+    # the chain the moment an empty response is detected.
+    #
+    # Design: "empty" is judged on *content*, not usage.output_tokens (which
+    # some backends report unreliably). A response is empty when its content
+    # list is empty, or every block is either a whitespace-only ``text`` block
+    # or a ``thinking`` block — i.e. nothing the client can act on. A single
+    # ``tool_use`` block or one non-whitespace ``text`` block makes it non-empty.
+    #
+    #   * ``off``      — no detection, no fallback, no log. Backward-compatible
+    #                    default (byte-for-byte identical to pre-⑧ behavior).
+    #   * ``warn``     — detect + emit an ``empty-response-detected`` log line
+    #                    only; the empty response is returned unchanged.
+    #   * ``fallback`` — on an empty response, log ``empty-response-detected``
+    #                    and continue to the next provider (the empty response
+    #                    is *not* recorded as an error). If every provider in
+    #                    the chain returns empty, the last empty response is
+    #                    returned as-is (a 200 blank is a legitimate answer)
+    #                    with ``chain_exhausted=True`` on the log line. On the
+    #                    streaming path, ``fallback`` buffers events until real
+    #                    content is observed, so an empty stream can be swapped
+    #                    to the next provider without the client seeing bytes.
+    #
+    # Default ``off`` preserves complete backward compatibility; the original
+    # request object is never mutated, so a later provider always receives the
+    # untouched request.
+    empty_response_action: Literal["off", "warn", "fallback"] = Field(
+        default="off",
+        description=(
+            "⑧ (empty-response): action when a provider returns a 200 with "
+            "empty content (no tool_use and no non-whitespace text; a "
+            "thinking-only response counts as empty). ``off`` (default) does "
+            "nothing. ``warn`` emits an ``empty-response-detected`` log only "
+            "and returns the empty response. ``fallback`` re-dispatches the "
+            "same request to the next provider (empty is not counted as an "
+            "error); if the whole chain returns empty, the last empty response "
+            "is returned unchanged. Streaming buffers events until real "
+            "content appears so empty streams can be swapped provider-side "
+            "without the client seeing any bytes."
+        ),
+    )
+
     # --- S2 (shim): tool_choice capability gate + emulation -----------------
     #
     # Anthropic clients (Claude Code, SDKs) can pin the model to a specific

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/coderouter/logging.py

@@ -1372,6 +1372,46 @@ def log_partial_stitch_surfaced(
     )
 
 
+# ---------------------------------------------------------------------------
+# ⑧ (empty-response): per-request empty-response fallback log shape
+#
+# Fired when a provider returns a 200 with content-empty output (no
+# tool_use and no non-whitespace text). The single event lane carries the
+# ``action`` (warn|fallback) that triggered it, whether the empty response
+# came from the streaming or non-streaming path, and ``chain_exhausted``
+# — True only when every provider in the chain returned empty and the last
+# empty response is being returned to the client as-is.
+# ---------------------------------------------------------------------------
+
+
+def log_empty_response_detected(
+    logger: logging.Logger,
+    provider: str,
+    *,
+    action: str,
+    stream: bool,
+    chain_exhausted: bool = False,
+) -> None:
+    """Emit an ``empty-response-detected`` warn line.
+
+    Fired when ``empty_response_action`` is ``warn`` or ``fallback`` and a
+    provider returns a structurally-valid but content-empty response. The
+    ``action`` field records which knob value produced the line; ``stream``
+    distinguishes the SSE path from the non-streaming path; ``chain_exhausted``
+    is True only when the whole chain returned empty and the last empty
+    response is being returned unchanged.
+    """
+    logger.warning(
+        "empty-response-detected",
+        extra={
+            "provider": provider,
+            "action": action,
+            "stream": stream,
+            "chain_exhausted": chain_exhausted,
+        },
+    )
+
+
 # ---------------------------------------------------------------------------
 # v2.0-I: Continuous probe log shapes
 #

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/coderouter/metrics/collector.py

@@ -52,6 +52,7 @@ Event inventory (dispatch table in :meth:`MetricsCollector._dispatch`)
     ``drift-promoted`` (v2.0-G)  → ``drift_promoted_total``
     ``drift-reload-attempted``   → ``drift_reload_total`` / success
     ``partial-stitch-surfaced``  → ``partial_stitch_surfaced_total`` (v2.0-H)
+    ``empty-response-detected``  → ``empty_responses_total`` + per-provider (⑧)
     ``probe-completed`` (v2.0-I) → ``probe_total`` / ``probe_success`` / ``probe_failure``
                                    + per-provider latency gauge
     ``probe-round-completed``    → ``probe_rounds_total`` (v2.0-I)
@@ -109,6 +110,7 @@ _KNOWN_EVENTS: Final[frozenset[str]] = frozenset(
         "drift-promoted",
         "drift-reload-attempted",
         "partial-stitch-surfaced",
+        "empty-response-detected",
         "probe-completed",
         "probe-round-completed",
         "probe-capabilities-drift",
@@ -272,6 +274,13 @@ class MetricsCollector(logging.Handler):
         # client instead of a generic error event.
         self._partial_stitch_surfaced_total: int = 0
 
+        # ⑧ (empty-response): count of per-request empty-response detections
+        # (both ``warn`` and ``fallback`` actions). Aggregate + per-provider
+        # so a dashboard can spot which backend blanks out. Counts every
+        # detection event, including the terminal chain-exhausted one.
+        self._empty_responses_total: int = 0
+        self._empty_responses_by_provider: Counter[str] = Counter()
+
         # v2.0-I: continuous probe counters. Per-provider probe attempts
         # and outcomes, plus a round counter for the dashboard's
         # "probes/min" panel.
@@ -329,6 +338,7 @@ class MetricsCollector(logging.Handler):
             "drift-promoted": self._on_drift_promoted,
             "drift-reload-attempted": self._on_drift_reload_attempted,
             "partial-stitch-surfaced": self._on_partial_stitch_surfaced,
+            "empty-response-detected": self._on_empty_response_detected,
             "probe-completed": self._on_probe_completed,
             "probe-round-completed": self._on_probe_round_completed,
             "probe-capabilities-drift": self._on_probe_capabilities_drift,
@@ -670,6 +680,17 @@ class MetricsCollector(logging.Handler):
         self._partial_stitch_surfaced_total += 1
         self._push_recent("partial-stitch-surfaced", extras, record)
 
+    def _on_empty_response_detected(
+        self, extras: dict[str, Any], record: logging.LogRecord
+    ) -> None:
+        # ⑧ (empty-response): a content-empty 200 was detected under a
+        # ``warn`` / ``fallback`` action.
+        self._empty_responses_total += 1
+        provider = _str(extras.get("provider"))
+        if provider:
+            self._empty_responses_by_provider[provider] += 1
+        self._push_recent("empty-response-detected", extras, record)
+
     def _on_probe_completed(
         self, extras: dict[str, Any], record: logging.LogRecord
     ) -> None:
@@ -888,6 +909,11 @@ class MetricsCollector(logging.Handler):
                     "drift_reload_success_total": self._drift_reload_success_total,
                     # v2.0-H (L6): partial stitch surfaced.
                     "partial_stitch_surfaced_total": self._partial_stitch_surfaced_total,
+                    # ⑧ (empty-response): per-request empty-response fallback.
+                    "empty_responses_total": self._empty_responses_total,
+                    "empty_responses_by_provider": dict(
+                        self._empty_responses_by_provider
+                    ),
                     # v2.0-I: continuous probe counters.
                     "probe_total": dict(self._probe_total),
                     "probe_success": dict(self._probe_success),
@@ -1042,6 +1068,9 @@ class MetricsCollector(logging.Handler):
             self._tokens_saved_by_mechanism.clear()
             # v2.0-H (L6)
             self._partial_stitch_surfaced_total = 0
+            # ⑧ (empty-response)
+            self._empty_responses_total = 0
+            self._empty_responses_by_provider.clear()
             # v2.0-I
             self._probe_total.clear()
             self._probe_success.clear()

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/coderouter/routing/fallback.py

@@ -76,6 +76,7 @@ from coderouter.logging import (
     log_chain_memory_pressure_blocked,
     log_chain_paid_gate_blocked,
     log_demote_unhealthy_provider,
+    log_empty_response_detected,
     log_memory_pressure_detected,
     log_skip_budget_exceeded,
     log_skip_memory_pressure,
@@ -973,6 +974,82 @@ def _warn_if_uniform_auth_failure(errors: list[AdapterError], *, profile: str) -
     )
 
 
+# ---------------------------------------------------------------------------
+# ⑧ (empty-response): content-based emptiness judgement + stream buffering
+# ---------------------------------------------------------------------------
+
+
+def _block_field(block: Any, key: str) -> Any:
+    """Read ``key`` from a content block whether it is a dict or a model.
+
+    Anthropic content blocks reach the engine as plain dicts (openai_compat
+    translation, most test fakes) or as pydantic-ish objects (native
+    passthrough). This mirrors the accessor pattern already used by the
+    drift fingerprint at the success-path tail.
+    """
+    if isinstance(block, dict):
+        return block.get(key)
+    return getattr(block, key, None)
+
+
+def _anthropic_response_is_empty(resp: AnthropicResponse) -> bool:
+    """Return True when ``resp`` carries no client-actionable content.
+
+    "Empty" is judged on content, never on ``usage.output_tokens`` (some
+    backends report that unreliably). A response is empty when:
+        - its ``content`` list is empty / falsy, or
+        - every block is either a whitespace-only ``text`` block or a
+          ``thinking`` block.
+
+    A single ``tool_use`` block, or one ``text`` block with any
+    non-whitespace character, makes the response non-empty. Unknown block
+    types are treated conservatively as *content* (non-empty) so a novel
+    actionable block is never silently swallowed.
+    """
+    content = getattr(resp, "content", None)
+    if not content:
+        return True
+    for block in content:
+        btype = _block_field(block, "type")
+        if btype == "text":
+            text = _block_field(block, "text") or ""
+            if text.strip():
+                return False
+            # whitespace-only text → keep scanning
+            continue
+        if btype == "thinking":
+            # thinking-only carries nothing the client can act on
+            continue
+        # tool_use or any other (unknown → conservatively non-empty) block
+        return False
+    return True
+
+
+def _stream_event_is_real_content(event: AnthropicStreamEvent) -> bool:
+    """Return True when ``event`` is the first byte of actionable content.
+
+    Real content is either a ``content_block_start`` opening a ``tool_use``
+    block, or a ``content_block_delta`` carrying a non-whitespace text
+    delta. ``message_start`` / ``ping`` / empty-text openers do not count —
+    they are the buffered preamble that an empty stream also emits.
+    """
+    data = getattr(event, "data", None) or {}
+    etype = data.get("type") if isinstance(data, dict) else None
+    if etype == "content_block_start":
+        block = data.get("content_block") or {}
+        return isinstance(block, dict) and block.get("type") == "tool_use"
+    if etype == "content_block_delta":
+        delta = data.get("delta") or {}
+        if isinstance(delta, dict):
+            # text_delta with non-whitespace text is real content;
+            # input_json_delta (tool args) is real content too.
+            if delta.get("type") == "text_delta":
+                return bool((delta.get("text") or "").strip())
+            if delta.get("type") == "input_json_delta":
+                return True
+    return False
+
+
 class FallbackEngine:
     """Sequential fallback router — the core of CodeRouter.
 
@@ -1717,6 +1794,20 @@ class FallbackEngine:
             getattr(profile, "cache_control_action", "off"),
         )
 
+    def _resolve_empty_response_action(self, profile_name: str | None) -> str:
+        """⑧ (empty-response): resolve ``empty_response_action`` for a profile.
+
+        Defaults to ``off`` so a missing / stub profile (e.g. a
+        ``__new__``-constructed test engine) yields the backward-compatible
+        no-op. Resolved once at the top of each Anthropic entry point.
+        """
+        chosen = profile_name or self.config.default_profile
+        try:
+            profile = self.config.profile_by_name(chosen)
+        except (KeyError, ValueError):
+            return "off"
+        return getattr(profile, "empty_response_action", "off")
+
     def _resolve_chain(self, profile_name: str | None) -> list[BaseAdapter]:
         """Return the list of adapters to try, in order, for this profile.
 
@@ -2431,6 +2522,11 @@ class FallbackEngine:
         tool_choice_action, cache_control_action = self._resolve_shim_actions(
             request.profile
         )
+        # ⑧ (empty-response): resolve the per-request empty-response action
+        # once. ``last_empty_resp`` holds the most recent content-empty 200
+        # under ``fallback`` so a fully-empty chain returns it verbatim.
+        empty_response_action = self._resolve_empty_response_action(request.profile)
+        last_empty_resp: AnthropicResponse | None = None
 
         for adapter, will_degrade in chain:
             is_native = isinstance(adapter, AnthropicAdapter)
@@ -2585,6 +2681,26 @@ class FallbackEngine:
                 stream=False,
                 response_fingerprint=_fp(_fp_text) if _fp_text else None,
             )
+            # ⑧ (empty-response): per-request empty-response handling. Runs
+            # after the drift observation above (which still gets the real
+            # output_tokens=0 signal for the windowed empty_response_rate)
+            # but before the cache-observed / observer fanout, so a
+            # ``fallback`` swap does not emit a "completed" line for a
+            # response the client never sees.
+            if empty_response_action != "off" and _anthropic_response_is_empty(resp):
+                if empty_response_action == "fallback":
+                    # Remember this empty response so a fully-empty chain can
+                    # return it verbatim, then try the next provider. Not
+                    # recorded in ``errors`` — a 200 blank is not a failure.
+                    last_empty_resp = resp
+                    log_empty_response_detected(
+                        logger, adapter.name, action="fallback", stream=False
+                    )
+                    continue
+                # ``warn``: log only; fall through and return the empty resp.
+                log_empty_response_detected(
+                    logger, adapter.name, action="warn", stream=False
+                )
             # v1.9-A: pair every successful Anthropic response with a
             # cache-observed log line. Native Anthropic / LM Studio
             # /v1/messages report cache_read_input_tokens /
@@ -2625,6 +2741,20 @@ class FallbackEngine:
             )
             return resp
 
+        # ⑧ (empty-response): the chain is exhausted. Under ``fallback``,
+        # if every provider that answered returned an empty 200, return the
+        # last empty response verbatim (a 200 blank is a legitimate answer)
+        # rather than raising — errors is empty in that case anyway.
+        if last_empty_resp is not None:
+            log_empty_response_detected(
+                logger,
+                last_empty_resp.coderouter_provider or "unknown",
+                action="fallback",
+                stream=False,
+                chain_exhausted=True,
+            )
+            return last_empty_resp
+
         profile = request.profile or self.config.default_profile
         _warn_if_uniform_auth_failure(errors, profile=profile)
         raise NoProvidersAvailableError(profile=profile, errors=errors)
@@ -2676,6 +2806,15 @@ class FallbackEngine:
         tool_choice_action, cache_control_action = self._resolve_shim_actions(
             request.profile
         )
+        # ⑧ (empty-response): resolve the per-request empty-response action.
+        # Only ``fallback`` changes the streaming path (it buffers events
+        # until real content appears); ``off`` / ``warn`` stream unchanged.
+        # ``last_empty_stream_buffer`` holds the buffered preamble of the
+        # most recent empty stream so a fully-empty chain can flush it and
+        # terminate normally rather than raising.
+        empty_response_action = self._resolve_empty_response_action(request.profile)
+        empty_fallback = empty_response_action == "fallback"
+        last_empty_stream_buffer: list[AnthropicStreamEvent] | None = None
 
         for adapter, will_degrade in chain:
             is_native = isinstance(adapter, AnthropicAdapter)
@@ -2828,15 +2967,59 @@ class FallbackEngine:
             self._observe_provider_success(
                 adapter.name, profile=request.profile
             )
-            acc.observe(first)
-            yield first
+            # ⑧ (empty-response): under ``fallback`` we withhold the opening
+            # events (message_start / empty content_block_start / ping) from
+            # the client until the *first real content* event is observed.
+            # Because no bytes have reached the client, an empty stream can
+            # still be swapped for the next provider. ``off`` / ``warn`` keep
+            # the legacy immediate-yield behavior (byte-for-byte unchanged).
+            #
+            # ``content_started`` flips True the moment real content is seen
+            # (or immediately, when the action is not ``fallback``); from
+            # then on events pass straight through and the mid-stream guard
+            # is live exactly as before.
+            content_started = not empty_fallback
+            buffer: list[AnthropicStreamEvent] = []
             # Mid-stream guard identical to stream() — any error after the
-            # first event is terminal.
+            # first *forwarded* event is terminal.
             try:
-                async for ev in event_iter:
+                # Seed the loop with the already-fetched ``first`` event,
+                # then drain the rest of the iterator.
+                pending = first
+                iterator = event_iter.__aiter__()
+                while True:
+                    ev = pending
                     acc.observe(ev)
-                    yield ev
+                    if content_started:
+                        yield ev
+                    else:
+                        buffer.append(ev)
+                        if _stream_event_is_real_content(ev):
+                            # Real content arrived — flush the withheld
+                            # preamble in order, then switch to passthrough.
+                            content_started = True
+                            for buffered_ev in buffer:
+                                yield buffered_ev
+                            buffer = []
+                    try:
+                        pending = await iterator.__anext__()
+                    except StopAsyncIteration:
+                        break
             except AdapterError as exc:
+                if not content_started:
+                    # Empty-stream failure before any real content under
+                    # ``fallback``: nothing reached the client, so treat it
+                    # like an empty response and try the next provider. Do
+                    # NOT raise MidStreamError (that would surface to the
+                    # client). Record the error-only observation for adaptive.
+                    self._adaptive.record_attempt(
+                        adapter.name, latency_ms=None, success=False
+                    )
+                    log_empty_response_detected(
+                        logger, adapter.name, action="fallback", stream=True
+                    )
+                    last_empty_stream_buffer = buffer
+                    continue
                 # M2: mid-stream failure — record an error-only
                 # observation (no latency; first-event success already
                 # recorded) so adaptive error rate reflects the breakage.
@@ -2868,6 +3051,17 @@ class FallbackEngine:
                 raise MidStreamError(
                     adapter.name, exc, partial_content=acc.partial_content
                 ) from exc
+            # ⑧ (empty-response): the stream ended cleanly but no real
+            # content was ever observed under ``fallback`` — the preamble is
+            # still buffered (never sent). Treat as an empty response: keep
+            # the buffer for a possible chain-exhausted flush and try the
+            # next provider.
+            if empty_fallback and not content_started:
+                log_empty_response_detected(
+                    logger, adapter.name, action="fallback", stream=True
+                )
+                last_empty_stream_buffer = buffer
+                continue
             # v2.0-G (L4): drift detection observation (stream success).
             # P1-4: compute response fingerprint for goal_progress_stall.
             _stream_fp_text = " ".join(
@@ -2923,6 +3117,23 @@ class FallbackEngine:
             )
             return
 
+        # ⑧ (empty-response): the chain is exhausted under ``fallback`` and
+        # every provider produced an empty stream. Flush the last buffered
+        # (empty) preamble so the client gets a well-formed, terminating SSE
+        # sequence (message_start … message_stop) instead of an error — a
+        # 200 blank is a legitimate answer. errors is empty in that case.
+        if last_empty_stream_buffer is not None:
+            log_empty_response_detected(
+                logger,
+                "unknown",
+                action="fallback",
+                stream=True,
+                chain_exhausted=True,
+            )
+            for buffered_ev in last_empty_stream_buffer:
+                yield buffered_ev
+            return
+
         profile = request.profile or self.config.default_profile
         _warn_if_uniform_auth_failure(errors, profile=profile)
         raise NoProvidersAvailableError(profile=profile, errors=errors)

{coderouter_cli-2.7.2 → coderouter_cli-2.7.3}/pyproject.toml

@@ -11,7 +11,7 @@
 # in plan.md §11.B; once granted, this name will become an alias and
 # `coderouter` will become the canonical distribution name.
 name = "coderouter-cli"
-version = "2.7.2"
+version = "2.7.3"
 description = "Local-first, free-first, fallback-built-in LLM router. Claude Code / OpenAI compatible."
 readme = "README.md"
 requires-python = ">=3.12"