core/agent: calibrate soft-compact with last-request prompt_tokens

After the first model call, scale char-based token estimates by
API/resolved prompt size so budget checks track near-real tokens.
/compact uses the same calibration when available.

src/plyngent/agent/budget.py

@@ -11,12 +11,14 @@ from plyngent.lmproto.openai_compatible.model import (
 )
 
 if TYPE_CHECKING:
-    from collections.abc import Sequence
+    from collections.abc import Callable, Sequence
 
     from plyngent.lmproto.openai_compatible.model import AnyChatMessage
 
+    type TokenMeasure = Callable[[Sequence[AnyChatMessage]], int]
+
 DEFAULT_TOOL_RESULT_MAX_CHARS = 32_000
-# Soft context budget in estimated tokens (~4 chars/token); not a hard model limit.
+# Soft context budget in tokens (API-calibrated when possible; else ~4 chars/token).
 DEFAULT_CONTEXT_MAX_TOKENS = 200_000
 DEFAULT_OLD_TOOL_RESULT_CHARS = 800
 DEFAULT_RECENT_TOOL_RESULTS = 4
@@ -64,12 +66,37 @@ def estimate_messages_chars(messages: Sequence[AnyChatMessage]) -> int:
 
 
 def estimate_messages_tokens(messages: Sequence[AnyChatMessage]) -> int:
-    """Char-based token estimate for soft context budget checks."""
+    """Char-based token estimate (fallback when no API calibration is available)."""
     from plyngent.agent.usage import chars_to_tokens
 
     return chars_to_tokens(estimate_messages_chars(messages))
 
 
+def measure_messages_tokens(
+    messages: Sequence[AnyChatMessage],
+    *,
+    prompt_tokens_hint: int | None = None,
+    sent_estimate_tokens: int | None = None,
+) -> int:
+    """Token size for budget checks.
+
+    When ``prompt_tokens_hint`` is the last request's API (or resolved) prompt
+    size and ``sent_estimate_tokens`` is the char-estimate of that same payload,
+    scale the current char-estimate by ``hint / sent_estimate`` so soft-compact
+    tracks near-real tokens after the first model call.
+    """
+    est = estimate_messages_tokens(messages)
+    if (
+        prompt_tokens_hint is not None
+        and prompt_tokens_hint > 0
+        and sent_estimate_tokens is not None
+        and sent_estimate_tokens > 0
+    ):
+        scaled = round(est * (prompt_tokens_hint / sent_estimate_tokens))
+        return max(1, scaled) if est > 0 else 0
+    return est
+
+
 def _shrink_tool(message: ToolChatMessage, max_chars: int) -> ToolChatMessage:
     if len(message.content) <= max_chars:
         return message
@@ -109,13 +136,14 @@ def _shrink_largest(
     *,
     max_tokens: int,
     shrink_cap: int,
+    measure: TokenMeasure,
 ) -> None:
     def tool_len(i: int) -> int:
         msg = messages[i]
         return len(msg.content) if isinstance(msg, ToolChatMessage) else 0
 
     for idx in sorted(tool_indices, key=tool_len, reverse=True):
-        if estimate_messages_tokens(messages) <= max_tokens:
+        if measure(messages) <= max_tokens:
             return
         tool_msg = messages[idx]
         if isinstance(tool_msg, ToolChatMessage):
@@ -128,17 +156,29 @@ def compact_messages_for_request(
     max_tokens: int = DEFAULT_CONTEXT_MAX_TOKENS,
     old_tool_result_chars: int = DEFAULT_OLD_TOOL_RESULT_CHARS,
     keep_recent_tool_results: int = DEFAULT_RECENT_TOOL_RESULTS,
-    # Deprecated alias: treated as token budget if max_tokens not overridden via callers.
+    prompt_tokens_hint: int | None = None,
+    sent_estimate_tokens: int | None = None,
+    # Deprecated alias: treated as token budget.
     max_chars: int | None = None,
 ) -> list[AnyChatMessage]:
     """Return a request-time copy with older tool dumps shrunk if over budget.
 
-    Budget is in **estimated tokens** (char/4). Does not mutate the original history.
+    Budget is in tokens. Prefer API-calibrated measurement via
+    ``prompt_tokens_hint`` / ``sent_estimate_tokens`` (last request); otherwise
+    fall back to char/4. Does not mutate the original history.
     ``max_tokens < 1`` disables compacting.
     """
     budget = max_tokens if max_chars is None else max_chars
+
+    def measure(msgs: Sequence[AnyChatMessage]) -> int:
+        return measure_messages_tokens(
+            msgs,
+            prompt_tokens_hint=prompt_tokens_hint,
+            sent_estimate_tokens=sent_estimate_tokens,
+        )
+
     out: list[AnyChatMessage] = list(messages)
-    if budget < 1 or estimate_messages_tokens(out) <= budget:
+    if budget < 1 or measure(out) <= budget:
         return out
 
     indices = _tool_indices(out)
@@ -147,7 +187,7 @@ def compact_messages_for_request(
 
     protect = _protect_indices(indices, keep_recent_tool_results)
     _shrink_except(out, indices, protect, old_tool_result_chars)
-    if estimate_messages_tokens(out) <= budget:
+    if measure(out) <= budget:
         return out
 
     _shrink_largest(
@@ -155,5 +195,6 @@ def compact_messages_for_request(
         indices,
         max_tokens=budget,
         shrink_cap=max(64, old_tool_result_chars // 2),
+        measure=measure,
     )
     return out

src/plyngent/agent/compact.py

@@ -68,9 +68,16 @@ def soft_compact_transcript(
     messages: Sequence[AnyChatMessage],
     *,
     max_tokens: int = DEFAULT_CONTEXT_MAX_TOKENS,
+    prompt_tokens_hint: int | None = None,
+    sent_estimate_tokens: int | None = None,
 ) -> str:
     """Soft-compact tool dumps then format as transcript text."""
-    compacted = compact_messages_for_request(messages, max_tokens=max_tokens)
+    compacted = compact_messages_for_request(
+        messages,
+        max_tokens=max_tokens,
+        prompt_tokens_hint=prompt_tokens_hint,
+        sent_estimate_tokens=sent_estimate_tokens,
+    )
     return format_transcript(compacted)
 
 
@@ -81,12 +88,19 @@ async def summarize_messages(
     model: str,
     max_context_tokens: int = DEFAULT_CONTEXT_MAX_TOKENS,
     temperature: float | None = 0.2,
+    prompt_tokens_hint: int | None = None,
+    sent_estimate_tokens: int | None = None,
 ) -> str:
     """Soft-compact history and ask the model for a dense summary (no tools)."""
     if not messages:
         msg = "nothing to compact"
         raise ValueError(msg)
-    transcript = soft_compact_transcript(messages, max_tokens=max_context_tokens)
+    transcript = soft_compact_transcript(
+        messages,
+        max_tokens=max_context_tokens,
+        prompt_tokens_hint=prompt_tokens_hint,
+        sent_estimate_tokens=sent_estimate_tokens,
+    )
     if not transcript.strip():
         msg = "nothing to compact"
         raise ValueError(msg)

src/plyngent/agent/loop.py

@@ -23,6 +23,7 @@ from .budget import (
     DEFAULT_CONTEXT_MAX_TOKENS,
     DEFAULT_TOOL_RESULT_MAX_CHARS,
     compact_messages_for_request,
+    estimate_messages_tokens,
     truncate_tool_result,
 )
 from .events import (
@@ -240,6 +241,9 @@ async def run_chat_loop(
 
     rounds_used = 0
     allowance = max_rounds
+    # Calibrate soft-compact from last model call's prompt_tokens (API preferred).
+    prompt_tokens_hint: int | None = None
+    sent_estimate_tokens: int | None = None
 
     while True:
         while rounds_used < allowance:
@@ -247,7 +251,10 @@ async def run_chat_loop(
             request_messages = compact_messages_for_request(
                 messages,
                 max_tokens=max_context_tokens,
+                prompt_tokens_hint=prompt_tokens_hint,
+                sent_estimate_tokens=sent_estimate_tokens,
             )
+            sent_est = estimate_messages_tokens(request_messages)
             param = ChatCompletionsParam(
                 messages=request_messages,
                 model=model,
@@ -257,6 +264,10 @@ async def run_chat_loop(
 
             pre_len = len(messages)
             async for event in _assistant_round(client, param, messages, stream=stream):
+                if isinstance(event, UsageEvent):
+                    # Next rounds scale char-estimates by real/resolved prompt size.
+                    prompt_tokens_hint = event.usage.prompt_tokens
+                    sent_estimate_tokens = sent_est
                 yield event
             assistant = _last_assistant(messages, pre_len)
             tool_calls = assistant.tool_calls

src/plyngent/cli/state.py

@@ -170,11 +170,22 @@ class ReplState:
             msg = "nothing to compact (empty history)"
             raise ValueError(msg)
 
+        # Prefer last API prompt_tokens to drive soft-compact toward real size.
+        hint: int | None = None
+        sent_est: int | None = None
+        if not self.agent.last_request_usage.is_zero():
+            from plyngent.agent.budget import estimate_messages_tokens
+
+            hint = self.agent.last_request_usage.prompt_tokens
+            # Approximate: calibrate against current full history char-est.
+            sent_est = estimate_messages_tokens(messages)
         summary = await summarize_messages(
             self.client,
             messages,
             model=self.model,
             max_context_tokens=self.agent.max_context_tokens,
+            prompt_tokens_hint=hint,
+            sent_estimate_tokens=sent_est,
         )
         session_name = name or f"compact-from-{old_id}"
         await self.new_session(name=session_name)