WEBDIENSTE/openswarm
1
0
Fork 0
mirror of https://github.com/openswarm-ai/openswarm.git synced 2026-09-22 01:24:52 +02:00
Code Issues Packages Projects Releases Wiki Activity

arena: v46 checkpoint self-verification done-gate (FCPAgent-style) -- capability work on long-horizon completion, pre-pilot

Browse Source
This commit is contained in:
ciregenz
2026-08-17 10:53:14 -07:00
parent 268f6948a7
commit 4b38269e51
1 changed files with 22 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
e2e/browser-v3/arena/llm_policy.py
+22
View File
@@ -198,6 +198,8 @@ class LlmPolicy:
answer_protocol: bool = False
# v45: read-your-writes doctrine (system rung, gated on state-change goals). Re-read + retry.
read_your_writes: bool = False
# v46: checkpoint self-verification done-gate (FCPAgent-style). Confirm each requirement before finishing.
done_gate: bool = False
# v43: answer-schema conformance gate (run.py-side). Validate send_msg against the task's own
# provided JSON schema and bounce non-conforming answers -- generic instruction-following.
schema_gate: bool = False
@@ -282,6 +284,9 @@ class LlmPolicy:
r"\b(post|comment|upvote|down ?vote|vote|create|add|edit|submit|reply|open an issue|change|update|delete|remove)\b", goal, re.I):
extra += ("\nSTATE-CHANGE TASK: after your write action, RE-READ and CONFIRM the change "
"is visible before finishing; if not visible, repeat the write.")
if getattr(self, "done_gate", False) and len(re.findall(r",|\band\b|;", goal)) >= 2:
extra += ("\nDONE-CHECK required before finishing: restate each requirement + its visible "
"evidence; if any lacks evidence and steps remain, keep working, do not finish.")
user = f"GOAL: {goal}{extra}\n\nACTIONS YOU ALREADY TOOK:\n{past}\n\nPAGE:\n{page}\n\nYour single next action:"
content: Any = user
if image_b64:
@@ -869,6 +874,13 @@ If the goal specifies a required response format or JSON schema (e.g. a FinalAge
fields like task_type and status), your FINAL send_msg_to_user MUST be exactly that JSON object
with all required fields filled -- not prose, not a bare value. Match the schema literally."""
OSW_SYSTEM_V46 = """
Before you FINISH a multi-part task, run an explicit completion check: restate EACH requirement the
task named, and for each state the concrete evidence on the page that it is done (the record exists,
the value shows, the count matches). If ANY requirement lacks evidence and you have steps left, DO
NOT finish -- go satisfy it. Only send your final answer once every requirement has visible evidence.
List the check as 'DONE-CHECK: [req1: evidence] [req2: evidence] ...' before finishing."""
OSW_SYSTEM_V45 = """
If the task CHANGES state (post, comment, vote, create, edit, submit, upload), a single click is
not enough -- the write often does not register on the first try. After the action, RE-READ: look
@@ -1114,6 +1126,16 @@ def build(name: str, model: str = "", endpoint: str = "", **_: Any) -> Any:
force_unblock=True, native_js_fallback=True, escape_token=True,
table_md=True, draw_circle=True, answer_protocol=True,
schema_gate=False, **v43) # DISABLED: gate falsely bounced valid JSON
if name == "osw-llm-v46": # v42 + checkpoint self-verification done-gate (FCPAgent-style)
v46 = dict(v7, system=OSW_SYSTEM_V8 + OSW_SYSTEM_V9_WIDGETS + OSW_SYSTEM_V16 + OSW_SYSTEM_V30
+ OSW_SYSTEM_V36 + OSW_SYSTEM_V46, max_tokens=900)
return OpenSwarmLlmPolicy(name=name, multi=True, vision="progressive", fastpath=True,
scripted_drag=True, auto_complete=True, som=False,
native_pickers=True, verify_terminal=True, post_mouse_vision=True,
multi_cap=6, fill_verify=True, dispatch=True, offscreen=True,
local_ctx=True, blocker_probe=True, suppress_wrappers=True,
force_unblock=True, native_js_fallback=True, escape_token=True,
table_md=True, answer_protocol=True, done_gate=True, **v46)
if name == "osw-llm-v45": # v42 + read-your-writes (state-change tasks: re-read + retry)
v45 = dict(v7, system=OSW_SYSTEM_V8 + OSW_SYSTEM_V9_WIDGETS + OSW_SYSTEM_V16 + OSW_SYSTEM_V30
+ OSW_SYSTEM_V36 + OSW_SYSTEM_V45, max_tokens=900)