工具观察原样进 prompt：去掉按 token 预算的脱敏与截断

- 删除 `bound_observation_for_prompt` 与 `truncate_to_estimated_tokens` - `prepare_game_creator_agent_runtime_prompt_history` 去掉 `tool_output_token_limit` 参数，观察原样进 prompt - 调用处同步去掉该参数：`provider_request_builders`、`tests/runtime_state` - 删除只验证旧有界行为的 `tool_output_is_bounded_by_token_limit` 测试 - 脱敏仍只发生在写审计/持久化时（`sanitize_game_creator_agent_runtime_context_observations_for_storage`）
2026-10-01 18:53:40 +08:00
parent 214055e319
commit cf6ed72f34
4 changed files with 25 additions and 126 deletions
@@ -708,7 +708,6 @@ fn build_game_creator_background_agent_context(
        &session_id,
        run_id,
        observations,
-        llm.tool_output_token_limit,
    )?;
    let repository_context = build_repository_startup_context_at(root)?;
    let repository_prompt = render_repository_startup_context_for_prompt(&repository_context);
@@ -21,7 +21,6 @@ fn agent_runtime_preview_infrastructure_failure_kind(error: &str) -> Option<&'st
 }

 fn agent_runtime_preview_infrastructure_observation(
-    root: &Path,
    revision: u64,
    failure_kind: &str,
    error: &str,
@@ -30,7 +29,7 @@ fn agent_runtime_preview_infrastructure_observation(
        "errorKind": AGENT_RUNTIME_PREVIEW_INFRASTRUCTURE_ERROR_KIND,
        "failureKind": failure_kind,
        "revision": revision,
-        "diagnostic": redact_agent_runtime_project_paths(root, error, 500),
+        "diagnostic": truncate_agent_runtime_text(error, 500),
    }))
    .ok();
    AgentRuntimeToolObservation {
@@ -93,7 +92,7 @@ pub(crate) fn observe_agent_runtime_preview_start(
        Err(error) => AgentRuntimeToolObservation {
            tool: "preview.start".to_string(),
            status: "failed".to_string(),
-            summary: redact_agent_runtime_project_paths(root, &error, 240),
+            summary: truncate_agent_runtime_text(&error, 240),
            detail: None,
        },
    }
@@ -158,7 +157,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: sanitize_agent_runtime_text(
+                summary: truncate_agent_runtime_text(
                    &format!("preview.validate 输入无效：{error}"),
                    240,
                ),
@@ -180,7 +179,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: redact_agent_runtime_project_paths(root, &error, 240),
+                summary: truncate_agent_runtime_text(&error, 240),
                detail: None,
            };
        }
@@ -194,7 +193,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                    status: "failed".to_string(),
                    summary: "preview.validate 执行身份或自主构建完成合同不可用，未执行浏览器试玩"
                        .to_string(),
-                    detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                    detail: Some(truncate_agent_runtime_text(&error, 500)),
                };
            }
        };
@@ -223,7 +222,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
                summary: "旧自主试玩回执无法失效，未执行新的浏览器试玩".to_string(),
-                detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                detail: Some(truncate_agent_runtime_text(&error, 500)),
            };
        }
    }
@@ -233,7 +232,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: redact_agent_runtime_project_paths(root, &error, 240),
+                summary: truncate_agent_runtime_text(&error, 240),
                detail: None,
            };
        }
@@ -254,7 +253,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                return AgentRuntimeToolObservation {
                    tool: "preview.validate".to_string(),
                    status: "failed".to_string(),
-                    summary: redact_agent_runtime_project_paths(root, &error, 240),
+                    summary: truncate_agent_runtime_text(&error, 240),
                    detail: None,
                };
            }
@@ -287,7 +286,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: redact_agent_runtime_project_paths(root, &error, 240),
+                summary: truncate_agent_runtime_text(&error, 240),
                detail: None,
            };
        }
@@ -310,7 +309,6 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
        Err(error) => {
            if let Some(failure_kind) = agent_runtime_preview_infrastructure_failure_kind(&error) {
                return agent_runtime_preview_infrastructure_observation(
-                    root,
                    revision_before.revision,
                    failure_kind,
                    &error,
@@ -319,7 +317,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: redact_agent_runtime_project_paths(root, &error, 240),
+                summary: truncate_agent_runtime_text(&error, 240),
                detail: None,
            };
        }
@@ -338,7 +336,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
            return AgentRuntimeToolObservation {
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
-                summary: redact_agent_runtime_project_paths(root, &error, 240),
+                summary: truncate_agent_runtime_text(&error, 240),
                detail: None,
            };
        }
@@ -382,7 +380,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                tool: "preview.validate".to_string(),
                status: "failed".to_string(),
                summary: "浏览器验证未通过，且当前验证凭证无法安全失效".to_string(),
-                detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                detail: Some(truncate_agent_runtime_text(&error, 500)),
            };
        }
    }
@@ -414,7 +412,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                        tool: "preview.validate".to_string(),
                        status: "failed".to_string(),
                        summary: "浏览器试玩已返回，但自主试玩回执无法形成".to_string(),
-                        detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                        detail: Some(truncate_agent_runtime_text(&error, 500)),
                    };
                }
            };
@@ -441,7 +439,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                        tool: "preview.validate".to_string(),
                        status: "failed".to_string(),
                        summary: "浏览器试玩已通过，但持久回执无法验证".to_string(),
-                        detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                        detail: Some(truncate_agent_runtime_text(&error, 500)),
                    };
                }
            };
@@ -457,7 +455,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                                tool: "preview.validate".to_string(),
                                status: "failed".to_string(),
                                summary: "浏览器试玩已通过，但首个可玩版本暂时无法登记".to_string(),
-                                detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                                detail: Some(truncate_agent_runtime_text(&error, 500)),
                            };
                        }
                    };
@@ -481,7 +479,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                            status: "failed".to_string(),
                            summary: "浏览器试玩已通过，但登记版本前无法复核项目 revision"
                                .to_string(),
-                            detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                            detail: Some(truncate_agent_runtime_text(&error, 500)),
                        };
                    }
                };
@@ -492,7 +490,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                        tool: "preview.validate".to_string(),
                        status: "failed".to_string(),
                        summary: "浏览器试玩已通过，但首个可玩版本无法登记".to_string(),
-                        detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                        detail: Some(truncate_agent_runtime_text(&error, 500)),
                    };
                }
            }
@@ -507,7 +505,7 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
                        tool: "preview.validate".to_string(),
                        status: "failed".to_string(),
                        summary: "浏览器验证已通过，但失败试玩凭证无法安全清除".to_string(),
-                        detail: Some(redact_agent_runtime_project_paths(root, &error, 500)),
+                        detail: Some(truncate_agent_runtime_text(&error, 500)),
                    };
                }
            }
@@ -562,13 +560,13 @@ pub(crate) async fn observe_agent_runtime_preview_validate(
        return AgentRuntimeToolObservation {
            tool: "preview.validate".to_string(),
            status: "failed".to_string(),
-            summary: redact_agent_runtime_project_paths(root, &error, 240),
+            summary: truncate_agent_runtime_text(&error, 240),
            detail: None,
        };
    }
    let detail = serde_json::to_string(&detail_value)
        .ok()
-        .map(|value| redact_agent_runtime_project_paths(root, &value, 3_600));
+        .map(|value| truncate_agent_runtime_text(&value, 3_600));
    AgentRuntimeToolObservation {
        tool: "preview.validate".to_string(),
        status: if result.passed { "ok" } else { "failed" }.to_string(),
@@ -592,12 +590,8 @@ mod tests {
            agent_runtime_preview_infrastructure_failure_kind(error),
            Some("browser-launch-failed")
        );
-        let observation = agent_runtime_preview_infrastructure_observation(
-            Path::new("/project"),
-            19,
-            "browser-launch-failed",
-            error,
-        );
+        let observation =
+            agent_runtime_preview_infrastructure_observation(19, "browser-launch-failed", error);
        assert_eq!(observation.status, "blocked");
        assert_eq!(
            agent_runtime_preview_infrastructure_blocker(&observation).as_deref(),
@@ -188,69 +188,6 @@ pub(crate) fn validate_game_creator_llm_request_context_budget(
    Ok(())
 }

-fn truncate_to_estimated_tokens(value: &str, token_limit: u64) -> String {
-    if token_limit == 0 {
-        return String::new();
-    }
-    let max_bytes = token_limit
-        .saturating_mul(AGENT_RUNTIME_CONTEXT_TOKEN_ESTIMATE_BYTES_PER_TOKEN)
-        .min(usize::MAX as u64) as usize;
-    if value.len() <= max_bytes {
-        return value.to_string();
-    }
-    let suffix = "\n...[tool output truncated by token budget]";
-    if max_bytes <= suffix.len() {
-        let mut boundary = max_bytes.min(value.len());
-        while boundary > 0 && !value.is_char_boundary(boundary) {
-            boundary -= 1;
-        }
-        return value[..boundary].to_string();
-    }
-    let retained_bytes = max_bytes.saturating_sub(suffix.len());
-    let mut boundary = retained_bytes.min(value.len());
-    while boundary > 0 && !value.is_char_boundary(boundary) {
-        boundary -= 1;
-    }
-    format!("{}{}", &value[..boundary], suffix)
-}
-
-fn bound_observation_for_prompt(
-    root: &Path,
-    observation: &AgentRuntimeToolObservation,
-    token_limit: u64,
-) -> AgentRuntimeToolObservation {
-    let mut bounded = sanitize_agent_runtime_context_observation(root, observation);
-    let summary_budget = token_limit.min(1_024).max(1);
-    bounded.summary = truncate_to_estimated_tokens(&bounded.summary, summary_budget);
-    let detail_budget = token_limit.saturating_sub(summary_budget).max(1);
-    bounded.detail = bounded
-        .detail
-        .as_deref()
-        .map(|detail| truncate_to_estimated_tokens(detail, detail_budget))
-        .filter(|detail| !detail.trim().is_empty());
-    while estimate_serialized_bytes_as_tokens(
-        bounded.summary.len() + bounded.detail.as_deref().map(str::len).unwrap_or_default(),
-    ) > token_limit
-    {
-        if let Some(detail) = bounded.detail.as_deref() {
-            let chars = detail.chars().count();
-            if chars > 32 {
-                bounded.detail = Some(detail.chars().take(chars / 2).collect());
-                continue;
-            }
-            bounded.detail = None;
-            continue;
-        }
-        let chars = bounded.summary.chars().count();
-        if chars <= 8 {
-            bounded.summary.clear();
-            continue;
-        }
-        bounded.summary = bounded.summary.chars().take(chars / 2).collect();
-    }
-    bounded
-}
-
 pub(crate) fn sanitize_game_creator_agent_runtime_context_observations_for_storage(
    root: &Path,
    observations: &[AgentRuntimeToolObservation],
@@ -457,7 +394,6 @@ pub(crate) fn prepare_game_creator_agent_runtime_prompt_history(
    session_id: &str,
    run_id: &str,
    observations: &[AgentRuntimeToolObservation],
-    tool_output_token_limit: u64,
 ) -> Result<AgentRuntimePromptHistory, String> {
    let (agent_messages, project_messages, sidecar) =
        prompt_history_sources(root, agent_id, session_id, run_id, observations)?;
@@ -499,10 +435,9 @@ pub(crate) fn prepare_game_creator_agent_runtime_prompt_history(
            project_tail = project_tail,
        ));
    }
-    let observations = observations[observation_start..]
-        .iter()
-        .map(|observation| bound_observation_for_prompt(root, observation, tool_output_token_limit))
-        .collect();
+    // 工具观察原样进 prompt：给模型看的内容不脱敏、不截断；脱敏只发生在写审计/持久化时
+    // （见 `sanitize_game_creator_agent_runtime_context_observations_for_storage`）。
+    let observations = observations[observation_start..].to_vec();
    Ok(AgentRuntimePromptHistory {
        context: sections.join("\n\n"),
        observations,
@@ -985,30 +920,6 @@ mod tests {
        );
    }

-    #[test]
-    fn tool_output_is_bounded_by_token_limit() {
-        let root = PathBuf::from("/tmp/context-compaction-token-bound");
-        let observation = AgentRuntimeToolObservation {
-            tool: "file.read".to_string(),
-            status: "ok".to_string(),
-            summary: "s".repeat(8_000),
-            detail: Some("d".repeat(20_000)),
-        };
-        let bounded = bound_observation_for_prompt(&root, &observation, 256);
-        let tokens = estimate_serialized_bytes_as_tokens(
-            bounded.summary.len() + bounded.detail.as_deref().map(str::len).unwrap_or_default(),
-        );
-        assert!(tokens <= 256);
-
-        let tiny = bound_observation_for_prompt(&root, &observation, 1);
-        let tiny_tokens = estimate_serialized_bytes_as_tokens(
-            tiny.summary.len() + tiny.detail.as_deref().map(str::len).unwrap_or_default(),
-        );
-        assert!(tiny_tokens <= 1);
-        assert_eq!(tiny.tool, "file.read");
-        assert_eq!(tiny.status, "ok");
-    }
-
    #[test]
    fn compaction_keeps_recent_tails_and_advances_only_for_new_source() {
        let project = context_compaction_test_project("tail-and-revision");
@@ -1038,7 +949,6 @@ mod tests {
            session_id,
            run_id,
            &observations,
-            1_000,
        )
        .expect("prepare compacted prompt history");
        assert!(history.context.contains("第一版安全摘要"));
@@ -1119,7 +1029,6 @@ mod tests {
            session_id,
            run_id,
            &tampered_observations,
-            1_000,
        )
        .expect_err("tampered observation prefix must fail");
        assert!(observation_error.contains("observation 前缀发生漂移"));
@@ -1148,7 +1057,6 @@ mod tests {
            session_id,
            run_id,
            &observations,
-            1_000,
        )
        .expect_err("tampered conversation prefix must fail");
        assert!(conversation_error.contains("对话前缀发生漂移"));
@@ -1327,7 +1235,6 @@ mod tests {
            session_id,
            run_id,
            &observations,
-            1_000,
        )
        .expect("prepare supervisor prompt history");
        assert!(!history.context.contains("LEGACY_PROJECT_MARKER_0"));
@@ -3496,7 +3496,6 @@ fn supervisor_background_task_persists_one_public_start_status_before_execution(
        &queued.state.session_id,
        run_id,
        &[],
-        1_000,
    )
    .expect("render Runtime prompt history");
    assert!(runtime_prompt_history