feat: make model generation budgets configurable

benchmarks/harbor/README.md

@@ -45,6 +45,14 @@ container's current directory, and saves combined stdout/stderr to
 `/logs/agent/nanopycodeagent.txt`. It uses the CLI's 50-turn default; override
 that with `--agent-kwarg max_turns=20`.
 
+For the per-reply generation limit, pass `--agent-kwarg max_tokens=32768`.
+This becomes `--max-tokens 32768` in the container and overrides the forwarded
+`ANTHROPIC_MAX_TOKENS` environment variable. When omitted, the adapter sends no
+token flag, so the installed agent's environment/settings/default applies
+(32768 in the version introducing this option). Older releases require omitting
+the new option. The effective budget is recorded in the startup log and in
+`agent.extra.max_tokens` in the ATIF trajectory.
+
 The adapter also asks the agent to write an ATIF-v1.7 trajectory directly to
 `/logs/agent/trajectory.json`. Harbor collects that file as the trial's native
 ATIF output and backfills prompt, completion, cache-token, and cost totals into

benchmarks/harbor/README.zh-CN.md

@@ -41,6 +41,12 @@ adapter 通过 stdin 发送任务指令，在 task 容器的当前目录中运
 的 stdout/stderr 保存到 `/logs/agent/nanopycodeagent.txt`。它默认沿用 CLI 的 50
 轮限制；可以通过 `--agent-kwarg max_turns=20` 覆盖此设置。
 
+每次回复的生成上限可通过 `--agent-kwarg max_tokens=32768` 指定，它会转换为容器内的
+`--max-tokens 32768`，优先于透传的 `ANTHROPIC_MAX_TOKENS` 环境变量。未指定时，
+adapter 不添加该 flag，沿用已安装 agent 的环境变量／配置文件／默认值（引入该参数的
+版本默认为 32768）。安装旧版本时需省略新参数。生效预算会记录到启动日志和 ATIF
+trajectory 的 `agent.extra.max_tokens`。
+
 adapter 还会要求 agent 将 ATIF-v1.7 trajectory 直接写入
 `/logs/agent/trajectory.json`。Harbor 会把该文件作为 trial 的原生 ATIF 输出采集，
 并将 prompt、completion、cache token 和 cost 汇总回填到 agent result。trajectory

benchmarks/harbor/src/harbor_adapter/adapter.py

@@ -37,7 +37,8 @@ class NanoPyCodeAgent(BaseInstalledAgent):
             cli="--max-turns",
             type="int",
             default=_DEFAULT_MAX_TURNS,
-        )
+        ),
+        CliFlag("max_tokens", cli="--max-tokens", type="int"),
     ]
 
     def __init__(self, *args, git_ref: str | None = None, **kwargs):
@@ -54,6 +55,9 @@ def __init__(self, *args, git_ref: str | None = None, **kwargs):
             raise ValueError("version and git_ref are mutually exclusive")
         self._git_ref = git_ref
         super().__init__(*args, **kwargs)
+        max_tokens = self._resolved_flags.get("max_tokens")
+        if max_tokens is not None and max_tokens < 1:
+            raise ValueError("max_tokens must be a positive integer")
 
     @staticmethod
     @override
@@ -112,6 +116,9 @@ def _runtime_env(self) -> dict[str, str]:
             model = self.model_name.split("/", 1)[-1]
         if model:
             env["ANTHROPIC_MODEL"] = model
+        max_tokens = self._get_env("ANTHROPIC_MAX_TOKENS")
+        if max_tokens is not None:
+            env["ANTHROPIC_MAX_TOKENS"] = max_tokens
         return env
 
     @override

benchmarks/harbor/tests/test_adapter.py

@@ -114,6 +114,7 @@ def test_run_pipes_the_instruction_and_forwards_anthropic_configuration(tmp_path
             "ANTHROPIC_MODEL": "deepseek/deepseek-v4-flash-0731",
         },
         max_turns=20,
+        max_tokens=65536,
     )
     environment = RecordingEnvironment()
 
@@ -124,6 +125,7 @@ def test_run_pipes_the_instruction_and_forwards_anthropic_configuration(tmp_path
     assert instruction not in command
     assert 'printf "%s" "$harbor_nanopycodeagent_instruction_' in command
     assert "nanoPyCodeAgent --max-turns 20" in command
+    assert "--max-tokens 65536" in command
     assert "--trajectory /logs/agent/trajectory.json" in command
     assert command.endswith("2>&1 | tee /logs/agent/nanopycodeagent.txt")
 
@@ -160,6 +162,21 @@ def test_run_normalizes_harbor_provider_configuration_for_the_anthropic_sdk(
     assert run_env["ANTHROPIC_BASE_URL"] == "https://openrouter.example/api"
     assert run_env["ANTHROPIC_MODEL"] == "deepseek/deepseek-v4-flash-0731"
     assert "OPENROUTER_API_KEY" not in run_env
+    assert "--max-tokens" not in environment.calls[-1]["command"]
+
+
+def test_run_forwards_the_environment_budget_without_forcing_a_cli_default(tmp_path):
+    adapter = make_adapter(tmp_path, extra_env={"ANTHROPIC_MAX_TOKENS": "16384"})
+    environment = RecordingEnvironment()
+    asyncio.run(adapter.run("fix it", environment, SimpleNamespace()))
+    assert environment.calls[-1]["env"]["ANTHROPIC_MAX_TOKENS"] == "16384"
+    assert "--max-tokens" not in environment.calls[-1]["command"]
+
+
+@pytest.mark.parametrize("value", [0, -1, True, "bad", 3.5])
+def test_adapter_rejects_invalid_generation_budgets(tmp_path, value):
+    with pytest.raises(ValueError, match="max_tokens"):
+        make_adapter(tmp_path, max_tokens=value)
 
 
 def test_adapter_declares_atif_support_and_populates_complete_context(tmp_path):

docs/changelogs/0.8.x.md

@@ -4,6 +4,12 @@ All notable changes in the **0.8.x** release series are documented here.
 
 ## [Unreleased]
 
+### Added
+- Configure per-reply generation budgets through `--max-tokens` or
+  `ANTHROPIC_MAX_TOKENS`, including the user settings file and Harbor adapter.
+  Record the effective limit in startup output, Event Journals, and ATIF
+  trajectories while retaining compatibility with older journals.
+
 ### Fixed
 - Stop explicitly when a model response reaches `max_tokens`, reporting
   `response_truncated` in Event Journals and ATIF trajectories instead of task
@@ -12,6 +18,7 @@ All notable changes in the **0.8.x** release series are documented here.
   New journals use schema v2, with v1 replay still supported.
 
 ### Changed
+- Increase the default per-reply generation limit from 8192 to 32768 tokens.
 - Moved detailed CLI and configuration guidance out of the bilingual READMEs
   into dedicated English and Chinese user references, keeping the READMEs
   concise while documenting task input, options, exit statuses, trajectories,