dialectical-framework 2.0.11__tar.gz → 2.1.0__tar.gz

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: dialectical-framework
-Version: 2.0.11
+Version: 2.1.0
 Summary: A dialectical framework for augmented intelligence. AI reasoning powered with dialectics supports humans in: system optimization (psychology, engineering, business, politics, etc.); dispute resolution (mediation, conflicts, negotiations, etc.); decision-making (dilemmas, challenging situations, win-win, etc.).
 License: MIT
 Keywords: dialectics,dialectical-reasoning,synthesis,thesis-antithesis,ai,artificial-intelligence,llm,reasoning-framework,philosophy,logic,argumentation,conflict-resolution,decision-making,critical-thinking,semantic-graph,mirascope,pydantic,perspectives,polarity-reasoning

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "dialectical-framework"
-version = "2.0.11"
+version = "2.1.0"
 description = "A dialectical framework for augmented intelligence. AI reasoning powered with dialectics supports humans in: system optimization (psychology, engineering, business, politics, etc.); dispute resolution (mediation, conflicts, negotiations, etc.); decision-making (dilemmas, challenging situations, win-win, etc.)."
 authors = ["Evaldas Taroza <evaldas@dialexity.com>"]
 readme = "README.md"

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/agents/consultant/consultant.py

@@ -381,7 +381,10 @@ class Consultant:
         return self._conversation._messages
 
     async def exploration_view(
-        self, focus: Optional[str] = None, attempts: Optional[int] = None
+        self,
+        focus: Optional[str] = None,
+        attempts: Optional[int] = None,
+        thesis: Optional[str] = None,
     ) -> ExplorationView:
         """The structure of this conversation as a view — drawn by this head.
 
@@ -402,6 +405,17 @@ class Consultant:
         think and building a tetrad in one shot is the heaviest reasoning a
         structured call is asked to do here.
 
+        `thesis` pins T to the host's own text, VERBATIM: the request says so
+        (`view_sketch.THESIS_IS_FIXED`) and the given words are then written
+        back over whatever each draw wrote, so the guarantee is code and not
+        compliance. For a host that shows the person their own sentence as
+        their position, a seven-word recomposition of it is a different
+        sentence. The rest of the tetrad is still built — and still built
+        around the pinned position, which is why the request carries the text
+        too rather than relying on the overwrite alone. Not the same contract
+        as `SketchTetrad(thesis=)`, which asks for the position condensed to
+        `component_length` because the graph stores a headline.
+
         Terminology-free by construction (no hash, alias or score exists),
         unchecked by construction (no HS gate, no validation, nothing
         persisted). Not a tool: the model cannot call this, and the prompt's
@@ -433,7 +447,9 @@ class Consultant:
         # utterances; the first app had measured the same 4.6 calls and +17 s).
         # The request->prose pair `history_ask` keeps OUT of the history is the
         # same trap; the long request as the last turn is what holds the format.
-        request = view_sketch_prompt(focus, self._conversation.settings.component_length)
+        request = view_sketch_prompt(
+            focus, self._conversation.settings.component_length, thesis=thesis
+        )
 
         async def draw() -> ViewSketchDto:
             # Each draw works on a COPY of the history: the turns run in
@@ -472,6 +488,15 @@ class Consultant:
             )
         if not sketches:
             raise failures[0]
+        # The pin, enforced: every draw's first tension carries the host's own
+        # words as T, so the winner and any runner-up a host serves later all
+        # show the sentence the person typed. Done before selection, so the
+        # coherence check judges the tetrad the host will actually get.
+        pinned = (thesis or "").strip()
+        if pinned:
+            for drawn_sketch in sketches:
+                if drawn_sketch.tensions:
+                    drawn_sketch.tensions[0].thesis = pinned
         # With several draws the framework's own coherence check picks the one
         # whose FIRST tension holds best (`concerns/tetrad_candidates.py`); a
         # draw with no complete first tension sorts last. The card a blindspot

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/concerns/aspect_generation.py

@@ -165,10 +165,35 @@ Generate aspect statements that fit the semantic structure."""
 # — small, and not re-measured at the archive's sample size. Do not re-add a
 # worked plus example without measuring BOTH coherence and restatement
 # (`tests/e2e/probe_aspect_variants.py`).
+#
+# Re-aimed on 2026-10-12, by measurement, from "supplies what the other pole is
+# for" to "leaves it possible". The old wording was the take-up half stated a
+# third time, and with a DECIDED thesis ("I'm not inviting my sister — done")
+# "A developed so that it supplies what T is for" IS the middle path: A+ came
+# back as the compromise ("Leave a path back that she must earn"), and the
+# transformation's Ac+, which turns T- into A+, then faithfully handed the
+# person back the thing they had refused (reported by the first app, promise
+# eval v12, 3 of 7 decided cases failing in every rep). On 16 decided stances
+# x 2 reps x 3 arms, Sonnet 5.5 (`tests/e2e/probe_decided_stance.py`,
+# `decided_stance-20261012-090438.json`): A+ a good A ITSELF is for, 39% ->
+# 94% (paired +55, 95% CI +33..+77), compromise pluses gone (plus_relation
+# `distinct` 69% -> 100%), and coherence ROSE, 59% -> 91% (+31, +9..+54) —
+# which is the theory's own reading, since a plus that carries the other pole
+# cannot satisfy "T+ without A+ yields T-". The arm that only reversed the
+# relation reached 75% / +12 CC; what took it to 94% was adding "a position,
+# not a move" (`view_sketch.TETRAD_BUILD_PROCEDURE` step 1), so the two
+# sentences ship together. Not the same lever as the two spent ones: pushing
+# take-up HARDER (the `*_takes_up` fields, 2026-10-02) cost coherence 25/40 ->
+# 15/40, and a hedged-parent CHECK (2026-09-30) moved nothing.
+# `ASPECT_DEFINITIONS` still states the relation the old way and was in both
+# arms' context — unchanged on purpose; editing it is a new measurement.
 PLUS_RESTATEMENT_CHECK = (
-    "Restated parent: if a plus only names what its own parent already "
-    "delivers, rewrite it so the parent stays the generative act while its "
-    "result also supplies what the other pole is for."
+    "Blended parent: a plus is its own parent developed, never the two sides "
+    "combined into one arrangement — not the parent made conditional on the "
+    "other pole, not a partial, scheduled or trial version of it, not the "
+    "middle between them. If a plus only names what its own parent already "
+    "delivers, rewrite it so the parent stays the generative act and what the "
+    "other pole is for is left possible by it."
 )
 
 def is_axis_name(axis: Optional[str]) -> bool:

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/concerns/transformation_sketch.py

@@ -116,9 +116,40 @@ from dialectical_framework.protocols.has_config import SettingsAware
 #: that writes the way through a tension in its own voice (the first app's
 #: card) — so none of them can restate the theory differently.
 TRANSFORMATION_POSITIONS = """1. The action (Ac): the concrete move that carries T toward A. The reflection (Re): the re-reading that carries A back toward T. These are the two poles of the transformation.
-2. Each developed well — Ac+ turns T's trap (T-) into A's strength (A+); Re+ turns A's trap (A-) into T's strength (T+).
+2. Each developed well — Ac+ turns T's trap (T-) into A's strength (A+); Re+ turns A's trap (A-) into T's strength (T+). Where T is something the person has DECIDED or refused, Ac and Ac+ are that decision carried out: what A is for shapes HOW they carry it out, never whether or how much. This governs the action and the action developed well only — Ac- still ends in A's trap, as point 3 says.
 3. Each overdeveloped one-sidedly — Ac- is the good action without the good reflection (Ac+ without Re+), which carries T's strength (T+) into A's trap (A-); Re- is the good reflection without the good action (Re+ without Ac+), which carries A's strength (A+) into T's trap (T-). They are the ACTION and the REFLECTION gone wrong: they END in the traps but are not the traps restated — a path is not its destination. Ac+ contradicts Re-, and Re+ contradicts Ac-."""
 
+#: THE DECIDED POSITION (2026-10-12). Point 2's second sentence, and the only
+#: part of this text that knows anything about the person's state. The first app
+#: reported the framework handing people back exactly what they had refused:
+#: given "Im not inviting my sister to the wedding … done.", Ac+ came back as
+#: "Give her a clear route back: apologise to Marco and she's welcome." Repairing
+#: A+ upstream (the compromise plus, `antithesis-selection.md`, same day) fixed
+#: the CORNER and not the move — `decision_stands` 62% before and after — so this
+#: layer has its own site, and it is not point 1's "carries T toward A":
+#: rewording that moved nothing (-3 paired, `tests/e2e/probe_decided_move.py`).
+#: What moved it is naming the state: 53% -> 88% of actions keeping the decision
+#: (+34, +18..+51, 32 frozen decided tetrads).
+#:
+#: **The scope clause is load-bearing, measured twice.** Whichever minus the rule
+#: is allowed to reach is the one that breaks. Stated unscoped here it kept 32 of
+#: 32 decisions and broke Ac- (-25, resolved): told the action must not dilute
+#: toward A, the writer wrote Ac- as the decision carried out COLDLY, which lands
+#: in T-, not A-. Moved onto the `action` / `ac_plus` field descriptions it scored
+#: the best endpoint of the three (+47) and broke Re- instead (-32, resolved).
+#: Naming what it does not govern is what left both landings alone.
+#:
+#: Price, on this text's own gate (`probe_transformation_sketch.py`, 25 tetrads
+#: x 2 reps, arm `no_decided_rule`): landings intact (Ac- +2, Re- -4, noise),
+#: validity composites within noise (-8, -10), S- third failure +8, and
+#: **`ac_own` 100% -> 90% (-10, -18..-2, resolved)** — the second accepted trade
+#: on that read, after the scaffold fields' 100% -> 93%. Named rather than
+#: swallowed: a third one needs an argument.
+#:
+#: What the sentence deliberately does NOT say is that a diluted action IS Ac-.
+#: That is a theory claim nothing here has checked, and the archive's record is
+#: that a gloss hardened into a rule gets argued with rather than obeyed.
+#:
 #: What the developments produce together. The S- sentences are the measured
 #: fix (`one-shot-transformation.md`: third-trap 4/17 → 12/17 on graph-built
 #: wheels once the synthesis prompt carried them) and are kept word for word,

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/concerns/view_sketch.py

@@ -111,9 +111,11 @@ class ViewSketchPerspectiveDto(BaseModel):
     )
     t_plus: str = Field(
         description=(
-            "T+ — the THESIS developed constructively, so that it also strengthens "
-            "what A offers. Derived from T; the positive end of the axis above. "
-            "Empty when not drawn."
+            "T+ — the THESIS developed constructively, so that what A is for stays "
+            "possible alongside it. T's own side developed — never T blended with A, "
+            "made conditional on it, or the middle between the two. Derived from T; "
+            "the positive end of the axis above. A position held, not a move to "
+            "make. Empty when not drawn."
         )
     )
     a_minus: str = Field(
@@ -131,9 +133,11 @@ class ViewSketchPerspectiveDto(BaseModel):
     )
     a_plus: str = Field(
         description=(
-            "A+ — the ANTITHESIS developed constructively, so that it also strengthens "
-            "what T offers. Derived from A; the positive end of the axis above. "
-            "Empty when not drawn."
+            "A+ — the ANTITHESIS developed constructively, so that what T is for "
+            "stays possible alongside it. A's own side developed — never A blended "
+            "with T, made conditional on it, or the middle between the two. Derived "
+            "from A; the positive end of the axis above. A position held, not a move "
+            "to make. Empty when not drawn."
         )
     )
     t_minus: str = Field(
@@ -162,6 +166,16 @@ class ViewSketchDto(BaseModel):
 #: the graph path gives `AspectGeneration._tetrad_prompt`, stated once here and
 #: shared by the view turn and `concerns/tetrad_sketch.py`, so the two
 #: one-shot builders cannot drift apart.
+#:
+#: Step 1's two measured sentences (2026-10-12, `probe_decided_stance.py`; the
+#: story is on `PLUS_RESTATEMENT_CHECK`, which carries the same rule as the
+#: step-3 check): a plus LEAVES the other pole's good possible rather than
+#: delivering it, and an aspect is a POSITION rather than a move. The second
+#: is what took A+ from 75% to 94% own-strength and coherence from 72% to 91%
+#: on decided stances — the failures it removes are imperatives ("Leave a path
+#: back that she must earn"), and a move that answers both sides at once is a
+#: compromise by construction. It is also where the layering is stated to the
+#: writer: positions live in the tetrad, moves in the Transformation.
 #: What the thesis IS, for a writer that finds it in free words. Measured
 #: without it (`tests/e2e/probe_question_entry.py`, 2026-10-04, 20 questions
 #: through the view turn): the tetrad was statement-grade (CC 17/20, a genuine
@@ -181,12 +195,29 @@ THESIS_IS_A_STANCE = """The thesis (T) is a STANCE: something the person holds,
 TETRAD_BUILD_PROCEDURE = f"""Building a tetrad, when you build one — as its two diagonal contradiction pairs (T+ vs A-, A+ vs T-).
 Each aspect has one fixed parent: T+ and T- develop T; A+ and A- develop A.
 For each pair:
-1. Derive each aspect from ITS OWN parent — a plus develops that parent so it also takes up what the other pole offers; a minus overdevelops that parent one-sidedly, with the other pole absent.
+1. Derive each aspect from ITS OWN parent — a plus develops that parent so that what the other pole is for stays POSSIBLE alongside it, not so that it delivers it; a minus overdevelops that parent one-sidedly, with the other pole absent. Each aspect is a POSITION held or a good someone values, never a move to make: the moves come later, from the way through the tension.
 2. Then name the **axis** — the single dimension on which the two aspects are opposite ends, so they cannot both hold at once.
 3. Re-read both aspects against step 1 for two distinct failures. (a) Wrong parent: if one is really the OTHER parent developed, rewrite it from its own parent. (b) {PLUS_RESTATEMENT_CHECK}"""
 
 
-def view_sketch_prompt(focus: Optional[str], max_words: int) -> str:
+#: What the thesis is when the HOST already has it and wants it kept word for
+#: word. Distinct from `tetrad_sketch_prompt`'s given-thesis line, which asks
+#: for the position "worded in the person's own terms, {max_words} words or
+#: fewer" — a condensation, which is right for `anchor` (the model named the
+#: position, the graph stores a headline) and wrong for a host that shows the
+#: person their OWN sentence as their position (the first app, 2026-10-12: the
+#: card displays the typed words while T was a seven-word recomposition of
+#: them). So this one forbids the rewording the other one invites, and the
+#: framework does not rely on it being obeyed: `Consultant.exploration_view`
+#: writes the given text back over the drawn one (the pin is in code; this
+#: text only stops the writer building the rest of the tetrad around a position
+#: it has quietly reworded).
+THESIS_IS_FIXED = """The thesis (T) is GIVEN, in the person's own words, and it is FIXED: carry it into the view exactly as it stands — do not reword, shorten, sharpen or generalise it, and do not put a different position in its place. Everything else in the tetrad is built around it as given."""
+
+
+def view_sketch_prompt(
+    focus: Optional[str], max_words: int, thesis: Optional[str] = None
+) -> str:
     """The structured turn's request, over the consultant's own conversation.
 
     Two instructions in one, because the ask is one or the other and the model
@@ -203,12 +234,42 @@ def view_sketch_prompt(focus: Optional[str], max_words: int) -> str:
     (docs/dev-notes/antithesis-selection.md, 2026-10-02). The thesis is asked
     for as a STANCE (`THESIS_IS_A_STANCE`): on 20 questions the turn echoed the
     question as T in 6 before the ask was here (the alsotrue app repo, docs/lab-notes-entry.md).
+
+    With `thesis` the host has the position already and wants it kept verbatim
+    (`THESIS_IS_FIXED`): the stance ask comes OFF — there is nothing to find —
+    and the view is asked for the ONE tension that position stands in, because
+    a pinned T says nothing about any other tension and a card shows one.
     """
+    pinned = (thesis or "").strip()
     focus_line = (
         f"What to show: {focus}\n\n"
         if focus and focus.strip()
         else "What to show: the tensions this conversation has worked out so far.\n\n"
     )
+    scope_line = (
+        "- ONE tension: the one the given position below stands in. How much of "
+        "it is drawn is still the ask's to say, by the rule above."
+        if pinned
+        else f"- At most {VIEW_SKETCH_MAX_PERSPECTIVES} tensions, most central "
+        "first. Nothing at all rather than a tension the conversation does not "
+        "support — a turn that carries no stance (a request for information) "
+        "supports none."
+    )
+    thesis_section = (
+        f'{THESIS_IS_FIXED}\n\nThe position: "{pinned}"'
+        if pinned
+        else THESIS_IS_A_STANCE
+    )
+    # The exception rides on the line that states the limit, not as a
+    # cross-reference from the section above it: a rule and its exception in
+    # two places is the competing-signal shape, and the limit is the last line
+    # the model reads.
+    length_exception = (
+        " The given position is the one exception: it stays exactly as it is, "
+        "whatever its length."
+        if pinned
+        else ""
+    )
     return f"""Draw the structure of this conversation as dialectical tetrads. Answer with ONE JSON object in the requested schema and nothing else — no prose. Earlier drawings in this conversation are recorded in prose; that is their record, not the format of this answer.
 
 {focus_line}{ASPECT_DEFINITIONS}
@@ -217,15 +278,15 @@ Rules for what goes in the view:
 - Show what this conversation has ESTABLISHED: a tension you named and the person took up is drawn as you both have it, in their terms. Do not re-derive what is already on the table.
 - Where the ask needs structure not yet worked out — a perspective for a thesis so far only stated, the antithesis of a position, the corners of a tension you only named — build it now, by the method below, and then draw it.
 - Where the ask stops short of a full tetrad (only the two poles, only the antitheses), leave the positions not asked for EMPTY. An empty corner is honest; an invented one is not.
-- At most {VIEW_SKETCH_MAX_PERSPECTIVES} tensions, most central first. Nothing at all rather than a tension the conversation does not support — a turn that carries no stance (a request for information) supports none.
+{scope_line}
 
-{THESIS_IS_A_STANCE}
+{thesis_section}
 
 {_OPPOSING_POSITION_ASK}
 
 {TETRAD_BUILD_PROCEDURE}
 
-Write every position in the person's own terms, {max_words} words or fewer each."""
+Write every position in the person's own terms, {max_words} words or fewer each.{length_exception}"""
 
 
 # --- The view shaping (pure) ----------------------------------------------

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/settings.py

@@ -288,8 +288,10 @@ class Settings(BaseModel):
     # so this is a ceiling on one generation; on a streamed round it is the gap
     # between two chunks, and a thinking turn streams its deltas (the
     # facilitator asks for `include_thoughts`), so the long silence is the
-    # non-streamed case only. 120s clears every unstalled generation in the
-    # archive (the longest measured conversational rounds sit under a minute on
+    # non-streamed case only. Since `llm_stream_awaited_calls` (below) every
+    # awaited call is streamed too, so this bounds the wait for the first
+    # content; a hang is caught earlier, when the response fails to open.
+    # 120s clears every unstalled generation in the archive (the longest measured conversational rounds sit under a minute on
     # Sonnet 5 with thinking) while cutting a stall from 600s to two minutes. A
     # host whose calls are short sets it lower: the first app's draws run ~15s,
     # texts ~8s, structured calls ~3s on Sonnet 5.5, and it runs at 30s. Too
@@ -299,6 +301,21 @@ class Settings(BaseModel):
     # timeout, when the provider registers.
     llm_read_timeout_s: float = Field(default=120.0, description="Read timeout for LLM calls, in seconds: the longest one call may wait for the provider's next byte. Lower where every call is short; a stall then surfaces as a timeout instead of holding the caller.")
 
+    # Every awaited Bedrock call (`BedrockAnthropicProvider._create_async`) is
+    # sent as a stream and assembled into the same `Message` before Mirascope
+    # decodes it, so a stall is caught where it happens instead of at the whole
+    # read timeout: the response must OPEN within 5s (`_STREAM_HEADERS_TIMEOUT_S`),
+    # content must start within `llm_read_timeout_s`, and a gap past 5s once it
+    # flows is a stall. Measured on the first app's promise run (2026-10-12,
+    # Sonnet 5.5, 59 cases, read timeout 20s): ~5.5% of calls hang, ALL before
+    # the response opens, while healthy calls open in a median 1.2s whatever
+    # their size; against a same-day baseline, cards drawn 43 -> 51, cards past
+    # the app's deadline 9 -> 0, p90 53s -> 33s, cost unchanged. The switch
+    # exists as an off switch for a model or provider the measurement did not
+    # cover (`docs/dev-notes/streaming-latency-and-retry.md`, "Hangs are caught
+    # at the opening"). Per request.
+    llm_stream_awaited_calls: bool = Field(default=True, description="Send awaited provider calls as streams and assemble the message, so a hung call is caught when the response fails to open (5s) rather than at the whole read timeout. On by default; off only for a model or provider where streaming misbehaves.")
+
     @classmethod
     def from_partial(cls, partial_settings: Optional[Settings] = None) -> Self:
         """
@@ -377,5 +394,6 @@ class Settings(BaseModel):
             **_env_thinking_level(),
             llm_connect_timeout_s=float(os.getenv("DIALEXITY_LLM_CONNECT_TIMEOUT_S", 30.0)),
             llm_read_timeout_s=float(os.getenv("DIALEXITY_LLM_READ_TIMEOUT_S", 120.0)),
+            llm_stream_awaited_calls=os.getenv("DIALEXITY_LLM_STREAM_AWAITED_CALLS", "true").lower() == "true",
             effect_log_dir=os.getenv("DIALEXITY_GRAPH_LOG_DIR"),
         )

{dialectical_framework-2.0.11 → dialectical_framework-2.1.0}/src/dialectical_framework/utils/bedrock_provider.py

@@ -1,5 +1,8 @@
 from __future__ import annotations
 
+import asyncio
+import logging
+import time
 from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any, cast
 
@@ -18,6 +21,7 @@ from typing_extensions import Unpack
 from dialectical_framework.enums.di import DI
 from dialectical_framework.exceptions.provider_errors import ModelRefusal
 from dialectical_framework.settings import Settings
+from dialectical_framework.settings_context import get_current_settings
 from dialectical_framework.utils.format_compat import (
     learn_format_mode_from_error,
     structured_output_format,
@@ -375,6 +379,135 @@ def raise_on_refusal(response: Any, model_name: str) -> None:
     )
 
 
+logger = logging.getLogger(__name__)
+
+#: Once a streamed awaited call has started writing content, the longest it may
+#: go without another event before it counts as a stall. Generation is a steady
+#: flow of deltas, so a silence this long mid-content is a hang rather than a
+#: slow token. Before the first content delta the wait is the read timeout:
+#: that window holds queueing, prefill and (where a call thinks) the thinking
+#: block. Fired once in 2,700 streamed calls on the first app's promise runs,
+#: so it is a backstop; the measured hangs are all caught at the opening.
+_STREAM_GAP_TIMEOUT_S = 5.0
+
+#: How long a streamed awaited call may wait for the response to OPEN (headers,
+#: `message_start`) before it counts as a stall. Measured on the first app's
+#: promise run (2026-10-12, Sonnet 5.5, 890 streamed calls): every stall (48,
+#: 5.4%) hung before the headers, while healthy calls opened at a median 1.2s,
+#: p99 3.6s, whatever their size (6 of 842 past 5s). Streamed, the opening is
+#: the one wait that does not grow with the output, so it is where a hang is
+#: caught: here at 5s rather than at the whole read timeout.
+_STREAM_HEADERS_TIMEOUT_S = 5.0
+
+#: Deltas that are the answer being written, as opposed to thinking or
+#: bookkeeping. The SDK's stream yields both the raw `content_block_delta` and
+#: a synthesized event per delta kind.
+_CONTENT_DELTAS = frozenset({"text_delta", "input_json_delta"})
+_CONTENT_EVENTS = frozenset({"text", "input_json"})
+
+
+class StreamReadTimeout(httpx.ReadTimeout):
+    """A streamed awaited call went quiet past its budget.
+
+    Subclasses `httpx.ReadTimeout` so `use_brain._is_stall_error` reads it as a
+    stall (the name ends in `ReadTimeout`) and re-asks it once, exactly like a
+    socket read timeout on the non-streamed path.
+    """
+
+
+def _is_content_event(event: Any) -> bool:
+    kind = getattr(event, "type", None)
+    if kind in _CONTENT_EVENTS:
+        return True
+    if kind == "content_block_delta":
+        return getattr(getattr(event, "delta", None), "type", None) in _CONTENT_DELTAS
+    return False
+
+
+async def _stream_to_message(
+    manager: _LearningStreamManager, model_name: str, read_timeout_s: float
+) -> AnthropicMessage:
+    """Consume one stream and return the assembled `Message`, policing silence.
+
+    Three budgets: the response must OPEN within `_STREAM_HEADERS_TIMEOUT_S`
+    (where the measured hangs are), the first content must follow within the
+    read timeout (queueing, prefill, thinking), and once content flows a gap
+    longer than `_STREAM_GAP_TIMEOUT_S` is a stall.
+
+    The provider's `ping`s do not count as progress: a stream that only pings
+    is a hang the socket's read timeout would never see. Every stall is logged
+    with the phase it happened in: the phase is what chose the opening as the
+    place to police (all 137 measured hangs were there), and a shift in it is
+    the first sign that choice needs revisiting.
+    """
+    loop = asyncio.get_running_loop()
+    started = time.monotonic()
+    headers_budget = min(_STREAM_HEADERS_TIMEOUT_S, read_timeout_s)
+    try:
+        stream = await asyncio.wait_for(manager.__aenter__(), timeout=headers_budget)
+    except asyncio.TimeoutError as e:
+        logger.warning(
+            "Stream stalled on %s before headers (%.1fs)",
+            model_name, time.monotonic() - started,
+        )
+        raise StreamReadTimeout(
+            f"stream did not open within {headers_budget:.0f}s"
+        ) from e
+    except Exception as e:
+        if type(e).__name__.endswith(("TimeoutError", "ReadTimeout")):
+            logger.warning(
+                "Stream stalled on %s before headers (%.1fs)",
+                model_name, time.monotonic() - started,
+            )
+        raise
+    events = 0
+    first_content_s: float | None = None
+    try:
+        iterator = stream.__aiter__()
+        deadline = loop.time() + read_timeout_s
+        while True:
+            try:
+                event = await asyncio.wait_for(
+                    iterator.__anext__(), timeout=max(deadline - loop.time(), 0.0)
+                )
+            except StopAsyncIteration:
+                break
+            except Exception as e:
+                timed_out_here = isinstance(e, asyncio.TimeoutError)
+                if not timed_out_here and not type(e).__name__.endswith(
+                    ("Timeout", "TimeoutError")
+                ):
+                    raise
+                elapsed = time.monotonic() - started
+                phase = (
+                    "before first content"
+                    if first_content_s is None
+                    else f"mid-content (first content at {first_content_s:.1f}s)"
+                )
+                logger.warning(
+                    "Stream stalled on %s %s: %.1fs, %d events",
+                    model_name, phase, elapsed, events,
+                )
+                if not timed_out_here:
+                    raise
+                raise StreamReadTimeout(
+                    f"stream went quiet {phase} after {elapsed:.1f}s and {events} events"
+                ) from e
+            if getattr(event, "type", None) == "ping":
+                continue
+            events += 1
+            if first_content_s is None and _is_content_event(event):
+                first_content_s = time.monotonic() - started
+            budget = read_timeout_s if first_content_s is None else _STREAM_GAP_TIMEOUT_S
+            deadline = loop.time() + budget
+        message = await stream.get_final_message()
+    except BaseException as e:
+        await manager.__aexit__(type(e), e, e.__traceback__)
+        raise
+    await manager.__aexit__(None, None, None)
+    return message
+
+
 class BedrockAnthropicProvider(AnthropicProvider):
     """Mirascope v2 provider that routes through AnthropicBedrock client (async-native).
 
@@ -394,6 +527,7 @@ class BedrockAnthropicProvider(AnthropicProvider):
     ) -> None:
         # Skip super().__init__() — parent creates Anthropic/AsyncAnthropic clients we don't need
         timeout = _client_timeout(connect_timeout_s, read_timeout_s)
+        self._read_timeout_s = timeout.read or 120.0
         self.client = AnthropicBedrock(timeout=timeout, max_retries=_SDK_MAX_RETRIES)
         self.async_client = AsyncAnthropicBedrock(
             timeout=timeout, max_retries=_SDK_MAX_RETRIES
@@ -468,8 +602,20 @@ class BedrockAnthropicProvider(AnthropicProvider):
         models reject. `with_thinking_compat` translates by model name; if the
         name heuristic is wrong the 400 itself says which shape is wanted, so
         one retry is both sufficient and self-correcting. See thinking_compat.
+
+        With `Settings.llm_stream_awaited_calls` the same request goes out as a
+        stream and comes back as the same assembled `Message`, so a stall is
+        caught by the gap between chunks (`_stream_to_message`) instead of by a
+        ceiling on the whole generation. The stream manager learns the thinking
+        shape the same way.
         """
         model_name = kwargs["model"]
+        if get_current_settings().llm_stream_awaited_calls:
+            return await _stream_to_message(
+                _LearningStreamManager(self.async_client, kwargs, params),
+                model_name,
+                self._read_timeout_s,
+            )
         try:
             return await self.async_client.messages.create(
                 **with_thinking_compat(model_name, kwargs, params)