WEBDIENSTE/openswarm
1
0
Fork 0
mirror of https://github.com/openswarm-ai/openswarm.git synced 2026-09-07 02:07:45 +02:00
Code Issues Packages Projects Releases Wiki Activity

[eric] browser: orchestrator translates the OUTCOME line into a plain human confirmation; drop hardcoded task-specific examples from prompts

Browse Source
This commit is contained in:
ciregenz
2026-06-07 23:44:39 -07:00
parent d5fc146123
commit 725578b2a8
2 changed files with 16 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/apps/agents/browser/browser_schema.py
+7 -5
View File
@@ -409,9 +409,10 @@ BROWSER_TOOLS_SCHEMA = [
"type": "string",
"description": (
"Optional. A specific button label, visible text, or CSS selector "
"you expect to appear (e.g. 'Haik Decie', 'Write a message', "
"'button[type=submit]'). The wait ends the moment it's present and "
"visible. Be specific, not a generic word like 'Message'."
"you expect to appear (e.g. the person's name you searched, a "
"placeholder like 'Write a message', or 'button[type=submit]'). The "
"wait ends the moment it's present and visible. Be specific, not a "
"generic word like 'Message'."
),
},
},
@@ -795,8 +796,9 @@ SYSTEM_PROMPT = (
"line: 'OUTCOME: DONE - <the verifiable result>' or 'OUTCOME: NOT DONE - <what is "
"missing and why>'. The parent agent decides whether to re-dispatch another browser "
"from that line alone, so a vague or missing outcome line costs a whole redundant "
"agent run. For irreversible actions, DONE requires the proof you observed "
"(e.g. 'message visible in thread at 9:31 PM'). Keep the rest of the summary brief."
"agent run. For irreversible actions, DONE requires the proof you observed: the exact "
"thing now on the page that confirms it, and where or when you saw it. Keep the rest "
"of the summary brief."
)
MAX_TURNS = 40
backend/apps/agents/manager/prompt/prompt_context.py
+9 -5
View File
@@ -139,12 +139,16 @@ def _build_browser_context(dashboard_id: str | None, selected_browser_ids: list[
"and faster than spawning one agent per item with BrowserAgents, use parallel "
"BrowserAgents only for genuinely DIFFERENT tasks, not for the same flow repeated.",
"",
"**Trust the agent's OUTCOME line; don't redo its work or re-narrate it.** "
"**Trust the agent's OUTCOME line for your DECISION; TRANSLATE it for the user.** "
"Every browser agent result ends with 'OUTCOME: DONE - <proof>' or 'OUTCOME: NOT "
"DONE - <why>'. DONE with proof means complete: reply to the user in 1-2 short "
"sentences passing that proof along, and do NOT dispatch a verification agent. "
"NOT DONE means re-dispatch with a sharper task (start from what the agent "
"reported), not a duplicate of the old one. Long restatements of what the agent "
"DONE - <why>'. That line is internal plumbing for YOU, never show it to the user. "
"DONE with proof means complete: confirm it to the user the way a helpful person "
"would, one short natural sentence saying what got done plus the human-meaningful "
"proof from the result (the name, the time, the title). NEVER echo the literal "
"'OUTCOME:' tag or the agent's UI mechanics (composer, textbox, element indices, "
"'value now empty'), those mean nothing to a user; and do NOT dispatch a "
"verification agent. NOT DONE means re-dispatch with a sharper task (start from "
"what the agent reported), not a duplicate. Long restatements of what the agent "
"already said just slow the user down.",
]