fix(agent): align progress updates with Codex cadence

This commit is contained in:
jxxghp
2026-08-07 07:01:03 +08:00
parent ea8d1f8d26
commit 6d3161f3cb
2 changed files with 12 additions and 4 deletions

View File

@@ -145,12 +145,16 @@ class PromptManager:
if not settings.AI_AGENT_VERBOSE:
verbose_spec = (
"\n\n[Important Instruction] STRICTLY ENFORCED: "
"Keep tool-related narration concise and useful. "
"Use a concise Codex-style progress cadence. "
"For a quick single-step tool call, call the tool directly without a filler acknowledgement. "
"For multi-step work or any task that may take noticeable time, send one brief user-facing progress update "
"before the first tool call, then another short update after a meaningful milestone or several tool calls. "
"State what you are checking or what you learned, without exposing hidden reasoning, raw tool arguments, "
"or repetitive per-tool narration. Continue working after each update and finish with a concise final result."
"before the first tool call. While work continues, send another update after a meaningful milestone or "
"several tool calls, and approximately every 30 to 60 seconds whenever you regain control. "
"Keep each update to one or two short sentences that state what you are doing, what you have learned, "
"and what comes next when those details are useful. Vary the wording and do not repeat an unchanged status. "
"Do not expose hidden reasoning, raw tool arguments, or repetitive per-tool narration. "
"Continue working after each update. The final reply must be self-contained and summarize the outcome "
"without relying on the user having read the progress updates."
)
# MoviePilot系统信息

View File

@@ -14,6 +14,10 @@ def test_non_verbose_prompt_keeps_long_running_tasks_user_visible() -> None:
assert "before the first tool call" in prompt
assert "after a meaningful milestone or several tool calls" in prompt
assert "approximately every 30 to 60 seconds whenever you regain control" in prompt
assert "one or two short sentences" in prompt
assert "do not repeat an unchanged status" in prompt
assert "Continue working after each update" in prompt
assert "The final reply must be self-contained" in prompt
assert "remain completely silent" not in prompt
assert "DO NOT output any intermediate content" not in prompt