Feat: nl2agent

[xuyaqist]

当资源不足时弹出资源不足卡片，支持用户创建skill，和配置mcp

支持绑定相关资源

[Copilot]

🟡 Changes recommended

Unresolved critical and moderate findings affect resource validation, binding, configuration, and installation flows.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds end-to-end NL2Agent resource-gap resolution, including resource discovery, installation, Skill/MCP configuration, verification, and binding.

Changes:

• Adds backend resource resolution, installation, configuration, and MCP tooling.
• Adds frontend resource-gap, installation, configuration, and binding flows.
• Updates prompts, localization, models, APIs, and regression tests.

Review findings:

• Critical (2 votes): backend/tool_collection/mcp/nl2agent_mcp_tools.py lines 971, 1020, and 1477 allow unvalidated gap-card construction.
• Moderate (2 votes): MCP cancellation, resource configuration payloads, and installation requests have state or request-handling issues.
• Moderate (1 vote): Skill binding state, resource hints, and inactive installation-card actions need correction.
• Nits (1 vote): English and Chinese prompt examples need safer candidate handling and consistent resource-task progression.

File summaries

File	Description
test/backend/services/test_nl2agent_service.py	Backend service regression tests
test/backend/services/test_mcp_internal_tool_search.py	MCP search and tool tests
test/backend/app/test_agent_app.py	Agent application endpoint tests
test/backend/agents/test_nl2agent_agent.py	NL2Agent integration tests
frontend/tests/nl2agentResourceResolution.test.ts	Frontend resource-resolution tests
frontend/services/api.ts	Shared API request handling
frontend/services/agentConfigService.ts	Agent configuration API client
frontend/public/locales/zh/common.json	Chinese localization strings
frontend/public/locales/en/common.json	English localization strings
frontend/lib/nl2agent-resource-resolution.ts	Resource-resolution helpers
frontend/lib/nl2agent-resource-gap.ts	Resource-gap helpers
frontend/lib/nl2agent-mcp-configuration.ts	MCP configuration helpers
frontend/contexts/nl2AgentFlow.tsx	NL2Agent workflow state
frontend/app/[locale]/newchat/ui/suggested-resource-installation-card.tsx	Resource installation UI
frontend/app/[locale]/newchat/ui/resource-gap-resolution-card.tsx	Resource-gap resolution UI
frontend/app/[locale]/newchat/ui/installed-resource-binding-card.tsx	Resource binding UI
frontend/app/[locale]/newchat/assistant-ui/thread.tsx	Chat thread integration
frontend/app/[locale]/newchat/adapter/remote-chat-model-adapter.ts	Resource payload normalization
frontend/app/[locale]/agents/components/agent-capability.tsx	Skill and MCP configuration UI
backend/tool_collection/mcp/nl2agent_mcp_tools.py	MCP schemas, validation, and tools
backend/tool_collection/mcp/nl2agent_mcp_service.py	MCP service integration
backend/services/skill_repository_service.py	Skill repository operations
backend/services/nl2agent_service.py	Resource resolution and installation logic
backend/prompts/nl2agent_zh.yaml	Chinese NL2Agent workflow prompts
backend/prompts/nl2agent_en.yaml	English NL2Agent workflow prompts
backend/consts/model.py	Model and resource constants
backend/apps/agent_app.py	Installation and configuration endpoints
backend/agents/nl2agent_agent.py	NL2Agent agent integration

Review details

Suppressed comments (10)

backend/prompts/nl2agent_en.yaml:81

• This transition requires the resource task to enter the same atomic Prompt-generation sequence, but the English prompt's atomic-action rule above limits updated_fields progression to full_generation (the Chinese prompt includes manage_resources). As written, a partial resource task can stop after binding instead of regenerating the required resource-dependent Prompt fields; align the English stage condition and numbered branches with this transition.

  7. When the backend returns `next_action="BIND"`, pass the complete unified resolution and the same `agent_id` unchanged to `{{ wrapper_name }}` with subtype `installed_resource_binding`.
  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.

backend/prompts/nl2agent_en.yaml:127

• The example unconditionally indexes the first resource. An uncovered result may have no resources, which raises an index error; if it has only weak references, this code incorrectly accepts a weak candidate despite the workflow requiring candidates whose facts prove the capability. Build the accepted list only from strong/recommended candidates after the model confirms their declared facts, and use an empty list otherwise.

  capability_verifications = [
      {"requirement_id": "verified_research", "accepted_candidate_refs": [verification_result["resources"][0]["candidate_ref"]]},
  ]

backend/prompts/nl2agent_zh.yaml:128

• The example unconditionally indexes the first resource. An uncovered result may have no resources, which raises an index error; if it has only weak references, this code incorrectly accepts a weak candidate despite the workflow requiring candidates whose facts prove the capability. Build the accepted list only from strong/recommended candidates after the model confirms their declared facts, and use an empty list otherwise.

  capability_verifications = [
      {"requirement_id": "verified_research", "accepted_candidate_refs": [verification_result["resources"][0]["candidate_ref"]]},
  ]

backend/services/nl2agent_service.py:1485

• Every installed candidate is marked is_bound=False here, even though the binding card uses this flag as the agent-specific persisted binding state. The frontend only synchronizes existing tool bindings, not skills, so an already-bound skill is shown as pending and is submitted again instead of remaining bound; load the target agent's binding state before emitting this summary.

                is_bound=False,

backend/tool_collection/mcp/nl2agent_mcp_tools.py:1482

• This re-resolution drops the capability_verifications selected by the model. If verification rejects every candidate, the backend can rediscover the same strong resource and return BIND/INSTALL (or make the wrapper subtype mismatch), so the final card no longer reflects the verified result. Preserve the accepted candidate refs when re-resolving, rather than passing an unrestricted request.

            verified = await resolve_resource_requirements_impl(
                requirements=[item.requirement for item in supplied.requirements],
                phase=supplied.phase,
                exclude_refs=[],
                tenant_id=tenant_id,
                user_id=user_id,

backend/tool_collection/mcp/nl2agent_mcp_tools.py:1023

• The prompt requires a verification pass with verification_required=True and then capability_verifications, but this ToolConfig exposes neither optional argument to the NL2Agent model. Because this config is installed into create_nl2agent_agent_config, the model cannot reliably perform the mandated verification call. Add both parameters to the advertised input schema.

                '{"agent_id":"int",'
                '"requirements":"list[ResourceRequirement]",'
                '"phase":"INITIAL | POST_INSTALL | POST_GAP",'
                '"exclude_refs":"list[str]"}'

backend/tool_collection/mcp/nl2agent_mcp_tools.py:432

• ResourceRequirement carries resource_name_hint, and the frontend's unchanged-gap action expects to preserve it, but this gap-card model omits the field. _build_v2_resource_payload therefore cannot emit the original hint, so confirming an unchanged requirement silently drops matching metadata despite the prompt requiring unchanged requirements to be preserved. Add the hint to this payload and populate it when building the card.

    requirement_id: str = Field(min_length=1, max_length=100)
    query: str = Field(min_length=1, max_length=500)
    search_terms: list[str] = Field(default_factory=list, max_length=8)
    weak_references: list[WeakResourceReference] = Field(

frontend/app/[locale]/agents/components/agent-capability.tsx:179

• The only transition that marks this request complete is the success callback. The modal's cancel handler merely closes the modal, leaving the request pending and disabling the gap card; because the request ID is already recorded, reopening the same request is also suppressed. Add a cancellation/reset transition to the flow state and invoke it on cancel so an abandoned creation remains retryable.

  const handleSkillBuildSuccess = useCallback(() => {
    invalidateSkills();
    if (
      skillCreationRequest &&
      skillCreationRequest.agentId === currentAgentId &&
      !skillCreationRequest.completed
    ) {
      completeSkillCreation(
        skillCreationRequest.agentId,
        skillCreationRequest.requestId
      );
    }

frontend/app/[locale]/newchat/ui/suggested-resource-installation-card.tsx:91

• canContinue ignores interactive, so a card that has become inactive can still enable Continue once its items reach terminal states and submit a duplicate action. Include the interaction state in this derived guard so only the active card can continue.

  const canContinue = items.every(
    (item) => item.status !== "idle" && item.status !== "installing"
  );

frontend/app/[locale]/newchat/ui/suggested-resource-installation-card.tsx:259

• The failed-item Skip button remains enabled when interactive is false. Users can mutate an inactive installation card after a newer card has taken over, producing stale skipped results. Disable this action for inactive cards.

                    <Button
                      type="button"
                      size="icon"
                      variant="ghost"
                      className="size-9"

• Files reviewed: 28/28 changed files
• Comments generated: 4
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.