Translate repair system prompt to English

tests/test_jinja_validation.py

@@ -47,3 +47,39 @@ def test_condition_field_literal_wrapped_as_jinja():
     assert errors == []
     assert summary.get("applied") in {False, None}
     assert normalized["nodes"][0]["params"]["expression"] == "{{ result_of.some_loop.exports.items | length > 0 }}"
+
+
+def test_malformed_conditional_literal_is_repaired():
+    workflow = _workflow({"status": "abnormal' if loop.item.temperature > 38 else 'normal"})
+
+    normalized, summary, errors = normalize_params_to_jinja(workflow)
+
+    assert errors == []
+    assert summary["applied"] is True
+    assert summary["replacements"] == [
+        {
+            "path": "params.status",
+            "from": "abnormal' if loop.item.temperature > 38 else 'normal",
+            "to": "{{ 'abnormal' if loop.item.temperature > 38 else 'normal' }}",
+        }
+    ]
+
+
+def test_malformed_conditional_literal_with_filter_is_repaired():
+    workflow = _workflow(
+        {
+            "status": "alert' if (loop.item.temperature is not none and (loop.item.temperature | float) > 38.0) else 'normal",
+        }
+    )
+
+    normalized, summary, errors = normalize_params_to_jinja(workflow)
+
+    assert errors == []
+    assert summary["applied"] is True
+    assert summary["replacements"] == [
+        {
+            "path": "params.status",
+            "from": "alert' if (loop.item.temperature is not none and (loop.item.temperature | float) > 38.0) else 'normal",
+            "to": "{{ 'alert' if (loop.item.temperature is not none and (loop.item.temperature | float) > 38.0) else 'normal' }}",
+        }
+    ]

velvetflow/planner/repair.py

@@ -366,20 +366,20 @@ def _make_error_key(e: ValidationError) -> str:
 
 
 _ERROR_TYPE_PROMPTS: Dict[str, str] = {
-    "MISSING_REQUIRED_PARAM": "补齐 params.<field>，优先绑定上游符合 arg_schema 的输出或使用 schema 默认值。示例：如果 action 需要 params.prompt 且缺失，可绑定上游文案生成节点的 result.output。",
-    "UNKNOWN_ACTION_ID": "替换为 Action Registry 中存在的 action_id，保持节点含义最接近并同步校正 params。示例：将未知的 action_id 改为 registry 中的 text.generate，并保留 prompt/temperature 等参数。",
-    "UNKNOWN_PARAM": "移除或重命名 params 中未在 arg_schema 声明的字段，使其与 schema 对齐。示例：action schema 仅接受 prompt/temperature，应删除意外出现的 max_token 字段。",
-    "DISCONNECTED_GRAPH": "连接无入度/出度节点，确保每个节点都位于可达路径上。示例：为孤立的 summary 节点添加来自前置生成节点的边，并把结果导向下游汇总节点。",
-    "INVALID_EDGE": "修复边的 from/to 以引用存在的节点，并保持条件字段合法。示例：若 edge.to 指向不存在的 node-3，可改为实际存在的 reviewer 节点。",
-    "SCHEMA_MISMATCH": "调整绑定字段或聚合方式，使类型与上游 output_schema/arg_schema 兼容。示例：如果上游输出是 list 而当前节点期望 string，可改为 join 列表或选择单个元素。",
-    "INVALID_SCHEMA": "按 DSL 结构补齐/修正字段类型（nodes/edges/params），避免 JSON 结构缺失。示例：在 workflow 缺少 edges 字段时补空数组，并确保 nodes 为列表。",
-    "INVALID_LOOP_BODY": "补齐 loop.body_subgraph.nodes（至少一个 action），并确保 exports 使用 {key: Jinja表达式} 结构引用 body_subgraph 节点字段。",
-    "STATIC_RULES_SUMMARY": "遵循静态规则摘要中的提示逐项修复，优先处理连通性和 schema 不匹配。示例：按摘要要求先修复未连接的边，再补齐缺失的必填参数。",
-    "EMPTY_PARAM_VALUE": "为空的参数应填入非空值或绑定，避免空字符串/空对象。示例：将空的 params.prompt 改为具体提示词或绑定上游结果。",
-    "EMPTY_PARAMS": "params 不能是空对象，为必填字段提供值或引用，并使用工具完成不确定的填充。示例：若节点类型为 action 但 params 为 {}，根据 arg_schema 填入 prompt/template 等必填字段。",
-    "SELF_REFERENCE": "移除节点对自身 result_of.<node> 的引用，改用上游输出或拆分节点并通过工具应用修改。示例：若 node-a.params.input 引用 result_of.node-a，可改为 result_of.previous-node。",
-    "SYNTAX_ERROR": "修正 DSL 语法（括号、逗号、引号等），使解析器能生成 AST。示例：补全缺失的引号或逗号，确保 JSON 能被解析。",
-    "GRAMMAR_VIOLATION": "遵循文法约束调整节点/字段位置，按期望的 token/产生式重新组织。示例：将误放在 edges 中的节点定义移回 nodes 列表。",
+    "MISSING_REQUIRED_PARAM": "Fill in params.<field>; prefer binding upstream outputs that match arg_schema or use schema defaults. Example: if an action requires params.prompt but it is missing, bind result.output from an upstream copywriting node.",
+    "UNKNOWN_ACTION_ID": "Replace with an action_id that exists in the Action Registry, keeping the node's intent closest and fixing params accordingly. Example: change an unknown action_id to text.generate in the registry and keep parameters such as prompt/temperature.",
+    "UNKNOWN_PARAM": "Remove or rename params fields not declared in arg_schema so they align with the schema. Example: if the action schema only accepts prompt/temperature, delete an unexpected max_token field.",
+    "DISCONNECTED_GRAPH": "Connect nodes with no in-degree/out-degree so every node sits on a reachable path. Example: add an edge from a prior generation node to an isolated summary node and route its result to downstream aggregation nodes.",
+    "INVALID_EDGE": "Fix edge.from/edge.to to reference existing nodes and keep condition fields valid. Example: if edge.to points to missing node-3, change it to an existing reviewer node.",
+    "SCHEMA_MISMATCH": "Adjust bound fields or aggregation so types are compatible with upstream output_schema/arg_schema. Example: if upstream output is a list but the current node expects a string, join the list or pick a single element.",
+    "INVALID_SCHEMA": "Fill or correct field types per the DSL structure (nodes/edges/params) to avoid missing JSON pieces. Example: if the workflow lacks an edges field, add an empty array and ensure nodes is a list.",
+    "INVALID_LOOP_BODY": "Fill loop.body_subgraph.nodes (at least one action) and ensure exports use the {key: Jinja expression} shape referencing body_subgraph node fields.",
+    "STATIC_RULES_SUMMARY": "Follow the static rules summary step by step, prioritizing connectivity and schema mismatches. Example: fix disconnected edges first per the summary, then fill missing required params.",
+    "EMPTY_PARAM_VALUE": "Populate empty parameters with non-empty values or bindings; avoid empty strings/objects. Example: change an empty params.prompt to a concrete prompt or bind it to an upstream result.",
+    "EMPTY_PARAMS": "params cannot be an empty object—provide values or references for required fields and use tools to fill uncertain parts. Example: if a node type is action but params is {}, fill required fields like prompt/template according to arg_schema.",
+    "SELF_REFERENCE": "Remove a node's reference to its own result_of.<node>; use upstream outputs instead, or split the node and apply tools. Example: if node-a.params.input references result_of.node-a, change it to result_of.previous-node.",
+    "SYNTAX_ERROR": "Fix DSL syntax (brackets, commas, quotes, etc.) so the parser can build the AST. Example: add missing quotes or commas to make the JSON parseable.",
+    "GRAMMAR_VIOLATION": "Rearrange nodes/fields according to grammar constraints, respecting expected tokens/productions. Example: move a node definition mistakenly placed in edges back to the nodes list.",
 }
 
 
@@ -569,69 +569,68 @@ def repair_workflow_with_llm(
         }
 
     system_prompt = """
-你是一个工作流修复助手。
-【Workflow DSL 语法与语义（务必遵守）】
-- workflow = {workflow_name, description, nodes: []}，只能返回合法 JSON（edges 会由系统基于节点绑定自动推导，不需要生成）。
-- node 基本结构：{id, type, display_name, params, action_id?, out_params_schema?, loop/subgraph/branches?}。
-  type 仅允许 action/condition/loop/parallel。无需 start/end/exit 节点。
-  action 节点必须填写 action_id（来自动作库）与 params；只有 action 节点允许 out_params_schema。
-  condition 节点需包含返回布尔值的 params.expression（合法 Jinja 表达式），以及 true_to_node/false_to_node（字符串或 null）。
-  loop 节点包含 loop_kind/iter/source/body_subgraph/exports，exports 的 value 必须引用 body_subgraph 节点字段（如 {{ result_of.node.field }}），循环外部只能引用 exports.<key>，body_subgraph 仅包含 nodes 数组。
-  loop.body_subgraph 内不需要也不允许显式声明 edges、entry 或 exit 节点，如发现请直接删除。
-  - params 必须直接使用 Jinja 表达式引用上游结果（如 {{ result_of.<node_id>.<field_path> }} 或 {{ loop.item.xxx }}），不得输出对象式绑定。
-    <node_id> 必须存在且字段需与上游 output_schema 或 loop.exports 对齐。
-  当前有一个 workflow JSON 和一组结构化校验错误 validation_errors。
-  validation_errors 是 JSON 数组，元素包含 code/node_id/field/message。
-  这些错误来自：
-  - action 参数缺失或不符合 arg_schema
-  - condition 条件不完整
-  - source 路径引用了不存在的节点
-  - source 路径与上游 action 的 output_schema 不匹配
-  - source 指向的数组元素 schema 中不存在某个字段
-
-  - previous_failed_attempts 会记录同一错误的历史修复尝试，请避免重复失败的方法，尝试不同的修复路径。
-
-  - workflow 结构不符合 DSL schema（例如节点 type 非法）
-
-  总体目标：在“尽量不改变工作流整体结构”的前提下，修复这些错误，使 workflow 通过静态校验。
-
-  具体要求（很重要，请严格遵守）：
- 1. 结构保持稳定：
-    - 不要增加或删除节点；
-    - edges 会由系统根据节点引用自动推导，不需要手动增删或调整 condition。
-
- 2. action 节点修复优先级：
-    - 首先根据 action_schemas[action_id].arg_schema 补齐 params 里缺失的必填字段，或修正错误类型；
-    - 如果 action_id 本身是合法的（存在于 action_schemas 中），优先“修 params”，不要改 action_id；
-    - 只有当 validation_errors 明确指出 action_id 不存在时，才考虑把 action_id 改成一个更合理的候选，
-      并同步更新该节点的 params 使之符合新的 arg_schema。
-
- 3. condition 节点修复：
-    - 确保 params.expression 存在且是合法的 Jinja 布尔表达式；
-    - expression 应直接编写判断逻辑，引用上游结果请使用 {{ result_of.<node>.<field_path> }} 或 {{ loop.item.xxx }}；
-    - 不要生成 kind/field/op/value 等字段，所有条件均需融入 expression。
-
- 4. loop 节点：确保 loop_kind（for_each/while）和 source 字段合法，并使用 exports 暴露 body 结果（表达式需引用 body_subgraph 节点字段），下游只能引用 result_of.<loop_id>.exports.<key>。
- 5. parallel 节点：branches 必须是非空数组。
-
-6. 参数绑定修复：
-   - 仅输出 Jinja 表达式或字面量，禁止回退到对象式绑定；
-   - 聚合/过滤/拼接逻辑请直接写在 Jinja 表达式或过滤器里，并确保字段路径与上游 schema 对齐；
-   - 常见错误：下游引用缺少 exports 段，请补齐 result_of.<loop_id>.exports.<key>。
-
- 7. 修改范围尽量最小化：
-    - 当有多种修复方式时，优先选择改动最小、语义最接近原意的方案（如只改一个字段名，而不是重写整个 params）。
-    - 当 validation_error_summary 提供 Schema 提示或路径信息时，优先按提示矫正字段类型/结构，避免多轮重复犯错。
-
- 8. 修复回合有限且必须闭环：
-    - 系统会在有限轮次内重新校验；如果你忽略任何一条 validation_error，流程将直接进入下一轮甚至终止。
-    - 请逐条对照 validation_errors，把所有问题修到为 0 再输出结果，避免留存隐患。
-
- 9. 输出要求：
-    - 保持顶层结构：workflow_name/description/nodes 不变（仅节点内部内容可调整，edges 由系统推导）；
-    - 节点的 id/type 不变；
-    - 必须使用工具链完成修改：通过修复工具调整 workflow 并直接使用最新结果，避免自然语言或 Markdown 代码块。
- 10. 可用工具：当你需要结构化修改时，优先调用提供的工具（无 LLM 依赖、结果确定），包含 fix_loop_body_references/fill_action_required_params/update_node_field/normalize_binding_paths/replace_reference_paths/drop_invalid_references。
+You are a workflow repair assistant.
+[Workflow DSL syntax and semantics (must follow)]
+- workflow = {workflow_name, description, nodes: []}. Only return valid JSON (edges will be auto-inferred by the system from node bindings; do not generate them).
+- Node base structure: {id, type, display_name, params, action_id?, out_params_schema?, loop/subgraph/branches?}.
+  type only allows action/condition/loop/parallel. No need for start/end/exit nodes.
+  action nodes must include action_id (from the action registry) and params; only action nodes may have out_params_schema.
+  condition nodes must include params.expression that returns a boolean (valid Jinja expression) and true_to_node/false_to_node (string or null).
+  loop nodes include loop_kind/iter/source/body_subgraph/exports. The value of exports must reference fields of nodes inside body_subgraph (e.g., {{ result_of.node.field }}). Outside the loop, only exports.<key> can be referenced. body_subgraph only contains a nodes array.
+  loop.body_subgraph does not need and must not explicitly declare edges, entry, or exit nodes—remove them if present.
+  - params must directly use Jinja expressions to reference upstream results (e.g., {{ result_of.<node_id>.<field_path> }} or {{ loop.item.xxx }}); do not output object-style bindings.
+    <node_id> must exist and the fields must align with upstream output_schema or loop.exports.
+  You have a workflow JSON and a set of structured validation_errors.
+  validation_errors is a JSON array whose elements contain code/node_id/field/message.
+  These errors come from:
+  - Missing action parameters or mismatches with arg_schema
+  - Incomplete condition definitions
+  - source paths referencing nonexistent nodes
+  - source paths not matching the upstream action's output_schema
+  - Fields missing from the array element schema referenced by source
+
+  - previous_failed_attempts records past repair attempts for the same error. Avoid repeating failed methods and try different fix paths.
+
+  - workflow structure not conforming to DSL schema (e.g., invalid node type)
+
+  Overall goal: repair these errors so the workflow passes static validation while “changing the overall workflow structure as little as possible.”
+
+  Specific requirements (very important, follow strictly):
+ 1. Keep the structure stable:
+    - Do not add or delete nodes;
+    - edges will be auto-derived by the system based on node references. Do not manually add/remove or adjust conditions.
+
+ 2. Action node repair priority:
+    - First, fill missing required fields in params or fix incorrect types based on action_schemas[action_id].arg_schema;
+    - If the action_id itself is valid (exists in action_schemas), prioritize “fixing params” instead of changing action_id;
+    - Only when validation_errors explicitly indicate the action_id does not exist should you consider changing action_id to a more reasonable candidate and simultaneously update that node’s params to match the new arg_schema.
+
+ 3. Condition node repair:
+    - Ensure params.expression exists and is a valid Jinja boolean expression;
+    - expression should directly express the logic. To reference upstream results use {{ result_of.<node>.<field_path> }} or {{ loop.item.xxx }};
+    - Do not generate kind/field/op/value fields; all conditions must be within expression.
+
+ 4. loop nodes: ensure loop_kind (for_each/while) and source are valid, and expose body results via exports (expressions must reference body_subgraph node fields). Downstream may only reference result_of.<loop_id>.exports.<key>.
+ 5. parallel nodes: branches must be a non-empty array.
+
+6. Parameter binding repairs:
+   - Only output Jinja expressions or literals; do not fall back to object-style bindings;
+   - Put aggregation/filtering/concatenation logic directly in Jinja expressions or filters and ensure field paths align with upstream schemas;
+   - Common error: downstream references missing the exports segment—add result_of.<loop_id>.exports.<key>.
+
+ 7. Minimize the scope of changes:
+    - When multiple fixes exist, prefer the minimal change closest to the original intent (e.g., rename a single field instead of rewriting the entire params).
+    - When validation_error_summary provides schema hints or path information, prioritize adjusting field types/structures per the hint to avoid repeating mistakes.
+
+ 8. Repair rounds are limited and must close the loop:
+    - The system revalidates within limited iterations; if you ignore any validation_error, the process may proceed to the next round or terminate.
+    - Address every item in validation_errors until there are zero issues before outputting results to avoid leaving problems behind.
+
+ 9. Output requirements:
+    - Keep the top-level structure workflow_name/description/nodes unchanged (only adjust node internals; edges are inferred by the system);
+    - Node id/type must remain unchanged;
+    - You must use the toolchain to apply modifications: adjust the workflow through repair tools and directly use the latest result—avoid natural language or Markdown code blocks.
+ 10. Available tools: when you need structured modifications, prioritize the provided tools (deterministic, no LLM dependency): fix_loop_body_references/fill_action_required_params/update_node_field/normalize_binding_paths/replace_reference_paths/drop_invalid_references.
 """
 
     working_workflow: Dict[str, Any] = broken_workflow