Improve agent system prompt for better token efficiency

Add explicit guidance on:
- Reading files with large ranges (500+ lines) instead of many small chunks
- Using read_file directly when target file is known vs search_and_read
- Cost awareness: planning refactors, avoiding redundant reads
- Tool call limits for complex tasks

1 files changed, 13 insertions, 4 deletions

diff --git a/Omni/Agent/Worker.hs b/Omni/Agent/Worker.hs
index 0e4071d..66f894d 100644
--- a/Omni/Agent/Worker.hs
+++ b/Omni/Agent/Worker.hs
@@ -399,15 +399,24 @@ buildBasePrompt ns repo =
     <> "- If tests pass, STOP. Do not verify again, do not double-check.\n\n"
     <> "# Tool Usage\n\n"
     <> "Your tools: read_file, write_file, edit_file, run_bash, search_codebase, search_and_read.\n\n"
-    <> "## Efficient Reading\n"
-    <> "- WRONG: read_file on a 2000-line file to find one function (wastes tokens).\n"
-    <> "- RIGHT: search_and_read with the function name (search + context in one call).\n"
-    <> "- RIGHT: search_codebase to find usages, then read_file with start_line/end_line.\n\n"
+    <> "## Efficient Reading (CRITICAL FOR BUDGET)\n"
+    <> "- Read files ONCE with large ranges (500+ lines), not many small 100-line chunks.\n"
+    <> "- WRONG: 10 separate read_file calls with 100-line ranges on the same file.\n"
+    <> "- RIGHT: 1-2 read_file calls with 500-1000 line ranges to cover the file.\n"
+    <> "- When you know the target file, use read_file directly with a path argument.\n"
+    <> "- WRONG: search_and_read across the whole repo when you know the file is Worker.py.\n"
+    <> "- RIGHT: read_file on Worker.py, or search_codebase with path='Worker.py'.\n"
+    <> "- search_and_read is for discovery when you DON'T know which file to look in.\n\n"
     <> "## Efficient Editing\n"
     <> "- Include enough context in old_str to match uniquely (usually 5+ lines).\n"
     <> "- If edit_file fails with 'old_str not found', you are hallucinating the content.\n"
     <> "- STOP. Call read_file on those exact lines to get fresh content. Then retry.\n"
     <> "- After 3 failed edits on the same file, reconsider your approach.\n\n"
+    <> "## Cost Awareness\n"
+    <> "- Each tool call costs tokens. Large file writes are expensive.\n"
+    <> "- For refactors: plan all new files first, then write them in order.\n"
+    <> "- Don't write a file, then immediately read it back - you just wrote it!\n"
+    <> "- Monitor your progress: if you're on tool call 30+ and not close to done, simplify.\n\n"
     <> "# Debugging\n"
     <> "If 'bild' fails, do NOT guess the fix.\n"
     <> "1. Read the error output carefully.\n"