Remove the main agent's ReAct turn cap; fix dark-mode legibility in the tool-call panel

src/backend/core/evolution/agent_profile.py

@@ -90,8 +90,13 @@
 # out-of-range value hides the fact that something proposed it.
 MIN_BUDGET_MULTIPLIER = 0.25
 MAX_BUDGET_MULTIPLIER = 3.0
+# 0 is the default and means "no turn cap" — the main agent is unbounded unless
+# a published profile deliberately fences it. Any other value must land in the
+# band below; the band starts at 3 because a 1-2 turn main agent cannot finish a
+# single tool round-trip.
 MIN_REACT_TURNS = 3
 MAX_REACT_TURNS = 80
+UNBOUNDED_REACT_TURNS = 0
 MIN_MEMORY_BUDGET_MS = 100
 MAX_MEMORY_BUDGET_MS = 5000
 MIN_LOOP_ATTEMPTS = 2
@@ -178,7 +183,10 @@ class AgentProfile:
     subagent_routes: List[SubagentRoute] = field(default_factory=list)
 
     memory_budget_ms: int = 600
-    max_react_turns: int = 50
+    # 0 = 不限（默认）。See UNBOUNDED_REACT_TURNS: the main agent's round ceiling
+    # was removed, so a turn budget here is an opt-in narrowing a published
+    # profile performs deliberately, not a floor every profile inherits.
+    max_react_turns: int = UNBOUNDED_REACT_TURNS
     budget_multiplier: float = 1.0
     reviewer_checkpoints: List[str] = field(default_factory=list)
     intervention_rules: List[InterventionRule] = field(default_factory=list)
@@ -225,7 +233,10 @@ def from_dict(cls, raw: Dict[str, Any]) -> "AgentProfile":
                 SubagentRoute.from_dict(r) for r in (raw.get("subagent_routes") or [])
             ],
             memory_budget_ms=int(raw.get("memory_budget_ms") or 600),
-            max_react_turns=int(raw.get("max_react_turns") or 20),
+            # Missing key → unbounded, not an invented ceiling: a profile
+            # published for unrelated reasons (tool allowlist, memory budget)
+            # must not silently re-fence the main loop.
+            max_react_turns=int(raw.get("max_react_turns") or UNBOUNDED_REACT_TURNS),
             budget_multiplier=float(raw.get("budget_multiplier") or 1.0),
             reviewer_checkpoints=[str(c) for c in (raw.get("reviewer_checkpoints") or [])],
             intervention_rules=[
@@ -253,11 +264,10 @@ def builtin_profile() -> AgentProfile:
         version="v0",
         tool_allowlist=None,
         memory_budget_ms=600,
-        # The main agent's existing cap, reproduced exactly. Installing profile
-        # resolution must change nothing until a profile is published; a
-        # "sensible default" here would be a behaviour change disguised as a
-        # refactor.
-        max_react_turns=50,
+        # The main agent's runtime default, reproduced exactly — which is now
+        # "no cap". This field only bounds the loop when a published profile
+        # sets it on purpose.
+        max_react_turns=UNBOUNDED_REACT_TURNS,
         budget_multiplier=1.0,
         loop_max_attempts_per_requirement=6,
         loop_strategy_change_after=2,
@@ -323,9 +333,12 @@ def validate_profile(
             f"预算倍数 {profile.budget_multiplier} 超出 "
             f"[{MIN_BUDGET_MULTIPLIER}, {MAX_BUDGET_MULTIPLIER}]"
         )
-    if not (MIN_REACT_TURNS <= profile.max_react_turns <= MAX_REACT_TURNS):
+    if profile.max_react_turns != UNBOUNDED_REACT_TURNS and not (
+        MIN_REACT_TURNS <= profile.max_react_turns <= MAX_REACT_TURNS
+    ):
         problems.append(
-            f"最大轮数 {profile.max_react_turns} 超出 [{MIN_REACT_TURNS}, {MAX_REACT_TURNS}]"
+            f"最大轮数 {profile.max_react_turns} 超出 "
+            f"[{MIN_REACT_TURNS}, {MAX_REACT_TURNS}]（{UNBOUNDED_REACT_TURNS} 表示不限）"
         )
     if not (MIN_MEMORY_BUDGET_MS <= profile.memory_budget_ms <= MAX_MEMORY_BUDGET_MS):
         problems.append(

src/backend/core/llm/agent_factory.py

@@ -80,6 +80,7 @@
     "确认后系统会自动逐条执行并把结果实时推送给用户，无需你重复调用。\n"
     "若请求确实只针对单一对象/单一概念，再走普通回答即可。\n"
 )
+
 from orchestration.registry import AgentSpec
 
 load_dotenv()
@@ -92,6 +93,32 @@
 _HTTP_MCP_FAIL_COOLDOWN_S = 60.0
 
 
+def _render_turn_budget_hint(max_iters: int) -> str:
+    """Tell an agent that *does* have a turn budget how to spend it.
+
+    Stating the budget up front is cheaper than shouting about it at the end:
+    the wrap-up reminder (:class:`IterBudgetReminderMiddleware`) can only ask a
+    loop that already burned its rounds one tool at a time to salvage
+    something, while this changes how the rounds get spent in the first place.
+    Parallel fan-out within one round is the lever — a round is one reasoning
+    step, not one tool call, so the budget binds serial round-trips, never the
+    total number of tool calls.
+
+    Only rendered where a bound actually exists (sub-agents, turbo, custom
+    agents, an explicit operator cap). The main agent is unbounded and must not
+    be told otherwise — a false scarcity claim would make it cut work short.
+    """
+    return (
+        "\n\n## 轮次预算\n"
+        f"本次运行最多 {max_iters} 轮「推理 → 工具调用」。一轮里可以同时发起任意多个"
+        "工具调用，所以受限的是串行往返次数，不是工具调用总数。\n"
+        "- 先想清楚需要哪些信息，再**在同一轮里并行发起**全部彼此独立的调用"
+        "（一次读多个文件、一次检索多个关键词），不要一轮只调一个。\n"
+        "- 只有后一步的参数确实要等前一步的结果时，才分成下一轮。\n"
+        "- 同一个操作连续失败两次就换思路或如实报告，不要用剩余轮次反复重试。\n"
+    )
+
+
 def _effective_mcp_server_keys(
     cfg,
     agent_spec: Optional[AgentSpec],
@@ -2053,29 +2080,50 @@ def _build_toolkit() -> Toolkit:
     # toolkit.get_agent_skill_prompt(), so no separate hook is needed.
 
     # ── Resolve agent name and max_iters ──
-    _DEFAULT_MAIN_ITERS = 50
+    #
+    # **The main agent has no turn cap.** A fixed round ceiling bounds the wrong
+    # axis: what a long task actually exhausts is context, not rounds, and any
+    # number picked here is simultaneously too low for report-scale work and too
+    # high to catch a genuine runaway. Neither of the two things a cap was
+    # supposed to buy needs it:
+    #   - runaway protection lives in the chat-run watchdog, which can see
+    #     wall-clock and output (CHAT_RUN_INACTIVITY_TIMEOUT_SEC /
+    #     CHAT_RUN_MAX_AGE_SEC / CHAT_RUN_HARD_MAX_AGE_SEC reap silent and
+    #     immortal runs regardless of how many rounds they took);
+    #   - context exhaustion is handled by compaction.
+    # ``_UNBOUNDED_ITERS`` is a loop backstop, not a budget — AgentScope's
+    # ReActConfig needs an int, and this one sits far above any real turn.
+    #
+    # Bounded budgets survive only where the bound is a deliberate contract:
+    # sub-agents (a delegated task that must come back), turbo's quick-lookup
+    # cap, a custom agent's own ``max_iters``, a published profile's turn
+    # budget, and the CHAT_MAIN_MAX_ITERS opt-in for operators who do want the
+    # main agent fenced.
+    _UNBOUNDED_ITERS = 100_000
     _DEFAULT_SUBAGENT_ITERS = 10
     _agent_name = "hugagent_agent"
-    # The active profile's turn budget governs the main agent. The built-in
-    # profile carries the same 50 this used, so resolution is a no-op until a
-    # profile is published.
-    _max_iters = profile.max_react_turns if user_agent is None else _DEFAULT_MAIN_ITERS
+    _max_iters = _UNBOUNDED_ITERS
     if max_iters is not None:
         _max_iters = max_iters
     elif user_agent is not None:
         _agent_name = (
             f"subagent_{user_agent.agent_id}" if isolated else f"agent_{user_agent.agent_id}"
         )
         _max_iters = user_agent.max_iters or (
-            _DEFAULT_SUBAGENT_ITERS if isolated else _DEFAULT_MAIN_ITERS
+            _DEFAULT_SUBAGENT_ITERS if isolated else _UNBOUNDED_ITERS
         )
     elif isolated:
         _max_iters = _DEFAULT_SUBAGENT_ITERS
     else:
-        # Main-agent env override for long-running tasks. Wins over the profile
-        # value: the profile validation range tops out at 80 turns, far below
-        # what e.g. a multi-hour report-generation run needs (container restart
-        # required to change, like all env config).
+        # Both main-agent caps are opt-in and absent by default: the built-in
+        # profile now carries 0 ("no cap"), so profile resolution only bounds the
+        # loop when someone publishes a profile that deliberately sets a turn
+        # budget. The env override still wins over it — the profile validation
+        # range tops out at 80 turns, far below what e.g. a multi-hour
+        # report-generation run needs (container restart required to change,
+        # like all env config).
+        if profile.max_react_turns:
+            _max_iters = profile.max_react_turns
         from core.config.settings import _env, _int
 
         _env_iters = _int(_env("CHAT_MAIN_MAX_ITERS"), 0)
@@ -2096,6 +2144,13 @@ def _build_toolkit() -> Toolkit:
         # else: 开了代码执行位且模式没配上限 → 不套极速的检索档硬顶（默认 4 轮
         # 会把跑代码的任务掐死），按 profile/env 的常规上限走。
 
+    # Budget spending policy, injected only where a budget exists. Appended last
+    # so it sees the final number after every narrowing above (turbo included);
+    # skipped without tools, where "spend your rounds on parallel calls" has
+    # nothing to describe.
+    if _max_iters < _UNBOUNDED_ITERS and not disable_tools:
+        system_prompt += _render_turn_budget_hint(_max_iters)
+
     # ── Create the Agent (AgentScope 2.0) ──
     # Note: long_term_memory is not passed — mem0 is fully stripped from the SSE
     # main path (manual non-blocking pipeline).

src/backend/tests/evolution/test_policies_and_memory_ops.py

@@ -22,8 +22,13 @@ def test_builtin_profile_reproduces_todays_constants():
     policy = PL.policy_from_profile(AP.builtin_profile())
     assert policy.max_attempts_per_requirement == 6
     assert policy.strategy_change_after == 2
-    # The main ReAct agent's existing cap, unchanged.
-    assert AP.builtin_profile().max_react_turns == 50
+    # The main ReAct agent has no turn cap; the built-in profile must not
+    # reintroduce one behind the runtime's back.
+    assert AP.builtin_profile().max_react_turns == AP.UNBOUNDED_REACT_TURNS
+    # A profile published without an explicit turn budget stays unbounded too —
+    # a missing key must not become an invented ceiling.
+    unspecified = AP.AgentProfile.from_dict({"profile_id": "p"})
+    assert unspecified.max_react_turns == AP.UNBOUNDED_REACT_TURNS
 
 
 def test_longer_stall_escalates_rather_than_repeating_the_mild_response():

src/backend/tests/test_iter_budget_middleware.py

@@ -143,3 +143,39 @@ def test_force_text_kill_switch(monkeypatch):
     agent = _fake_agent(max_iters=10, cur_iter=9)
     out = _force(mw, agent, {"tool_choice": None})
     assert out.get("tool_choice") is None
+
+
+# ── Turn-budget hint (static policy, injected only where a budget exists) ──
+
+
+def test_turn_budget_hint_states_the_actual_number():
+    from core.llm.agent_factory import _render_turn_budget_hint
+
+    hint = _render_turn_budget_hint(12)
+    assert "12 轮" in hint
+    # The point of the hint is the parallel fan-out strategy, not the scary
+    # number — one that only announces a limit makes the model cautious
+    # rather than efficient.
+    assert "并行" in hint
+
+
+def test_turn_budget_hint_is_byte_stable():
+    """Prefix-cache safety: same budget → same bytes, every request."""
+    from core.llm.agent_factory import _render_turn_budget_hint
+
+    assert _render_turn_budget_hint(30) == _render_turn_budget_hint(30)
+
+
+def test_unbounded_profile_is_valid_and_bad_bounds_are_not():
+    from core.evolution import agent_profile as AP
+
+    unbounded = AP.builtin_profile()
+    assert unbounded.max_react_turns == AP.UNBOUNDED_REACT_TURNS
+    ok, problems = AP.validate_profile(unbounded)
+    assert ok, problems
+
+    fenced = AP.builtin_profile()
+    fenced.max_react_turns = 1  # below MIN_REACT_TURNS, and not the 0 sentinel
+    ok, problems = AP.validate_profile(fenced)
+    assert ok is False
+    assert any("最大轮数" in p for p in problems)

src/frontend/src/styles/chat.css

@@ -598,13 +598,32 @@
   color:transparent;
   animation:jx-kf-shimmer 1.6s linear infinite;
 }
+/* 深色下把流光方向掰回来：上面那条渐变是「slate-400 打底 + slate-600 高光」，
+   靠"把字压暗"做高光——浅色底成立，暗底上就完全反了，高光处 (#475569) 对比度不到 2，
+   于是「深度拥抱中…」会朝着看不见的方向脉动。深色档改成暗→亮，高光才是最亮的一点。 */
+:root[data-theme="dark"] .jx-turnStatus-label{
+  background-image:linear-gradient(90deg,
+    color-mix(in srgb, var(--color-text) 52%, transparent) 0%,
+    color-mix(in srgb, var(--color-text) 96%, transparent) 50%,
+    color-mix(in srgb, var(--color-text) 52%, transparent) 100%);
+  background-size:200% 100%;
+  /* ⚠️ 必须用 background-image 而不是 background 简写：简写会把 background-clip 一并
+     重置回 border-box，这行字正是靠 clip:text 把渐变裁进字形的，重置后整行会变成一块
+     实心灰方块、文字消失（已踩）。这两行是保险，即便将来改回简写也不至于塌掉。 */
+  -webkit-background-clip:text;
+  background-clip:text;
+}
 .jx-turnStatus-clock{
   margin-left:2px;
   font-size:12px;
   font-weight:400;
   font-variant-numeric:tabular-nums;
   color:var(--color-text-placeholder);
 }
+/* 与折叠摘要同理：placeholder 令牌在深色档只有 2.65 对比度，计时数字看不清 */
+:root[data-theme="dark"] .jx-turnStatus-clock{
+  color:var(--color-text-tertiary);
+}
 
 
 /* ── 上传文件预览卡片（输入区 & 用户消息） ──────── */
@@ -4162,8 +4181,8 @@ textarea.jx-composer{
 .jx-deploySwitcherMenu {
   position: absolute; top: calc(100% + 6px); left: 50%; transform: translateX(-50%);
   min-width: 288px; z-index: 50; padding: 6px;
-  background: var(--color-bg-white, #fff);
-  border: 1px solid var(--color-border, #e5e6eb);
+  background: var(--color-bg-elevated);
+  border: 1px solid var(--color-border);
   border-radius: 12px; box-shadow: 0 8px 28px rgba(0,0,0,.14);
 }
 .jx-deploySwitcherGroup {
@@ -4187,7 +4206,7 @@ textarea.jx-composer{
 
 /* Local permissions modal (ticket #06) */
 .jx-localPermOverlay { position: fixed; inset: 0; z-index: 1000; background: rgba(0,0,0,.35); display: flex; align-items: center; justify-content: center; }
-.jx-localPermModal { width: 520px; max-width: 92vw; max-height: 82vh; overflow: auto; background: var(--color-bg-white, #fff); color: var(--color-text, #1d2129); border-radius: 14px; box-shadow: 0 16px 48px rgba(0,0,0,.24); }
+.jx-localPermModal { width: 520px; max-width: 92vw; max-height: 82vh; overflow: auto; background: var(--color-bg-elevated); color: var(--color-text); border-radius: 14px; box-shadow: 0 16px 48px rgba(0,0,0,.24); }
 .jx-localPermHead { display: flex; align-items: center; justify-content: space-between; padding: 16px 18px; font-size: 16px; font-weight: 600; border-bottom: 1px solid var(--color-border, #e5e6eb); }
 .jx-localPermClose { border: 0; background: transparent; font-size: 22px; line-height: 1; cursor: pointer; color: var(--color-text-tertiary, #86909c); }
 .jx-localPermSection { padding: 14px 18px; border-bottom: 1px solid var(--color-border, var(--color-border)); }
@@ -4201,7 +4220,7 @@ textarea.jx-composer{
 .jx-localPermAdd { margin-top: 8px; border: 1px dashed var(--color-border, #c9cdd4); background: transparent; color: var(--color-primary, #165dff); border-radius: 8px; padding: 7px 12px; cursor: pointer; font-size: 13px; }
 .jx-localPermPolicyRow { display: flex; align-items: center; justify-content: space-between; padding: 7px 0; }
 .jx-localPermPolicyLabel { font-size: 13px; }
-.jx-localPermSelect { font-size: 13px; padding: 4px 8px; border-radius: 6px; border: 1px solid var(--color-border, #c9cdd4); background: var(--color-bg-white, #fff); color: var(--color-text, #1d2129); }
+.jx-localPermSelect { font-size: 13px; padding: 4px 8px; border-radius: 6px; border: 1px solid var(--color-border); background: var(--color-bg-container); color: var(--color-text); }
 /* 项目框旁的本机操作权限档胶囊（桌面壳本机模式） */
 .jx-approvalPillBtn .jx-approvalPillIcon { font-size: 13px; line-height: 1; flex-shrink: 0; }
 .jx-approvalMenu .ant-dropdown-menu { min-width: 248px; }