PowerX (service) - Natural Language Task Parsing & Plugin Matching

Usecase Overview

• Business Goal: Quickly convert business users' natural language instructions into structured task plans, automatically match appropriate plugin/tool combinations, and provide traceable context for subsequent parallel execution and risk control.
• Success Metrics: Return executable plans within 2 seconds on average; plugin matching accuracy ≥ 90%; low-confidence prompts accurately guide supplementary information; all plans written to audit.
• Scenario Linkage: Corresponds to main scenario Stage 1「Intent Parsing & Capability Planning」, the entry point of the task chain, directly determining subsequent execution efficiency and security boundaries.

The Planner needs to balance NLU accuracy with capability coverage, maintaining second-level response experience while outputting auditable and controllable task DAGs.

Context & Assumptions

• Prerequisites
  • agent-orchestrator-v2 and capability-graph-service Feature Flags enabled.
  • Plugin metadata (version, input/output, tenant availability, sensitivity level) registered in capability graph.
  • Conversation/instruction center can provide tenant, user, language and other context fields.
  • Audit and metrics pipeline available to record Planner output and risk prompts.
• Inputs/Outputs
  • Input: Natural language instructions, context entities (customer, billing period, channel), historical conversation snippets, tenant policies, Feature Flags.
  • Output: Task DAG (nodes, dependencies, parameter mapping), candidate plugin list, plugin scores/confidence, risk and approval suggestions, audit record ID.
• Boundaries
  • Not responsible for plugin capability implementation or testing.
  • Does not cover ReAct Prompt design (handled by ReAct scenario).
  • Does not handle execution retries or human collaboration in this usecase.

Solution Blueprint

System Decomposition

Component	Responsibility	Description
Intent Parser	NLU, entity extraction, confidence evaluation	Structured processing of input statements based on multi-language models + rules.
Constraint Extractor	Constraint identification, context merging	Extract SLA, budget, sensitivity level and other constraints, merge with tenant policies.
Capability Graph Service	Plugin capability retrieval and scoring	Multi-dimensional scoring based on task type, data requirements, plugin health.
Plan Builder	Build task DAG and step descriptions	Output node dependencies, input/output mapping, risk labeling, approval strategy.
Audit & Telemetry Writer	Record plans and risk metrics	Write to agent_plan table, publish agent.plan.created event.

Process & Sequence

1. Step 1 – Input Parsing: Intent Parser performs semantic analysis, entity extraction, and confidence calculation on natural language.
2. Step 2 – Constraint Merge: Constraint Extractor merges input constraints with tenant policies, filling missing fields.
3. Step 3 – Capability Retrieval: Capability Graph loads plugin candidates based on task type and context, scoring with health signals and historical success rates.
4. Step 4 – Plan Generation: Plan Builder generates task DAG, determines node order, call parameters, callbacks, approval strategy and risk labels.
5. Step 5 – Output & Audit: Planner returns plan to Orchestrator, writes to audit, metrics and risk prompts; requests user to supplement information if confidence is low.

Contracts & Interfaces

• Inbound:
  • POST /internal/agent/intents:parse — Called by conversation/command center, includes tenant, user, utterance, context.
  • EVENT agent.intent.created — Prompt asynchronous Planner processing, suitable for batch requests.
• Outbound:
  • POST /internal/capabilities/search — Retrieve plugin capabilities based on task tags, data domain, tenant availability.
  • POST /audit/agent-plan — Write plans, risks, plugin lists.
  • POST /notifications/agent/need-context — Request supplementary information when confidence is too low.
• Configuration/Scripts:
  • config/agent/intent_rules.yaml — Intent templates and fallback rules.
  • config/agent/capability_weights.yaml — Plugin scoring factors.
  • scripts/qa/intent-regression.mjs — Parser regression test script.

Implementation Checklist

Item	Description	Status	Owner
Parser multi-language support	Introduce multi-language models and fallback rules	[ ]	Agent Platform Guild
Capability graph scoring	Integrate health signals, tenant whitelists	[ ]	Plugin Guild
Risk annotation	Support sensitive task approval prompts	[ ]	Ops Reliability Center
Audit output	Write plans, plugins, constraints to unified audit channel	[ ]	Agent Platform Guild
Low confidence supplementation	Auto-generate clarification questions and notify users	[ ]	Agent Platform Guild

Testing Strategy

• Unit: Intent parsing, entity extraction, scoring functions, DAG dependency topology validation.
• Integration: End-to-end Parser + Capability Graph + Plan Builder, verify normal tasks, sensitive tasks, no available plugins three paths.
• End-to-End: Initiate real tasks from sandbox conversation entry, check Planner output, audit logs, alert prompts.
• Non-functional: 200 QPS pressure test; inject Graph slow queries to verify timeout protection; Chaos simulate partial plugin health signal loss.

Observability & Ops

• Metrics: agent.plan.latency_p95, agent.plan.success_rate, agent.plan.low_confidence_total, agent.plan.audit_write_total.
• Logs: Record plan_id, intent, confidence, selected_plugins, risk_flags; mask PII.
• Alerts: Plan time >5s (5-minute window), matching failure rate >5%, audit write failure >1%; push through Grafana + PagerDuty.
• Dashboard: Grafana「Agent Planner」, Datadog Trace「planner.*」, internal audit replay panel.

Rollback & Failure Handling

• When Planner upgrade fails, can rollback to previous container image and restore old weight configuration.
• If capability graph unavailable, degrade to rule table matching or prompt manual process.
• When widespread low confidence alerts occur, enable planner-safe-mode to only allow whitelisted tasks through.

Follow-ups & Risks

Risk	Impact	Mitigation	ETA
Plugin health signals not fully integrated, affecting scoring stability	Plan selection errors	Align metric fields with Plugin Guild, publish health signal SDK	2025-03-10
Insufficient multi-language support	Parsing failures for some tenants	Expand example corpus, gray launch by region	2025-03-05

References & Links

• Scenario Document: docs/scenarios/agent-orchestration/SCN-AGENT-TASK-EXEC-001.md
• Design Draft: docs/meta/scenarios/powerx/agent-and-automation/agent-orchestration/agent-task-execution/primary.md
• Related Standards: docs/standards/powerx/backend/integration/09_agent/Agent_Adaptor_and_Transport_Spec.md