perf: preserve conversation prefixes across agent turns

tests/unit/agents/test_flash_runner_ledger.py

@@ -283,13 +283,13 @@ def test_injected_instruction_body_outlives_its_wrapper(mock_context):
     assert guidance.wrapper.startswith(USER_GUIDANCE_MARKER)
     assert "don't send messages" not in guidance.wrapper
 
-    # Scrub edge parity: once the turn has left the live position (a newer
-    # committed observation exists) the wrapper is gone and the body is
-    # still in the committed turn.
+    # Normal turns preserve guidance verbatim; an explicit compaction removes
+    # the old wrapper but preserves the user's standing instruction.
     ledger.stage_turn([tail, AIMessage(content="ok")])
     ledger.commit_staged(step_key="s1")
     ledger.stage_turn([runner._build_tail(ledger, 3, b"IMG", "[1] Send"), AIMessage(content="ok")])
     ledger.commit_staged(step_key="s2")
+    ledger.compact_active()
     rendered = ledger.render([runner._build_tail(ledger, 4, b"IMG", "[1] Send")])
     committed = next(
         m
@@ -679,8 +679,8 @@ async def test_run_calibrates_the_ledger_only_from_measured_usage(mock_context):
     assert record.call_count == 2
     first_call, second_call = record.call_args_list
     assert first_call.args[0] == 1200 and first_call.kwargs["messages"] is not None
-    # The report turn carries no usage metadata: estimated, recorded, not calibrated.
-    assert second_call.args[0] >= 1 and second_call.kwargs["messages"] is None
+    # Missing usage stays unknown; the next render computes its own estimate.
+    assert second_call.args[0] is None and second_call.kwargs["messages"] is None
 
 
 @pytest.mark.asyncio

tests/unit/agents/test_operator_prompt_estimate.py

@@ -30,7 +30,7 @@ from artemis.agents.operator.operator import OperatorNode
 from artemis.config.agent import MemoryTranscriptConfig
 from artemis.context import ArtemisContext
 from artemis.memory.step_memory import StepMemoryService
-from artemis.memory.transcript import TranscriptLedger, estimate_prompt_tokens
+from artemis.memory.transcript import TranscriptLedger
 
 SCREENSHOT_B64 = base64.b64encode(b"fake-jpeg-bytes").decode("utf-8")
 
@@ -92,7 +92,7 @@ def _llm(captured: list, usage_metadata=None):
 
 
 @pytest.mark.asyncio
-async def test_operator_records_an_estimate_when_the_provider_reports_no_usage():
+async def test_operator_uses_projected_context_without_fabricating_provider_usage():
     ctx = _ctx()
     captured: list = []
     with patch("artemis.agents.operator.operator.get_llm", return_value=_llm(captured)):
@@ -102,8 +102,8 @@ async def test_operator_records_an_estimate_when_the_provider_reports_no_usage()
     ledger = ctx.transcript_ledger
     assert isinstance(ledger, TranscriptLedger)
     assert captured, "the operator made no model call"
-    assert ledger.last_prompt_tokens == estimate_prompt_tokens(captured[-1])
-    assert ledger.last_prompt_tokens > 0
+    assert ledger.last_prompt_tokens is None
+    assert ledger.projected_prompt_tokens > 0
     # No measurement, no calibration.
     assert ledger.chars_per_token == 4.0
 

tests/unit/agents/test_operator_transcript.py

@@ -355,7 +355,7 @@ async def test_transcript_mode_two_turns_build_four_regions(tmp_path):
         # Text edge at depth 1: the committed observation has lost its plan
         # recitation as soon as the next tail exists (its screenshot stays
         # until the screenshot edge).
-        assert PLAN_RECITATION_MARKER not in str(committed_obs.content)
+        assert PLAN_RECITATION_MARKER in str(committed_obs.content)  # retained until compaction
         # The live tail still recites the plan.
         assert PLAN_RECITATION_MARKER in str(turn2_messages[-1].content)
         # Message layout is S + A + tail.