aru-code 0.14.0__tar.gz → 0.14.1__tar.gz

{aru_code-0.14.0/aru_code.egg-info → aru_code-0.14.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aru-code
-Version: 0.14.0
+Version: 0.14.1
 Summary: A Claude Code clone built with Agno agents
 Author-email: Estevao <estevaofon@gmail.com>
 License-Expression: MIT

aru_code-0.14.1/aru/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.14.1"

{aru_code-0.14.0 → aru_code-0.14.1}/aru/context.py

@@ -11,7 +11,7 @@ from __future__ import annotations
 # ── Constants ──────────────────────────────────────────────────────
 
 # Pruning: minimum chars that must be freeable to justify a prune pass
-PRUNE_MINIMUM_CHARS = 20_000  # ~5.7K tokens
+PRUNE_MINIMUM_CHARS = 12_000  # ~3K tokens (lower = prune sooner)
 # Placeholder that replaces evicted content
 PRUNED_PLACEHOLDER = "[previous output cleared to save context]"
 # User messages larger than this threshold are truncated when outside protection window
@@ -24,16 +24,18 @@ PRUNE_PROTECT_TURNS = 2
 PRUNE_PROTECTED_MARKERS = {"[SubAgent-", "delegate_task"}
 
 # Truncation: universal limits for any tool output
-TRUNCATE_MAX_LINES = 500
-TRUNCATE_MAX_BYTES = 20 * 1024  # 20 KB
-TRUNCATE_KEEP_START = 350  # lines to keep from the start
-TRUNCATE_KEEP_END = 100  # lines to keep from the end
+TRUNCATE_MAX_LINES = 300
+TRUNCATE_MAX_BYTES = 15 * 1024  # 15 KB (was 20KB — tighter to prevent context bloat)
+TRUNCATE_KEEP_START = 200  # lines to keep from the start
+TRUNCATE_KEEP_END = 60  # lines to keep from the end
 TRUNCATE_MAX_LINE_LENGTH = 2000  # chars per individual line (prevents minified files)
 
 # Compaction: trigger when per-run input tokens exceed this fraction of model limit
-COMPACTION_THRESHOLD_RATIO = 0.85
+COMPACTION_THRESHOLD_RATIO = 0.70  # was 0.85 — compact earlier to avoid hitting limits
 # Compaction: target post-compaction size as fraction of model context limit
 COMPACTION_TARGET_RATIO = 0.15
+# Compaction: reserve buffer for the compaction process itself (like OpenCode's 20K)
+COMPACTION_BUFFER_TOKENS = 20_000
 # Default model context limits (input tokens)
 MODEL_CONTEXT_LIMITS: dict[str, int] = {
     # Anthropic
@@ -103,13 +105,13 @@ def _get_prune_protect_chars(model_id: str = "default") -> int:
     """Scale protection window based on model context size.
 
     Larger models get more protection; smaller models prune more aggressively
-    to delay compaction. Returns ~10% of the model's context in chars (~3.5 chars/token).
+    to prevent context overflow. Returns ~7% of the model's context in chars.
     """
     limit = MODEL_CONTEXT_LIMITS.get(model_id, MODEL_CONTEXT_LIMITS["default"])
-    # ~3.5 chars per token, protect ~10% of context
-    protect = int(limit * 0.10 * 3.5)
-    # Clamp between 20K (minimum usable) and 80K (diminishing returns)
-    return max(20_000, min(protect, 80_000))
+    # ~4 chars per token, protect ~7% of context (was 10% — tighter budget)
+    protect = int(limit * 0.07 * 4)
+    # Clamp between 15K (minimum usable) and 60K (diminishing returns)
+    return max(15_000, min(protect, 60_000))
 
 
 def prune_history(
@@ -278,7 +280,10 @@ def should_compact(
     history_or_tokens: int | list[dict[str, str]],
     model_id: str = "default",
 ) -> bool:
-    """Check if the conversation should be compacted (reactive, post-run).
+    """Check if the conversation should be compacted.
+
+    Uses OpenCode's approach: usable = model_limit - buffer, then
+    trigger when tokens >= usable * threshold_ratio.
 
     Accepts either an estimated token count (int) or the history list
     (from which tokens are estimated via char count).
@@ -288,7 +293,8 @@ def should_compact(
     else:
         tokens = history_or_tokens
     limit = MODEL_CONTEXT_LIMITS.get(model_id, MODEL_CONTEXT_LIMITS["default"])
-    threshold = int(limit * COMPACTION_THRESHOLD_RATIO)
+    usable = limit - COMPACTION_BUFFER_TOKENS
+    threshold = int(usable * COMPACTION_THRESHOLD_RATIO)
     return tokens >= threshold
 
 

{aru_code-0.14.0 → aru_code-0.14.1}/aru/runner.py

@@ -249,12 +249,15 @@ async def run_agent_capture(agent, message: str, session=None, lightweight: bool
             run_input_tokens = getattr(run_output.metrics, "input_tokens", 0) or 0
             if should_compact(run_input_tokens, session.model_id):
                 try:
+                    # Always prune first to shrink history before compaction
+                    session.history = prune_history(session.history, model_id=session.model_id)
                     session.history = await compact_conversation(
                         session.history, session.model_ref, session.plan_task,
                         model_id=session.model_id,
                     )
                     console.print("[dim]Context compacted to save tokens.[/dim]")
                 except Exception:
+                    # Even if compaction fails, keep the pruned history
                     pass
 
         final_content = accumulated or final_content

{aru_code-0.14.0 → aru_code-0.14.1}/aru/tools/codebase.py

@@ -54,23 +54,23 @@ def _format_diff(old_string: str, new_string: str) -> Group:
 
 
 
-# Hard ceiling per tool result (~15K tokens). Even max_size=0 respects this per chunk.
-_READ_HARD_CAP = 60_000  # bytes
+# Hard ceiling per tool result (~10K tokens). Even max_size=0 respects this per chunk.
+_READ_HARD_CAP = 40_000  # bytes (was 60K — tighter to protect context)
 
 def clear_read_cache():
     """Clear the read cache. Call after file mutations to avoid stale data."""
     get_ctx().read_cache.clear()
 
 
-def read_file(file_path: str, start_line: int = 0, end_line: int = 0, max_size: int = 15_000) -> str:
+def read_file(file_path: str, start_line: int = 0, end_line: int = 0, max_size: int = 12_000) -> str:
     """Read file contents. Returns chunked output for large files.
 
     Args:
         file_path: Path to the file (absolute or relative).
         start_line: First line (1-indexed, inclusive). 0 = beginning.
         end_line: Last line (1-indexed, inclusive). 0 = end.
-        max_size: Max bytes before truncation. Default 15KB.
-            Set to 0 to read the full file in chunks — each chunk up to ~60KB.
+        max_size: Max bytes before truncation. Default 12KB.
+            Set to 0 to read the full file in chunks — each chunk up to ~40KB.
             The first chunk includes a continuation hint so you can call again
             with start_line to get the next chunk.
     """
@@ -204,7 +204,7 @@ async def read_file_smart(file_path: str, query: str) -> str:
         query: The specific question you want answered about this file.
     """
     # Read raw content first (reuse existing read_file logic)
-    raw = read_file(file_path, max_size=20_000)
+    raw = read_file(file_path, max_size=15_000)
 
     if raw.startswith("Error:"):
         return raw

{aru_code-0.14.0 → aru_code-0.14.1/aru_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aru-code
-Version: 0.14.0
+Version: 0.14.1
 Summary: A Claude Code clone built with Agno agents
 Author-email: Estevao <estevaofon@gmail.com>
 License-Expression: MIT

{aru_code-0.14.0 → aru_code-0.14.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "aru-code"
-version = "0.14.0"
+version = "0.14.1"
 description = "A Claude Code clone built with Agno agents"
 readme = "README.md"
 license = "MIT"

aru_code-0.14.0/aru/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.14.0"