HOTFIX ENG-1081: give the completion verifier room to answer (silent task death, 25% of verdicts)

anton/core/backends/scratchpad_boot.py

@@ -217,6 +217,8 @@ def generate_object(
                 """
                 from anton.core.llm.structured import (
                     build_structured_tool,
+                    looks_truncated,
+                    raise_unusable_tool_call,
                     unwrap_structured_response,
                 )
 
@@ -233,11 +235,24 @@ def generate_object(
                 )
 
                 if not response.tool_calls:
-                    raise ValueError("LLM did not return structured output.")
+                    # Same classification as the async path (ENG-1081).
+                    # Nothing retries here, but the message reaches the model as
+                    # a traceback, so "you ran out of budget" is actionable
+                    # where "did not return structured output" was not.
+                    raise_unusable_tool_call(
+                        response, tool_name=tool["name"], budget=max_tokens
+                    )
 
-                return unwrap_structured_response(
-                    response.tool_calls[0].input, validator_class, is_list
-                )
+                try:
+                    return unwrap_structured_response(
+                        response.tool_calls[0].input, validator_class, is_list
+                    )
+                except Exception:
+                    if looks_truncated(response, max_tokens):
+                        raise_unusable_tool_call(
+                            response, tool_name=tool["name"], budget=max_tokens
+                        )
+                    raise
 
         _scratchpad_llm_instance = _ScratchpadLLM()
 

anton/core/llm/client.py

@@ -199,28 +199,41 @@ async def _generate_object_with(
         """
         from anton.core.llm.structured import (
             build_structured_tool,
+            looks_truncated,
+            raise_unusable_tool_call,
             unwrap_structured_response,
         )
 
         tool, validator_class, is_list = build_structured_tool(schema_class)
 
+        budget = max_tokens or self._max_tokens
+
         response = await provider.complete(
             model=model,
             system=system,
             messages=messages,
             tools=[tool],
             tool_choice={"type": "tool", "name": tool["name"]},
-            max_tokens=max_tokens or self._max_tokens,
+            max_tokens=budget,
         )
 
         if not response.tool_calls:
-            raise ValueError(
-                f"LLM did not return a tool call for forced schema {tool['name']}."
-            )
+            # Shared with the scratchpad's sync twin so both paths classify the
+            # failure identically — see `raise_unusable_tool_call` (ENG-1081).
+            raise_unusable_tool_call(response, tool_name=tool["name"], budget=budget)
 
-        return unwrap_structured_response(
-            response.tool_calls[0].input, validator_class, is_list
-        )
+        try:
+            return unwrap_structured_response(
+                response.tool_calls[0].input, validator_class, is_list
+            )
+        except Exception:
+            # The budget can also run out *inside* the tool call's JSON, so
+            # validation fails here instead. Same cause, same retry (ENG-1081).
+            if looks_truncated(response, budget):
+                raise_unusable_tool_call(
+                    response, tool_name=tool["name"], budget=budget
+                )
+            raise
 
     async def generate_object(
         self,
@@ -263,8 +276,11 @@ async def generate_object(
             input was a list annotation.
 
         Raises:
-            ValueError: If the LLM fails to produce a tool call (rare —
-                forced tool_choice usually prevents this).
+            StructuredOutputError: If the LLM fails to produce a tool call.
+                A ``ValueError`` subclass, so callers that catch
+                ``ValueError`` are unaffected. Check ``.truncated`` to tell a
+                blown ``max_tokens`` budget (retry with more room) from a
+                genuine failure (ENG-1081).
             pydantic.ValidationError: If the tool's input doesn't match
                 the schema.
 

anton/core/llm/provider.py

@@ -312,6 +312,36 @@ class TokenLimitExceeded(Exception):
     """Raised when the LLM returns 429 due to billing/token limits."""
 
 
+class StructuredOutputError(ValueError):
+    """Raised when a forced-tool-call structured-output call yields no usable call.
+
+    ``truncated`` says whether a retry can help: ``True`` means the model spent
+    the ``max_tokens`` budget narrating before it reached the tool call (the
+    narrating aliases — ``mindshub_air``/``kimi``, ``deepseek``, ``qwen`` — do
+    this deterministically under a tight budget, ENG-1081); ``False`` means the
+    provider errored, refused, or returned nothing, and a bigger budget won't
+    fix it. See `structured.looks_truncated` for how that's decided.
+
+    Subclasses ``ValueError`` so call sites catching the documented ``ValueError``
+    from ``generate_object``/``generate_object_code`` keep working.
+    """
+
+    def __init__(
+        self,
+        message: str,
+        *,
+        truncated: bool = False,
+        output_tokens: int = 0,
+        max_tokens: int = 0,
+        stop_reason: str | None = None,
+    ):
+        super().__init__(message)
+        self.truncated = truncated
+        self.output_tokens = output_tokens
+        self.max_tokens = max_tokens
+        self.stop_reason = stop_reason
+
+
 class TransientProviderError(ConnectionError):
     """Raised when the provider fails in a way that a retry might fix.
 

anton/core/llm/structured.py

@@ -27,7 +27,89 @@
 
 from __future__ import annotations
 
-from typing import Any
+from typing import Any, NoReturn
+
+from .provider import StructuredOutputError
+
+
+def looks_truncated(response, budget: int) -> bool:
+    """True if `response` was cut off by the `max_tokens` budget.
+
+    Token count first, because the MindsHub gateway reports
+    ``finish_reason: "stop"`` at the cap for most aliases (ENG-1082) — the
+    standard ``"length"`` check can't be relied on there. Both dialects are
+    honoured when reported: OpenAI says ``"length"``, Anthropic ``"max_tokens"``.
+    No usage information → ``False``; without evidence we don't buy a retry.
+    """
+    usage = getattr(response, "usage", None)
+    output_tokens = getattr(usage, "output_tokens", 0) or 0
+    stop_reason = getattr(response, "stop_reason", None)
+    return stop_reason in ("length", "max_tokens") or (
+        budget > 0 and output_tokens >= budget
+    )
+
+
+def raise_unusable_tool_call(response, *, tool_name: str, budget: int) -> NoReturn:
+    """Raise `StructuredOutputError` explaining *why* the tool call is unusable.
+
+    Covers both shapes of the same underlying problem:
+
+    - **No tool call at all** — the model narrated in plain ``content`` until
+      the budget ran out.
+    - **A damaged tool call** — the budget ran out *inside* the call's JSON
+      arguments, so ``safe_parse_tool_input`` salvaged a partial dict and set
+      ``parse_error``, and validation then fails. Without this, that case
+      surfaces as a bare ``ValidationError``, which reads like a schema bug and
+      never gets the retry a truncation deserves (ENG-1081).
+
+    Lives here, next to `build_structured_tool`/`unwrap_structured_response`,
+    because both structured-output paths must classify the failure the same
+    way — the async `LLMClient._generate_object_with` and the sync
+    `_ScratchpadLLM.generate_object` in the scratchpad subprocess. Keeping it
+    in one of the two callers is how they drift (ENG-1081).
+
+    A forced tool call can come back empty for two very different reasons:
+
+    - **Truncated** — the model narrated in plain ``content`` and ran out of
+      ``budget`` before reaching the call. Retrying with more room usually
+      works. Models served through MindsHub's Fireworks aliases
+      (``mindshub_air``/``kimi``, ``deepseek``) narrate before acting, so a
+      tight budget fails them.
+    - **Anything else** — the provider errored, refused, or returned nothing.
+      A bigger budget won't help.
+
+    See `looks_truncated` for how truncation is detected.
+
+    Args:
+        response: The provider's ``LLMResponse``.
+        tool_name: Name of the forced tool, for the message.
+        budget: The ``max_tokens`` the call was given.
+
+    Raises:
+        StructuredOutputError: Always. ``.truncated`` carries the verdict.
+    """
+    usage = getattr(response, "usage", None)
+    output_tokens = getattr(usage, "output_tokens", 0) or 0
+    stop_reason = getattr(response, "stop_reason", None)
+    truncated = looks_truncated(response, budget)
+    what = (
+        "returned an unusable tool call for"
+        if getattr(response, "tool_calls", None)
+        else "did not return a tool call for"
+    )
+    detail = (
+        f" (truncated: {output_tokens}/{budget} output tokens spent before the "
+        "call was complete)."
+        if truncated
+        else "."
+    )
+    raise StructuredOutputError(
+        f"LLM {what} forced schema {tool_name}{detail}",
+        truncated=truncated,
+        output_tokens=output_tokens,
+        max_tokens=budget,
+        stop_reason=stop_reason,
+    )
 
 
 def build_structured_tool(schema_class) -> tuple[dict, type, bool]:

anton/core/session.py

@@ -39,6 +39,7 @@
     StreamTaskProgress,
     StreamTextDelta,
     StreamToolResult,
+    StructuredOutputError,
     TokenLimitExceeded,
     ToolCall,
     TransientProviderError,
@@ -158,6 +159,81 @@ class _VerifierVerdict(BaseModel):
     reason: str = Field(description="One brief sentence explaining the verdict.")
 
 
+# Output budgets for the verdict call: first attempt, then the retry used when
+# it comes back truncated (ENG-1081).
+#
+# Models that narrate before acting (`mindshub_air`/`kimi`, `deepseek`, `qwen`)
+# spend the budget on prose and never reach the forced tool call. The original
+# 256 truncated them on essentially every call — 98.6% of `mindshub_air` verdicts
+# in prod returned no tool call, which the fail-safe below turned into a silent
+# "task complete".
+#
+# 2048 is sized from a measured distribution, not one sample: 16 identical calls
+# spanned 245–1654 output tokens (median ~290). That 6.7x per-call spread is also
+# why there is no "this model can't do it" latch — one truncation is a tail
+# sample, not proof about the next turn — and why the 4096 retry exists rather
+# than a single bigger budget. 1024 was measurably too small. Nothing pays for
+# headroom it doesn't use; first-party models answer in 43–115 tokens either way.
+_VERIFIER_TOKEN_BUDGETS = (2048, 4096)
+
+# Appended to the verifier system prompt. Shortens the preamble on narrating
+# models but is not sufficient alone (0/3 at 256 with it), so it pairs with the
+# budgets above. Tool name comes off the schema class so it can't go stale.
+_VERIFIER_NO_PREAMBLE = (
+    f"Call the {_VerifierVerdict.__name__} tool immediately as your first "
+    "action. Do not think out loud, restate the conversation, or explain your "
+    "reasoning before calling it — put your one-sentence justification in the "
+    "tool's `reason` field."
+)
+
+
+def _safe_error_detail(exc: BaseException) -> str:
+    """Describe an exception for logs without copying model or user content.
+
+    Neither ``str(exc)`` nor ``repr(exc)`` is safe here: a Pydantic
+    ``ValidationError`` embeds the rejected ``input_value`` — for the verifier
+    that's model-generated text derived from the user's conversation — and
+    provider exceptions can carry response bodies. Emits the exception type,
+    plus for validation errors the field locations and error codes only, which
+    is what actually identifies the failure.
+    """
+    # This runs *inside* an `except` handler, so it must not raise: an exception
+    # escaping here would turn a gracefully-handled verifier failure into a dead
+    # turn. Everything below is therefore wrapped, including the attribute reads
+    # (a custom exception can expose `status_code`/`errors` as a property that
+    # raises).
+    try:
+        name = type(exc).__name__
+    except Exception:  # pragma: no cover — defensive
+        return "unavailable"
+    try:
+        status = getattr(exc, "status_code", None)
+        if status is not None:
+            return f"{name}(status={status})"
+        errors = getattr(exc, "errors", None)
+        if callable(errors):
+            try:
+                details = errors(
+                    include_input=False, include_url=False, include_context=False
+                )
+            except TypeError:  # pydantic v1 has no include_* kwargs
+                details = errors()
+            # Only `loc` and `type` are ever read — `msg`/`input`/`ctx` can
+            # quote the rejected value.
+            fields = ",".join(
+                f"{'.'.join(str(p) for p in e.get('loc') or ())}:{e.get('type', '?')}"
+                for e in details
+            )
+            if fields:
+                return f"{name}({fields})"
+    except Exception:
+        # Anything odd about this exception object — a property that raises, an
+        # `errors()` that misbehaves — degrades to the type name, never a crash
+        # and never the message.
+        pass
+    return name
+
+
 def _render_tool_result_content(content, cap: int) -> str:
     """Render a tool_result's content as bounded plain text.
 
@@ -2656,25 +2732,64 @@ async def _stream_and_handle_tools(
                 "You are a task-completion verifier. Decide whether the user's "
                 "request is complete, the assistant is waiting on the user, the work "
                 "is unfinished, or the assistant is blocked. Follow the status "
-                "definitions exactly."
+                "definitions exactly.\n\n"
+                # Models that narrate before acting spend the whole budget on
+                # prose and never reach the tool call (ENG-1081). Asking for the
+                # call first shortens the preamble; it does not eliminate it,
+                # which is why the budget below is generous as well.
+                + _VERIFIER_NO_PREAMBLE
             )
-            try:
-                verdict = await self._llm.generate_object_code(
-                    _VerifierVerdict,
-                    system=verifier_system,
-                    messages=verify_messages,
-                    max_tokens=256,
-                )
+            verdict = None
+            for attempt, budget in enumerate(_VERIFIER_TOKEN_BUDGETS):
+                try:
+                    verdict = await self._llm.generate_object_code(
+                        _VerifierVerdict,
+                        system=verifier_system,
+                        messages=verify_messages,
+                        max_tokens=budget,
+                    )
+                    break
+                except StructuredOutputError as exc:
+                    # A truncated verdict is a budget problem, not a verdict: the
+                    # model narrated past `budget` before it reached the tool call
+                    # (ENG-1081). Retry with more room. Any other structured-output
+                    # failure won't be fixed by a bigger budget, so don't pay for it.
+                    retrying = exc.truncated and attempt + 1 < len(
+                        _VERIFIER_TOKEN_BUDGETS
+                    )
+                    _verifier_log.info(
+                        "completion-verifier verdict=%s budget=%d output_tokens=%d "
+                        "stop_reason=%s retrying=%s",
+                        "TRUNCATED" if exc.truncated else "NO_TOOL_CALL",
+                        budget, exc.output_tokens, exc.stop_reason, retrying,
+                    )
+                    if not retrying:
+                        break
+                except Exception as exc:
+                    # Enough to tell the failure modes apart — four used to
+                    # collapse into one "verifier unavailable" line — but never
+                    # the exception message, which can carry conversation
+                    # content (ENG-1081).
+                    _verifier_log.info(
+                        "completion-verifier verdict=ERROR budget=%d error=%s",
+                        budget, _safe_error_detail(exc),
+                    )
+                    break
+
+            if verdict is not None:
                 status = verdict.status
                 reason = verdict.reason.strip()
-            except Exception:
+            else:
                 # Verifier failed — fail safe by treating the turn as done rather
                 # than forcing a continuation the user never asked for.
                 status, reason = "COMPLETE", "verifier unavailable"
 
             _verifier_log.info(
-                "completion-verifier verdict=%s continuation=%d/%d tool_rounds=%d reason=%s",
-                status, continuation, self._max_continuations, tool_round, reason,
+                # No `reason` — it's the model's free-text justification, derived
+                # from the user's conversation, and this is an ordinary app log.
+                # It stays in the Langfuse trace, where that content belongs.
+                "completion-verifier verdict=%s continuation=%d/%d tool_rounds=%d",
+                status, continuation, self._max_continuations, tool_round,
             )
 
             if status in ("COMPLETE", "WAITING"):