Merge pull request #52 from pewdiepie-archdaemon/feature/runtime-context-resolution

feat(runtime): resolve compact context once at turn preparation

src/agent_loop.py

@@ -84,6 +84,7 @@ from src.tool_types import ToolBlock
 from src.turn_contract import selected_tools_for_request, with_turn_contract
 from src.agent_runtime.journal import propose_action, execute_action
 from src.agent_runtime.completion import with_completion_gate
+from src.agent_runtime.runtime_selection import is_compact_preview_contract
 from src.teacher_escalation import with_teacher_takeover, request_teacher_takeover
 from src.tool_utils import _truncate, get_mcp_manager
 from src.agent_tools import (
@@ -20425,6 +20426,7 @@ async def stream_agent_loop(
     reasoning_effort: Optional[str] = None,
     _parent_run_id: Optional[str] = None,
     request_authority=MISSING_AUTHORITY,
+    context_resolution=None,
 ) -> AsyncGenerator[str, None]:
     """Streaming agent loop generator.
 
@@ -20460,7 +20462,7 @@ async def stream_agent_loop(
                     ),
                 )
 
-    if turn_contract is not None and turn_contract.selection_mode == 'clean_compact_v3_preview':
+    if is_compact_preview_contract(turn_contract):
         from src.clean_agent_preview import stream_preview
         async for chunk in stream_preview(
             endpoint_url=endpoint_url, model=model, messages=messages, headers=headers,
@@ -20477,6 +20479,9 @@ async def stream_agent_loop(
             max_rounds=max_rounds,
             max_tool_calls=max_tool_calls,
             temperature=temperature,
+            # The route's prepared, typed context window; the compact runtime
+            # must not resolve it a second time.
+            context_resolution=context_resolution,
         ):
             yield chunk
         return

src/agent_runtime/context_resolution.py

@@ -0,0 +1,519 @@
+"""Effective model context window, resolved once per logical turn.
+
+A turn resolves the window it budgets against at preparation time, before any
+model request, and keeps the value together with the evidence that chose it.
+Terminal metrics report that stored resolution; they never start discovery.
+
+Evidence classes are kept apart instead of being folded into one "known" flag:
+
+* ``runtime_confirmed``: the serving process reported its active window
+  (llama.cpp ``/slots`` or ``/props``) or rejected a request of this turn with
+  an explicit limit.
+* ``provider_advertised``: the provider's model catalog lists a window.
+* ``operator_declared``: the client or operator declared a transport window.
+  It caps runtime or provider evidence and replaces weaker evidence.
+* ``known_table``: the static ``KNOWN_CONTEXT_WINDOWS`` fallback.
+* ``unknown``: nothing above is available. The value is 0, never a default.
+
+Any disagreement between sources is recorded as a conflict. An operator value
+below a measured value is a cap, not a contradiction; an operator value above
+it is a contradiction.
+
+Context sizing is not authority: nothing here grants or denies an operation.
+"""
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass, field, replace
+from enum import Enum
+import hashlib
+import json
+import logging
+import time
+from typing import Any, Mapping, Optional
+from urllib.parse import urlparse
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+# Upper bound for all provider metadata I/O of one turn preparation. A slow
+# or unreachable metadata endpoint costs at most this much before the turn
+# proceeds with whatever evidence it has.
+PROBE_DEADLINE_SECONDS = 3.0
+# Remote provider metadata changes rarely; failures are retried sooner so a
+# transient outage does not pin a turn to weaker evidence for long. Local
+# servers are always re-probed because they can restart with another window.
+PROBE_CACHE_TTL_SECONDS = 600.0
+PROBE_FAILURE_TTL_SECONDS = 60.0
+
+# Headers that describe the chat request body rather than the caller.
+_REQUEST_ONLY_HEADERS = frozenset({"content-type", "content-length", "accept", "accept-encoding"})
+
+
+class ContextEvidence(str, Enum):
+    RUNTIME_CONFIRMED = "runtime_confirmed"
+    PROVIDER_ADVERTISED = "provider_advertised"
+    OPERATOR_DECLARED = "operator_declared"
+    KNOWN_TABLE = "known_table"
+    UNKNOWN = "unknown"
+
+
+@dataclass(frozen=True)
+class ContextObservation:
+    evidence: ContextEvidence
+    value: int
+    source: str
+
+    def to_dict(self) -> dict:
+        return {"evidence": self.evidence.value, "value": self.value, "source": self.source}
+
+
+@dataclass(frozen=True)
+class ContextConflict:
+    first: ContextObservation
+    second: ContextObservation
+
+    def to_dict(self) -> dict:
+        return {"first": self.first.to_dict(), "second": self.second.to_dict()}
+
+
+@dataclass(frozen=True)
+class ContextResolution:
+    """The effective window of one turn and why it was chosen."""
+
+    effective: int
+    evidence: ContextEvidence
+    source: str
+    observations: tuple[ContextObservation, ...] = ()
+    conflicts: tuple[ContextConflict, ...] = ()
+    provider_io: bool = False
+    cached: bool = False
+    probe_errors: tuple[str, ...] = ()
+    # The route this resolution describes. Empty for resolutions built
+    # directly from observations by internal callers. The URL can carry
+    # credentials, so it stays out of repr() and to_dict().
+    endpoint_url: str = field(default="", repr=False)
+    model: str = ""
+
+    @property
+    def mismatch(self) -> bool:
+        return bool(self.conflicts)
+
+    @property
+    def budget_limit(self) -> int:
+        """Window the runtime may budget against; 0 means budget reactively."""
+        return self.effective if self.evidence is not ContextEvidence.UNKNOWN else 0
+
+    @property
+    def shaping_window(self) -> int:
+        """Window for the legacy history compaction/trim helpers.
+
+        Those helpers predate typed evidence and always size against some
+        window, using DEFAULT_CONTEXT when none is known. This only feeds them
+        a number; it never creates provenance for that number.
+        """
+        if self.budget_limit:
+            return self.budget_limit
+        from src.model_context import DEFAULT_CONTEXT
+        return DEFAULT_CONTEXT
+
+    def applies_to(self, endpoint_url: str, model: str) -> bool:
+        """Whether this resolution may be reused for the given route."""
+        if not self.endpoint_url and not self.model:
+            return True
+        return self.endpoint_url == endpoint_url and self.model == model
+
+    def observe_runtime_limit(self, limit: Any, source: str = "provider_rejection") -> "ContextResolution":
+        """Fold a limit the provider stated during this turn. Performs no I/O."""
+        try:
+            value = int(limit or 0)
+        except (TypeError, ValueError):
+            return self
+        if value <= 0:
+            return self
+        observation = ContextObservation(ContextEvidence.RUNTIME_CONFIRMED, value, source)
+        if observation in self.observations:
+            return self
+        combined = combine_observations((*self.observations, observation))
+        return replace(
+            combined,
+            provider_io=self.provider_io,
+            cached=self.cached,
+            probe_errors=self.probe_errors,
+            endpoint_url=self.endpoint_url,
+            model=self.model,
+        )
+
+    def to_dict(self) -> dict:
+        return {
+            "effective": self.effective,
+            "evidence": self.evidence.value,
+            "source": self.source,
+            "mismatch": self.mismatch,
+            "conflicts": [conflict.to_dict() for conflict in self.conflicts],
+            "observations": [observation.to_dict() for observation in self.observations],
+            "provider_io": self.provider_io,
+            "cached": self.cached,
+            "probe_errors": list(self.probe_errors),
+        }
+
+
+UNRESOLVED_CONTEXT = ContextResolution(0, ContextEvidence.UNKNOWN, "none")
+
+_MEASURED = (ContextEvidence.RUNTIME_CONFIRMED, ContextEvidence.PROVIDER_ADVERTISED)
+
+
+def _conflicts(observations: tuple[ContextObservation, ...]) -> tuple[ContextConflict, ...]:
+    conflicts = []
+    for index, first in enumerate(observations):
+        for second in observations[index + 1:]:
+            if first.value == second.value:
+                continue
+            classes = {first.evidence, second.evidence}
+            if ContextEvidence.OPERATOR_DECLARED in classes:
+                operator, other = (
+                    (first, second) if first.evidence is ContextEvidence.OPERATOR_DECLARED
+                    else (second, first)
+                )
+                # A declared window replaces the static table and may cap a
+                # measured window. Only a declaration above what the runtime
+                # or provider supports contradicts it.
+                if other.evidence not in _MEASURED or operator.value < other.value:
+                    continue
+            conflicts.append(ContextConflict(first, second))
+    return tuple(conflicts)
+
+
+def _strongest(observations, evidence: ContextEvidence) -> Optional[ContextObservation]:
+    matching = [observation for observation in observations if observation.evidence is evidence]
+    return min(matching, key=lambda observation: observation.value) if matching else None
+
+
+def combine_observations(observations) -> ContextResolution:
+    """Choose the effective window deterministically from observations.
+
+    The smallest runtime-confirmed value wins, else the smallest provider
+    value. An operator declaration caps either, and replaces the known table
+    or an unknown window. The known table is used only when nothing stronger
+    exists. No observation yields an unknown window of 0.
+    """
+    observations = tuple(observations)
+    measured = (
+        _strongest(observations, ContextEvidence.RUNTIME_CONFIRMED)
+        or _strongest(observations, ContextEvidence.PROVIDER_ADVERTISED)
+    )
+    operator = _strongest(observations, ContextEvidence.OPERATOR_DECLARED)
+    if measured and operator:
+        chosen = operator if operator.value < measured.value else measured
+    else:
+        chosen = measured or operator or _strongest(observations, ContextEvidence.KNOWN_TABLE)
+    conflicts = _conflicts(observations)
+    if chosen is None:
+        return replace(UNRESOLVED_CONTEXT, observations=observations, conflicts=conflicts)
+    return ContextResolution(
+        chosen.value, chosen.evidence, chosen.source,
+        observations=observations, conflicts=conflicts,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Provider metadata probe
+# ---------------------------------------------------------------------------
+
+@dataclass(frozen=True)
+class _ProbeResult:
+    observations: tuple[ContextObservation, ...] = ()
+    errors: tuple[str, ...] = ()
+    io: bool = False
+
+
+_probe_cache: dict[tuple[str, str, str], tuple[float, _ProbeResult]] = {}
+
+
+def clear_probe_cache() -> None:
+    _probe_cache.clear()
+
+
+_DEFAULT_PORTS = {"http": 80, "https": 443}
+
+
+def _origin(url: str) -> tuple[str, str, Optional[int]]:
+    parsed = urlparse(url or "")
+    scheme = parsed.scheme.lower()
+    try:
+        port = parsed.port
+    except ValueError:
+        return ("", "", None)
+    return (scheme, (parsed.hostname or "").lower(), port or _DEFAULT_PORTS.get(scheme))
+
+
+def _http_client(timeout: float):
+    # Credentials must never follow a redirect to another location.
+    return httpx.AsyncClient(timeout=timeout, follow_redirects=False)
+
+
+def _provider_urls(endpoint_url: str) -> tuple[Optional[str], str]:
+    """Models catalog URL and the server-resolved form of the endpoint.
+
+    Both come from the existing endpoint resolver, which may rewrite an
+    unresolvable host to its Tailscale address. Blocking (DNS, subprocess);
+    call it off the event loop.
+    """
+    from src.endpoint_resolver import build_models_url, resolve_url
+
+    return build_models_url(endpoint_url), resolve_url(endpoint_url)
+
+
+def _probe_headers(trusted_origins, target_url: str, headers: Optional[Mapping[str, Any]]) -> dict:
+    """Forward the turn's provider credentials only to the provider's origin."""
+    origin = _origin(target_url)
+    if not headers or not origin[1] or origin not in trusted_origins:
+        return {}
+    return {
+        str(name): str(value) for name, value in headers.items()
+        if value is not None and str(name).lower() not in _REQUEST_ONLY_HEADERS
+    }
+
+
+def _auth_fingerprint(headers: Optional[Mapping[str, Any]]) -> str:
+    if not headers:
+        return ""
+    material = json.dumps(
+        sorted((str(k).lower(), str(v)) for k, v in headers.items()
+               if v is not None and str(k).lower() not in _REQUEST_ONLY_HEADERS),
+        separators=(",", ":"),
+    )
+    return hashlib.sha256(material.encode("utf-8")).hexdigest()[:16]
+
+
+def _serving_base(endpoint_url: str) -> str:
+    # Same derivation the regular runtime uses for llama.cpp server routes.
+    return endpoint_url.split("/v1")[0] if "/v1" in endpoint_url else endpoint_url.rsplit("/", 1)[0]
+
+
+async def _get_json(client, url, headers, errors, label):
+    try:
+        response = await client.get(url, headers=headers)
+    except httpx.TimeoutException:
+        errors.append(f"{label}:timeout")
+        return None
+    except httpx.TransportError:
+        errors.append(f"{label}:transport_error")
+        return None
+    status = getattr(response, "status_code", 0)
+    if not (200 <= int(status or 0) < 300):
+        errors.append(f"{label}:http_{status}")
+        return None
+    try:
+        return response.json()
+    except Exception:
+        errors.append(f"{label}:invalid_payload")
+        return None
+
+
+def _positive_int(value) -> int:
+    if isinstance(value, bool) or not isinstance(value, (int, float)) or value <= 0:
+        return 0
+    return int(value)
+
+
+async def _probe(endpoint_url, model, headers, is_local, observations, errors, timeout):
+    from src.copilot import is_copilot_base
+    from src.model_context import _model_ctx_from_entry
+
+    # Credentials go only to the configured provider's origin, or to the
+    # form of that same endpoint the server-owned resolver produced.
+    trusted = {_origin(endpoint_url)}
+    async with _http_client(timeout) as client:
+        if is_local:
+            base = _serving_base(endpoint_url)
+            slots = await _get_json(
+                client, f"{base}/slots", _probe_headers(trusted, f"{base}/slots", headers),
+                errors, "slots",
+            )
+            n_ctx = _positive_int(slots[0].get("n_ctx")) if (
+                isinstance(slots, list) and slots and isinstance(slots[0], dict)
+            ) else 0
+            if not n_ctx:
+                props = await _get_json(
+                    client, f"{base}/props", _probe_headers(trusted, f"{base}/props", headers),
+                    errors, "props",
+                )
+                generation = props.get("default_generation_settings") if isinstance(props, dict) else None
+                n_ctx = _positive_int(generation.get("n_ctx")) if isinstance(generation, dict) else 0
+                source = "llamacpp_props"
+            else:
+                source = "llamacpp_slots"
+            if n_ctx:
+                observations.append(
+                    ContextObservation(ContextEvidence.RUNTIME_CONFIRMED, n_ctx, source)
+                )
+
+        # Copilot's catalog needs headers this layer does not own; an
+        # unauthenticated probe only fails. Its models are table-covered.
+        if is_copilot_base(endpoint_url):
+            errors.append("models:unsupported_endpoint")
+            return
+        # URL building may resolve the host (DNS, tailscale lookup); keep that
+        # off the event loop and inside the probe deadline.
+        models_url, resolved_endpoint = await asyncio.to_thread(_provider_urls, endpoint_url)
+        if not models_url:
+            errors.append("models:unsupported_endpoint")
+            return
+        trusted.add(_origin(resolved_endpoint))
+        payload = await _get_json(
+            client, models_url, _probe_headers(trusted, models_url, headers),
+            errors, "models",
+        )
+        if payload is None:
+            return
+        entries = payload.get("data") if isinstance(payload, dict) else None
+        if not isinstance(entries, list):
+            errors.append("models:invalid_payload")
+            return
+        wanted = model.split("/")[-1]
+        for entry in entries:
+            if not isinstance(entry, dict):
+                continue
+            entry_id = str(entry.get("id") or "")
+            if entry_id == model or entry_id.split("/")[-1] == wanted:
+                value = _model_ctx_from_entry(entry)
+                if value:
+                    observations.append(ContextObservation(
+                        ContextEvidence.PROVIDER_ADVERTISED, int(value), "models_catalog",
+                    ))
+                else:
+                    errors.append("models:no_window_listed")
+                return
+        errors.append("models:model_not_listed")
+
+
+async def probe_provider_context(
+    endpoint_url: str,
+    model: str,
+    *,
+    headers: Optional[Mapping[str, Any]] = None,
+    deadline_seconds: float = PROBE_DEADLINE_SECONDS,
+    is_local: Optional[bool] = None,
+) -> _ProbeResult:
+    """Query provider metadata once, bounded by ``deadline_seconds``.
+
+    Never raises: every failure is reported as a short, secret-free error code
+    so a turn can continue with other evidence.
+    """
+    observations: list[ContextObservation] = []
+    errors: list[str] = []
+    if is_local is None:
+        is_local = await _is_local(endpoint_url)
+    timeout = max(0.1, float(deadline_seconds))
+    try:
+        await asyncio.wait_for(
+            _probe(endpoint_url, model, headers, is_local, observations, errors, timeout),
+            timeout=timeout,
+        )
+    except asyncio.TimeoutError:
+        errors.append("deadline_exceeded")
+    except Exception as exc:
+        logger.debug("Context window probe failed: %s", type(exc).__name__)
+        errors.append("probe_failed")
+    return _ProbeResult(tuple(observations), tuple(errors), io=True)
+
+
+async def _is_local(endpoint_url: str) -> bool:
+    from src.model_context import is_local_endpoint
+
+    try:
+        # Reads configured endpoints from the local database on the calling
+        # thread, as the regular runtime does. Moving it to worker threads
+        # gives SQLite sessions per-thread connections the app never uses.
+        return bool(is_local_endpoint(endpoint_url))
+    except Exception:
+        return False
+
+
+async def _cached_probe(endpoint_url, model, headers, deadline_seconds, clock):
+    is_local = await _is_local(endpoint_url)
+    key = (endpoint_url, model, _auth_fingerprint(headers))
+    if not is_local:
+        cached = _probe_cache.get(key)
+        if cached and cached[0] > clock():
+            return cached[1], True
+    result = await probe_provider_context(
+        endpoint_url, model, headers=headers, deadline_seconds=deadline_seconds,
+        is_local=is_local,
+    )
+    if not is_local:
+        ttl = PROBE_CACHE_TTL_SECONDS if result.observations else PROBE_FAILURE_TTL_SECONDS
+        _probe_cache[key] = (clock() + ttl, result)
+    return result, False
+
+
+def declared_context_window(client_runtime_context: Any) -> int:
+    """Operator/client declared transport window, or 0."""
+    if not isinstance(client_runtime_context, Mapping):
+        return 0
+    try:
+        value = int(client_runtime_context.get("model_context_window") or 0)
+    except (TypeError, ValueError):
+        return 0
+    return value if value > 0 else 0
+
+
+async def resolve_effective_context(
+    endpoint_url: str,
+    model: str,
+    *,
+    headers: Optional[Mapping[str, Any]] = None,
+    client_runtime_context: Any = None,
+    deadline_seconds: float = PROBE_DEADLINE_SECONDS,
+    probe: bool = True,
+    clock=time.monotonic,
+) -> ContextResolution:
+    """Resolve the effective context window for one turn preparation."""
+    from src.model_context import _lookup_known
+
+    observations: list[ContextObservation] = []
+    errors: tuple[str, ...] = ()
+    provider_io = cached = False
+    if probe and endpoint_url and model:
+        result, cached = await _cached_probe(
+            endpoint_url, model, headers, deadline_seconds, clock,
+        )
+        observations.extend(result.observations)
+        errors = result.errors
+        provider_io = result.io and not cached
+    declared = declared_context_window(client_runtime_context)
+    if declared:
+        observations.append(ContextObservation(
+            ContextEvidence.OPERATOR_DECLARED, declared, "client_runtime_context",
+        ))
+    known = _lookup_known(model or "")
+    if known:
+        observations.append(ContextObservation(ContextEvidence.KNOWN_TABLE, int(known), "known_table"))
+    resolution = combine_observations(observations)
+    resolution = replace(
+        resolution, provider_io=provider_io, cached=cached, probe_errors=errors,
+        endpoint_url=endpoint_url or "", model=model or "",
+    )
+    if resolution.mismatch:
+        logger.info(
+            "Context window sources disagree for %s: %s",
+            model, [conflict.to_dict() for conflict in resolution.conflicts],
+        )
+    return resolution
+
+
+def context_metrics(resolution: Optional[ContextResolution], request_tokens: int) -> dict:
+    """Metrics fields derived only from a stored resolution. Performs no I/O."""
+    resolution = resolution or UNRESOLVED_CONTEXT
+    length = resolution.budget_limit
+    percent = (
+        min(round((request_tokens / length) * 100, 1), 100.0)
+        if length and request_tokens else 0
+    )
+    return {
+        "context_length": length,
+        "context_percent": percent,
+        "context_resolution": resolution.to_dict(),
+    }

src/agent_runtime/runtime_selection.py

@@ -0,0 +1,44 @@
+"""Whether a turn runs on the compact (clean v3) preview runtime.
+
+The chat route decides this once, from facts known before context
+preparation, and uses that one value both to prepare the turn (its typed
+context resolution) and to stamp the turn contract's selection mode. The
+agent loop dispatches on that stamp. Keeping both sides here, with no other
+imports, means preparation and dispatch read one rule and cannot drift.
+
+Runtime selection is not authority: it grants or denies no operation.
+"""
+
+COMPACT_PREVIEW_MODE = "clean_compact_v3_preview"
+
+
+def uses_compact_preview_runtime(
+    *,
+    clean_route_requested: bool,
+    turn_contract_enabled: bool,
+    agent_mode: bool,
+    agent_permitted: bool,
+    image_generation: bool,
+) -> bool:
+    """The single compact-runtime eligibility rule for one turn.
+
+    ``turn_contract_enabled`` is the route's contract policy for this turn
+    (exact approvals, TUI surface and full-schema routes opt out).
+    ``agent_permitted`` is false when the user's privileges demote the turn
+    to plain chat; image generation sessions run their own execution path.
+    """
+    return bool(
+        clean_route_requested
+        and turn_contract_enabled
+        and agent_mode
+        and agent_permitted
+        and not image_generation
+    )
+
+
+def is_compact_preview_contract(turn_contract) -> bool:
+    """Whether a turn contract was stamped for the compact runtime."""
+    return (
+        turn_contract is not None
+        and getattr(turn_contract, "selection_mode", None) == COMPACT_PREVIEW_MODE
+    )

src/clean_agent_preview.py

@@ -21,6 +21,7 @@ import httpx
 import jsonschema
 
 from src.context_compactor import prune_multimodal_images, trim_for_context
+from src.agent_runtime.runtime_selection import COMPACT_PREVIEW_MODE
 from src import agent_runs
 from src.agent_evidence import command_has_mutation_effect, workspace_artifact_is_usable
 from src.tool_capabilities import ToolEffect, ToolRunSecurityContext, capabilities_for_action
@@ -47,7 +48,7 @@ from src.model_profiles import (
 )
 
 ENDPOINT_ID = 'cleanv3'
-MODE = 'clean_compact_v3_preview'
+MODE = COMPACT_PREVIEW_MODE
 
 
 class ProviderStreamError(Exception):
@@ -4904,7 +4905,9 @@ async def preview_model_response(client, endpoint_url, headers, request, recover
     )
     transport_attempts = 0
     while True:
-        limit = recovery.get('context_limit')
+        # A provider-stated limit learned this turn overrides the window
+        # resolved at turn preparation.
+        limit = recovery.get('context_limit') or recovery.get('budget_limit')
         if limit:
             message_context = max(1, limit
                 - estimate_tool_schema_tokens(request.get('tools'))
@@ -4995,6 +4998,7 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                          client_runtime_context=None, max_tokens=768, max_rounds=8,
                          max_tool_calls=0,
                          external_tool_schemas=None, temperature=0.0,
+                         context_resolution=None,
                          request_authority=MISSING_AUTHORITY, **ignored):
     from src.generation_sampling import validate_temperature
     temperature = validate_temperature(temperature)
@@ -5249,7 +5253,18 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
     static_fetch_failed_urls = set()
     entity_result_links = {}
     calendar_create_confirmation = ''
-    context_recovery = {}
+    # Resolve the effective window once, before any model request, so the
+    # turn budgets against it and metrics report exactly what it ran under.
+    # The chat route prepares it; only callers arriving without one (or with
+    # one for a different route) resolve here. Terminal metrics must never
+    # start this discovery themselves.
+    if context_resolution is None or not context_resolution.applies_to(endpoint_url, model):
+        from src.agent_runtime.context_resolution import resolve_effective_context
+        context_resolution = await resolve_effective_context(
+            endpoint_url, model, headers=headers,
+            client_runtime_context=client_runtime_context,
+        )
+    context_recovery = {'budget_limit': context_resolution.budget_limit}
     successful_write = False
     editor_batch_pending = False
     editor_suggested_finds = []
@@ -8026,9 +8041,15 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
         return
     elapsed = time.monotonic() - started
     ttft = first_token - started if first_token else None
+    from src.agent_runtime.context_resolution import context_metrics
+    # Report the stored resolution plus any limit the provider stated during
+    # this turn. This is pure bookkeeping; no metadata request happens here.
+    context_resolution = context_resolution.observe_runtime_limit(
+        context_recovery.get('context_limit'))
     yield event({'type': 'metrics', 'data': {
         'email_task_scope': {**intent_accounting, 'failed': intent_scope_failed},
         'model': model, 'input_tokens': usage_in, 'output_tokens': usage_out,
+        **context_metrics(context_resolution, last_request_tokens),
         'total_tokens': usage_in + usage_out, 'response_time': round(elapsed, 3),
         'time_to_first_token': round(ttft, 3) if ttft is not None else None,
         'tokens_per_second': round(usage_out / elapsed, 2) if elapsed > 0 else 0,

src/context_compactor.py

@@ -481,12 +481,17 @@ async def maybe_compact(
     persist: bool = True,
     compaction_state: Optional[Dict[str, Any]] = None,
     deterministic: bool = False,
+    context_length: Optional[int] = None,
 ) -> tuple:
     """Check context usage and compact if above threshold.
 
+    ``context_length`` lets a caller that already resolved the turn's window
+    supply it, so this helper does not query the endpoint a second time.
+
     Returns (messages, context_length, was_compacted).
     """
-    context_length = get_context_length(endpoint_url, model)
+    if context_length is None:
+        context_length = get_context_length(endpoint_url, model)
     used = estimate_tokens(messages)
     pct = (used / context_length) * 100 if context_length else 0
     threshold = auto_compact_threshold_percent()