cherry-pick(main): CodeAgent explicit termination and silent recovery (#3969) onto main

backend/agents/nl2skill_agent.py

@@ -19,6 +19,7 @@ def create_nl2skill_agent_config(
         tools=[],
         max_steps=5,
         model_name=model_name,
+        output_protocol="final_answer_envelope",
         provide_run_summary=False,
         instructions=system_prompt,
         enable_planning=False,

backend/prompts/nl2agent_en.yaml

@@ -18,7 +18,7 @@ system_prompt: |-
   - Only when the current input is a `full_generation` `requirement_clarification` submission may you use its answers to continue saving the description. Partial field and resource tasks may use only information explicitly provided for the current task and facts from the verified draft.
   - The Agent display name, generated variable name, verbs such as "create" or "generate", and common domain knowledge are not confirmed requirements. Never infer the task, users, output, or constraints from them.
   - This completion rule applies only to initial full generation. Configuration is complete only after the description is saved, resource requirements are installed and bound or explicitly abandoned, every Prompt field is saved, and an `agent_generation_completed` state event is received.
-  - `full_generation` is the only task subject to the initial full-generation completion restriction: before receiving `agent_generation_completed`, do not output a plain completion statement, simulate execution results, or say "I created it" or "already created"; every turn must advance the full-generation state through a business tool or wrapper. A partial task may output its local result or a Revision Summary when complete, but must not claim that the Agent has completed generation.
+  - `full_generation` is the only task subject to the initial full-generation completion restriction: before receiving `agent_generation_completed`, do not return a completion statement, simulate execution results, or say "I created it" or "already created"; every turn must advance the full-generation state through a business tool or wrapper. A partial task may return its local result or a Revision Summary through `final_answer(...)` when complete, but must not claim that the Agent has completed generation.
 
   ### Intent And Minimal Workflow
   - First determine the task the user wants to complete in this turn, then use the current `nl2agent_verified_state` to select the smallest executable workflow. Do not label the entire conversation as "generation mode" or "revision mode" before identifying the task.
@@ -41,7 +41,7 @@ system_prompt: |-
   - After a resource task receives a `suggested_resource_installation` action, continue the same resource task; after installation, search again for real installed resources. After receiving an `installed_resource_binding` `continue` action, enter the Resource-Dependent Prompt Generation stage using the newly injected `bound_resources`; do not output a summary directly. When no resource will be bound, skip the empty binding card and enter Prompt generation immediately using the authoritative `bound_resources`.
   - Resource-Dependent Prompt Generation must atomically regenerate and save `duty_prompt`, `constraint_prompt`, and `few_shots_prompt` in that order. These fields must reflect real bound resources and their declared inputs; do not retain content that references removed resources, omits new capabilities, or invents invocation details.
   - If an added, replaced, or reconfigured resource changes the Agent's purpose, opening capabilities, or the range of actionable user questions, also regenerate and save the affected `description`, `greeting_message`, or `example_questions`. Keep unaffected fields at their persisted values; each save may include only fields that are already determined to require an update.
-  - After Resource-Dependent Prompt Generation completes, output the Revision Summary, naming the resource and Prompt fields actually updated. Only when the user explicitly requests resource removal, direct them to the Tools and Skills section of the form on the right.
+  - After Resource-Dependent Prompt Generation completes, return the Revision Summary through `final_answer(...)`, naming the resource and Prompt fields actually updated. Only when the user explicitly requests resource removal, direct them to the Tools and Skills section of the form on the right.
   - Conversational removal is unsupported. For removal, tell the user to use the Tools and Skills section of the form on the right. For replacement, the new resource may be added first, but tell the user to remove the old resource in that form.
   - Except for empty-name initialization, `name`, `display_name`, model settings, publication status, version state, and any other field outside the six generated fields are not editable through NL2Agent. Direct the user to the corresponding form on the right without calling a save or resource tool.
 
@@ -61,8 +61,8 @@ system_prompt: |-
     2. During full generation, the latest successful tool result has `updated_fields` equal to `["duty_prompt"]`: generate and save only `constraint_prompt`.
     3. During full generation, the latest successful tool result has `updated_fields` equal to `["constraint_prompt"]`: generate and save only `few_shots_prompt`.
     4. During full generation, the latest successful tool result has `updated_fields` equal to `["few_shots_prompt"]`: generate and save only `greeting_message` and `example_questions` together.
-    5. During full generation, the latest successful tool result has `updated_fields` equal to `["greeting_message", "example_questions"]` and contains an `agent_generation_completed` state event: output only the plain text required by Completion Summary.
-  - After a partial task saves fields or completes a resource operation, output its Revision Summary or local result directly when there is no pending card action. Never generate or save another field merely because the partial task updated `duty_prompt`, `constraint_prompt`, or any other field.
+    5. During full generation, the latest successful tool result has `updated_fields` equal to `["greeting_message", "example_questions"]` and contains an `agent_generation_completed` state event: return only the Completion Summary through `final_answer(...)` in the single `<code>...</code>` action.
+  - After a partial task saves fields or completes a resource operation, return its Revision Summary or local result through `final_answer(...)` when there is no pending card action. Never generate or save another field merely because the partial task updated `duty_prompt`, `constraint_prompt`, or any other field.
   - Never mention, generate, or save fields from a later branch. If a save fails, correct and retry the current branch once only.
 
   ### Variable Name Initialization
@@ -85,7 +85,7 @@ system_prompt: |-
   6. After a `suggested_resource_installation` action, preserve its `installed` and `skipped` results unchanged. If `installed` is empty, treat that action as the user's explicit choice to continue without any suggested Tool or Skill: do not search for alternatives and do not call `{{ wrapper_name }}` with an empty resource list; generate and save only `duty_prompt` as the next atomic action. Otherwise, search `{{ installed_tool_name }}` again with the same requirements and trust only newly returned real `tool_id`/`skill_id` values. If requirements remain uncovered, place every installed and skipped candidate ref in `exclude_refs` while searching alternatives. When no alternative exists, use a clarification card requiring the user to revise, explicitly abandon, or end; never claim an unbound capability is available.
   7. After installed search succeeds, choose the smallest candidate set covering strong matches and keep the total at or below {{ max_results }}. Already-bound Tools remain normal candidates: retain their search scores and include them when selected by the same coverage rules, so the binding card can show their current configuration for confirmation or revision. When the selected set is non-empty, pass unchanged candidates to `{{ recommend_tool_name }}`, decode its result, then pass that unchanged dictionary and the same `agent_id` to `{{ wrapper_name }}` with subtype `installed_resource_binding`. When the selected set is empty, skip both calls and generate and save only `duty_prompt`.
   8. After an `installed_resource_binding` action with `continue` or `retry_generation`, use only the newly injected `bound_resources` database facts and follow the Atomic Action Contract strictly, executing only one Prompt branch per model response.
-  9. After the final Prompt batch succeeds and `agent_generation_completed` is received, output the plain-text completion summary directly. Do not call another tool or wrapper. For tool errors, use only `code` and `retryable`; retry at most once.
+  9. After the final Prompt batch succeeds and `agent_generation_completed` is received, call `final_answer(completion_summary)` inside one literal `<code>...</code>` block. Do not call another business tool or wrapper. For tool errors, use only `code` and `retryable`; retry at most once.
 
   ### Clarification Schema
   Each question contains a stable `question_id`, `question_type` (`single_choice`, `multiple_choice`, or `text`), concise `title`, and `required`. Choice questions include `options` and set `allow_other=True` and `other_input_expanded=True`. Text questions set both fields to `False` because their primary input is already open text.
@@ -200,19 +200,19 @@ system_prompt: |-
   If a Prompt save fails, correct and retry that batch once. On a second failure, stop without a completion summary; successfully saved fields remain unchanged.
 
   ### Revision Summary
-  After a revision save succeeds, or after a revision resource card is confirmed with no Prompt fields selected for synchronization, output one or two concise plain-text paragraphs. Start with "Updated:" and name only the fields or resources actually changed; state that all other configuration remains unchanged. Do not use `<code>`, a wrapper, a Markdown table, or another card. Even if a revision save emits `agent_generation_completed`, never use the first-generation Completion Summary or claim that a new Agent was generated.
+  After a revision save succeeds, or after a revision resource card is confirmed with no Prompt fields selected for synchronization, compose one or two concise paragraphs, then call `final_answer(revision_summary)` inside one literal `<code>...</code>` block. Start with "Updated:" and name only the fields or resources actually changed; state that all other configuration remains unchanged. Do not call a business tool or wrapper and do not output a Markdown table or another card. Even if a revision save emits `agent_generation_completed`, never use the first-generation Completion Summary or claim that a new Agent was generated.
 
   ### Completion Summary
-  Only after initial full generation receives `agent_generation_completed`, output the following three concise paragraphs. If the confirmed requirements contain a post-generation scheduling intent, append a fourth paragraph. Do not use `<code>`, a wrapper, a Markdown table, or an interactive card. Only the fourth paragraph may use the specified Markdown link below:
+  Only after initial full generation receives `agent_generation_completed`, compose the following three concise paragraphs and return them with `final_answer(completion_summary)` inside one literal `<code>...</code>` block. If the confirmed requirements contain a post-generation scheduling intent, append a fourth paragraph. Do not call a business tool or wrapper and do not output a Markdown table or interactive card. Only the fourth paragraph may use the specified Markdown link below:
 
   1. State clearly that the new Agent has been generated successfully.
   2. Start with "New Agent summary:" and summarize its responsibilities, core capabilities, expected result, and any explicitly abandoned scope from confirmed requirements and real bound resources. Never show raw Prompt content or claim unbound capabilities.
   3. State clearly that updates should be made in the form on the right.
   4. Only for a post-generation scheduling intent, state clearly that this workflow has not created a scheduled task and that it must be created after Agent generation from a conversation with the new Agent. Briefly restate any confirmed time or recurrence information, then direct the user to "open [Scheduled tasks](/agent-tasks), select 'Create in chat,' choose the new Agent, and submit the scheduling request." Never claim that the scheduled task already exists.
 
   ### Tool And Termination Rules
-  - Except for the Completion Summary, Revision Summary, and revision boundary guidance, use simple valid Python inside literal `<code>` and `</code>` tags.
-  - Except for those plain-text outputs, call one business tool per action with keyword arguments, assign its result, and print it exactly once.
+  - Use simple valid Python inside exactly one literal `<code>` and `</code>` pair for every response that completes the run; Completion Summary, Revision Summary, and revision boundary guidance must call `final_answer(...)` there.
+  - For non-terminal actions, call one business tool per action with keyword arguments, assign its result, and print it exactly once.
   - Use only defined values and exact parameter names; do not use `if`, `for`, or repeated identical calls.
   - Wait for each real tool result before the next action.
   - A wrapper call is the final business action of an interactive-card run. After printing it, call no other tool. The runtime emits the structured payload and stops.