Stop estimate_tokens from double-counting text content (#868)

estimate_tokens added message.all_text() and then serialized every content
block (including text) with model_dump_json(), so text was counted twice and
a pure-text message estimated at ~2x its real size. The docstring says it
counts "text plus serialized non-text payloads"; skip content already covered
by all_text() in the per-content loop so text is counted once, while non-text
payloads are still serialized.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: chuenchen309

src/fast_agent/history/compaction.py

@@ -23,6 +23,7 @@
 from fast_agent.constants import FAST_AGENT_COMPACTION_CHANNEL
 from fast_agent.core.logging.logger import get_logger
 from fast_agent.event_progress import ProgressAction
+from fast_agent.mcp.helpers.content_helpers import get_text
 from fast_agent.mcp.prompt import Prompt
 from fast_agent.types.llm_stop_reason import LlmStopReason
 
@@ -185,6 +186,10 @@ def estimate_tokens(messages: list[PromptMessageExtended]) -> int:
     for message in messages:
         chars += len(message.all_text())
         for content in message.content:
+            if get_text(content) is not None:
+                # Text is already counted via all_text() above; only serialize
+                # non-text payloads here (matches the "non-text" docstring).
+                continue
             try:
                 chars += len(content.model_dump_json())
             except Exception:

tests/unit/fast_agent/history/test_compaction.py

@@ -16,6 +16,7 @@
 from fast_agent.config import CompactionSettings, Settings, get_settings
 from fast_agent.context import Context
 from fast_agent.history.compaction import (
+    _CHARS_PER_TOKEN,
     DEFAULT_COMPACTION_PROMPT,
     FAST_AGENT_COMPACTION_CHANNEL,
     CompactionSkipped,
@@ -359,6 +360,12 @@ def test_counts_non_text_content_and_channels(self):
 
         assert estimate_tokens([msg]) > plain + 5_000
 
+    def test_does_not_double_count_text(self):
+        # Text is counted once via all_text(); the per-content loop must skip it
+        # (docstring: "Count text plus serialized non-text payloads").
+        text = "x" * 1200
+        assert estimate_tokens([_user(text)]) == len(text) // _CHARS_PER_TOKEN
+
 
 class _FakeLLM:
     def __init__(self, summary: str = "SUMMARY OF WORK") -> None: