space/screenjob
1
0
Fork 0
Code Issues Pull Requests Actions Releases Activity

docs: update context compaction prompt with observe-decide-act-verify loop
Some checks failed
CI / test (push) Failing after 8s
Details

Browse Source
This commit is contained in:
Space-Banane
2026-05-31 20:52:49 +02:00
parent 4123765aba
commit d514fe161c
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/agent.py
View File

@@ -183,6 +183,7 @@ def build_context_compaction_prompt(
"The retained visuals already represent the latest image tool calls and their results. Reuse them first; do not call see_screen again only because compaction happened.\n" "The retained visuals already represent the latest image tool calls and their results. Reuse them first; do not call see_screen again only because compaction happened.\n"
"Determine the current surface kind, foreground app, likely focus, pending dialog/browser state, and what changed since the last action before acting.\n" "Determine the current surface kind, foreground app, likely focus, pending dialog/browser state, and what changed since the last action before acting.\n"
"Use classify -> choose control channel -> execute one meaningful transition -> verify.\n" "Use classify -> choose control channel -> execute one meaningful transition -> verify.\n"
"Follow observe -> decide -> act -> verify.\n"
f"{_prohibited_key_combo_prompt(prohibited_key_combos)}" f"{_prohibited_key_combo_prompt(prohibited_key_combos)}"
"Re-observe with see_screen only when the retained visuals are stale or the UI likely changed. Add enhance for small or text-heavy details, and use get_active_window, detect_dialog, wait helpers, clipboard_get, or command stdout when they are the better verification channel.\n" "Re-observe with see_screen only when the retained visuals are stale or the UI likely changed. Add enhance for small or text-heavy details, and use get_active_window, detect_dialog, wait helpers, clipboard_get, or command stdout when they are the better verification channel.\n"
"If a fresh focus check or retained visual already proves a text field or editor is ready, act without demanding another screenshot.\n" "If a fresh focus check or retained visual already proves a text field or editor is ready, act without demanding another screenshot.\n"
@@ -1011,7 +1012,7 @@ class ScreenJobAgent:
"type": "function", "type": "function",
"name": "get_active_window", "name": "get_active_window",
"description": ( "description": (
"Return metadata for the current foreground window to verify focus, active app, surface kind, " "Return metadata for the current foreground window to verify focus and active app, surface kind, "
"browser state, and dialog classification." "browser state, and dialog classification."
), ),
"parameters": { "parameters": {