KRISHI-NEXUS backend foundation

Phase 12 adds real retrieval/assessment components while leaving the deterministic FSM, safety engine, reviewer, memory, and the Synthesizer mock in place.

Phase 12.2: Field Context Agent

ToolFieldContextAgent is a deterministic, mockable retrieval layer between the FSM and three tool boundaries:

• farm.get_state(farmId, plotId)
• weather.get_forecast(location, startTime, hours)
• farm.get_history(farmId, plotId, fromDate, toDate)

It copies only retrieved facts into CaseState.farm, environment, and history. CaseState.context separately records retrieved field names, derived context (currently empty), tool statuses, and conflicts. Missing, failed, timed-out, or conflicting context never receives fabricated fallback values; the workflow safely falls back before risk, synthesis, or safety evaluation.

Phase 12.3: Risk & Economics Agent

ToolRiskEconomicsAgent deterministically converts Perception and Field Context facts into disease, weather, crop-stress, and overall risk factors. It retrieves money inputs through economics.get_data(farmId, plotId, crop, cropStage) and returns a typed crop value, expected loss, intervention cost, justification, and decision gate: INTERVENE, MONITOR, WAIT, or SEEK_CONFIRMATION.

No model selects prices, loss amounts, product names, chemicals, or dosages. Every monetary input must be supplied by the economic tool boundary. Missing, failed, timed-out, negative/non-finite, or conflicting values are recorded as structured flags and result in SEEK_CONFIRMATION; no value is manufactured. The built-in data is explicitly SEEDED_DEMO, is used only by deterministic tests/demo flow, and is not live market information. When cost is not justified, the mock synthesizer retains its role but emits MONITOR; concrete interventions remain downstream of the vetted knowledge base and safety engine.

Perception model configuration

The Perception Agent reads OPENAI_API_KEY from the environment and uses OPENAI_MODEL when supplied (default: gpt-4o). It sends farmer text, an optional voice transcript, and image URLs/data URLs to the Responses API, requesting structured observations only. It records confidence and uncertainty, and has no treatment, chemical, dosage, or action fields.

$env:OPENAI_API_KEY = "..."
$env:OPENAI_MODEL = "gpt-4o" # optional
npm run dev

The test suite injects a fake model boundary; it never uses an API key or calls a live model.

Run

Requires Node.js 24+ (for built-in TypeScript stripping).

npm run dev

The server listens on http://localhost:3000.

Invoke-RestMethod http://localhost:3000/api/v1/demo/tomato -Method Post

Or submit an intake payload:

Invoke-RestMethod http://localhost:3000/api/v1/decisions -Method Post -ContentType 'application/json' -Body '{"farmId":"FARM-001","plotId":"PLOT-A","language":"kn","farmerText":"Tomato leaves have dark spots and the affected area is increasing."}'

Test

npm test

The seeded demo deliberately proposes a foliar biological intervention during a rain window. WX-001 blocks it, the mock planner replans to WAIT_FOR_SAFE_WEATHER_WINDOW, and the reviewer approves the safe plan. This provides an auditable end-to-end demonstration of the veto/replan architecture.

Layout

• src/types: CaseState and decision contracts
• src/agents: replaceable mocked reasoning components
• src/tools: in-process MCP-compatible tool interfaces plus seeded data
• src/rules: deterministic safety gate
• src/orchestrator: finite-state decision pipeline
• src/memory: in-memory decision repository (replaceable by a database adapter)
• src/api: HTTP request handling

All price, weather, and treatment data are demo/seeded values only, not agronomic advice.