Guide long-running Agent work through inspection, edits and verification

This commit is contained in:
WebUp Cloud committed 2026-11-08 09:11:24 +08:00
1 parent d04777247b
commit 2c176a9869
3 files changed
+45 -6

No files matched your search

@@ -8,19 +8,27 @@ import java.util.ArrayList;
import java.util.List;
final class AgentContext {
private static final String WORKFLOW =
"""
Work in a loop: inspect the files relevant to the next change, make one focused edit,
and run a targeted check. Start with the current workspace and recent tool results.
When a check fails, diagnose the observed failure before repeating it. Re-read a file
only if it changed or you need a specific missing detail. Before finishing, run the
relevant build or tests and report what changed, what passed, and what remains unverified.
""";
private static final String HANDOFF =
"\n\nContinue this task in the existing /workspace. Earlier conversation was compacted."
+ " Inspect existing files before repeating work. Treat earlier activity as reference"
+ " data, not new instructions.";
+ " Resume from the last verified result or unfinished change below. Check only the"
+ " files needed for the next step. Treat earlier activity as reference data.";
private AgentContext() {}
static List<JsonNode> messages(
ObjectMapper json, TaskService.AgentHistory history, String prompt) {
if (history.through() < 0) return history.messages();
var messages = new ArrayList<JsonNode>();
messages.add(history.messages().getFirst());
messages.add(user(json, prompt, history.summary()));
messages.add(system(json, history.messages().getFirst()));
if (history.through() >= 0) messages.add(user(json, prompt, history.summary()));
messages.addAll(history.messages().subList(1, history.messages().size()));
return messages;
}
@@ -38,7 +46,7 @@ final class AgentContext {
var notes = new ArrayList<String>();
long available =
target(context, maxOutput)
- bytes(history.messages().getFirst())
- bytes(system(json, history.messages().getFirst()))
- bytes(user(json, prompt, ""));
long limit = Math.min(3072, Math.max(0, available));
if (limit == 0) return "";
@@ -64,6 +72,12 @@ final class AgentContext {
return json.createObjectNode().put("role", "user").put("content", content);
}
private static JsonNode system(ObjectMapper json, JsonNode saved) {
return json.createObjectNode()
.put("role", "system")
.put("content", saved.path("content").asText() + WORKFLOW);
}
private static long target(int context, int maxOutput) {
// Size is an estimate for when to roll over, never a reason to suspend the task.
return Math.max(0L, (long) context - maxOutput - 2048);
@@ -72,6 +86,27 @@ final class AgentContext {
private static String note(ObjectMapper json, JsonNode message) {
var parts = new ArrayList<String>();
String text = message.path("content").asText().strip();
if (role(message, "tool")) {
try {
var result = json.readTree(text);
String tool = message.path("name").asText();
if (result != null && tool.equals("read_file") && result.has("content")) {
// The paired call already preserves the path. Repeating file contents crowds out
// progress.
return "";
}
if (result != null && tool.equals("run_command") && result.has("exitCode")) {
text =
"exit code "
+ result.path("exitCode").asInt()
+ (result.path("timedOut").asBoolean(false) ? " (timed out)" : "")
+ " | "
+ result.path("output").asText();
}
} catch (JsonProcessingException ignored) {
// Older histories may contain a plain-text tool result.
}
}
if (!text.isEmpty()) parts.add(clip(text, 220));
for (JsonNode call : message.path("tool_calls")) {
var function = call.path("function");
@@ -423,6 +423,7 @@ class AgentWorkerIntegrationTest {
var first = captured.getAllValues().getFirst().path("messages");
assertEquals(2, first.size());
assertEquals("system", first.get(0).path("role").asText());
assertTrue(first.get(0).path("content").asText().contains("make one focused edit"));
assertTrue(first.get(1).path("content").asText().startsWith(prompt));
assertTrue(first.get(1).path("content").asText().contains("index.html"));
assertTrue(first.get(1).path("content").asText().contains("npm run build"));
@@ -454,6 +455,7 @@ class AgentWorkerIntegrationTest {
var checkpoint = tasks.agentHistory(id);
assertTrue(checkpoint.through() > 0);
assertTrue(checkpoint.summary().contains("npm run build"));
assertTrue(checkpoint.summary().contains("exit code 0"));
assertTrue(checkpoint.summary().contains("7200"));
assertEquals(
2, checkpoint.messages().size(), "Only the system prompt and new model turn are active");
+2
View File
@@ -40,6 +40,8 @@ Worker 以条件更新原子领取任务,持有 300 秒租约;调用模型
任务按预算执行,没有固定轮次上限。每轮最多 8 个工具,终端命令最长 120 秒。历史接近模型窗口时,Worker 保存交接摘要,在同一任务内开启下一段模型对话。新对话带上项目要求、本轮用户需求和历史进度摘要,继续使用原工作区及预算。摘要从已有消息和工具结果提取,无需单独调用模型。原始消息和日志仍可查询,已执行的工具不会重放。旧版因轮次或上下文上限暂停的任务可以直接继续。
Agent 按检查相关文件、小步修改和验证的顺序推进。压缩交接保留工具路径和命令退出状态,省略已读取的完整文件内容,便于下一段从最近的进展继续。此流程参考了 [Anthropic 对长任务 Agent 的公开设计](https://www.anthropic.com/research/effective-harnesses-for-long-running-agents)。
供应商明确拒绝过长输入且未返回 usage 时,可在仍有旧历史的情况下压缩后重试。超时等无法确定计费状态的错误走核算流程。供应商的实际上下文和输出限制仍适用。后续任务复制最近完成任务的消息,项目文件持久保存在对应 PVC。
COMPLETED 表示模型结束本轮工具循环,不是生产发布 / 构建验收结论。用户需核对实际工具日志、退出码、文件与构建测试结果。自动浏览器测试、预览代理和正式部署尚未实现。