# SPDX-License-Identifier: Apache-2.0
# yaml-language-server: $schema=https://nika.sh/spec/v1/workflow.schema.json
#
# 01 · Hello world — the smallest COMPLETE Nika workflow.
#
# Four questions, four answers, and nothing else: what does it run on, what
# may it touch, what does it do, what does it hand back. A file that skips
# one of the four still executes — it just stops being a contract.
#
# Demonstrates ·
#   - the envelope · `nika: v1` (the contract version) + `workflow.id`
#   - `model:` · one local model · no API key, nothing leaves the machine
#   - `permits: {}` · the DECLARED zero · this workflow may touch nothing
#   - one task · the `infer:` verb, bounded by `max_tokens:`
#   - `outputs:` · the value a caller gets back
#
# Needs · ollama with the model pulled · `ollama pull qwen3.5:4b`.
#   No ollama? The dry twin runs anywhere, zero setup ·
#   `nika run examples/01-hello.nika.yaml --model mock/echo`
#
# Run · nika run examples/01-hello.nika.yaml

nika: v1
workflow:
  id: hello

model: ollama/qwen3.5:4b   # local · zero key · swap for any provider in the catalog

# The declared zero. `permits:` is default-deny once present, so an empty
# block is a statement, not a formality: this workflow opens no file, reaches
# no host, runs no program. Leaving the block out says something different —
# that no boundary was declared — and both forms refuse an effecting tool at
# run, with different verdicts: an absent block reports NIKA-AUTH-006 (no
# boundary to judge against), `permits: {}` reports NIKA-SEC-004 (outside the
# boundary you drew). Declaring the zero is what puts the checker's PERMITS
# and TRIFECTA judges on duty — under-declaring quietly stands them down.
permits: {}

tasks:
  greet:
    infer:
      prompt: "Say hello in French, in one short sentence."
      # The ceiling. On a priced model this is what turns the cost report from
      # UNBOUNDED into a hard worst-case number.
      #
      # 2048 for two words looks absurd until you measure it. `max_tokens` caps
      # think + answer, and qwen3.5 is a reasoning model that deliberates before
      # it speaks — on this exact prompt: 64 → "", 256 → "", 512 → "", 1024 →
      # "Bonjour.", 2048 → "Bonjour !". Under the threshold the run is GREEN and
      # the output is EMPTY, because the budget ran out mid-thought. Size the
      # cap for the thinking, or pick a non-thinking model (llama3.2:3b) when
      # you want a tight one.
      max_tokens: 2048

# The return value. It does NOT appear on the live run — that surface shows
# progress. `--output json` is the export contract: the typed `outputs:` as
# ONE JSON object on stdout, which is how a caller reads a workflow's result
# (`exec: nika run sub.nika.yaml --output json` + `capture: stdout`).
#
#   nika run examples/01-hello.nika.yaml --output json
#   → {"greeting":"Bonjour !"}
outputs:
  greeting: ${{ tasks.greet.output }}

the path · 01

Hello world — the smallest COMPLETE Nika workflow.

run itnika try 01-hello

01-hello.nika.yaml

1# SPDX-License-Identifier: Apache-2.02# yaml-language-server: $schema=https://nika.sh/spec/v1/workflow.schema.json3#4# 01 · Hello world — the smallest COMPLETE Nika workflow.5#6# Four questions, four answers, and nothing else: what does it run on, what7# may it touch, what does it do, what does it hand back. A file that skips8# one of the four still executes — it just stops being a contract.9#10# Demonstrates ·11#   - the envelope · `nika: v1` (the contract version) + `workflow.id`12#   - `model:` · one local model · no API key, nothing leaves the machine13#   - `permits: {}` · the DECLARED zero · this workflow may touch nothing14#   - one task · the `infer:` verb, bounded by `max_tokens:`15#   - `outputs:` · the value a caller gets back16#17# Needs · ollama with the model pulled · `ollama pull qwen3.5:4b`.18#   No ollama? The dry twin runs anywhere, zero setup ·19#   `nika run examples/01-hello.nika.yaml --model mock/echo`20#21# Run · nika run examples/01-hello.nika.yaml2223nika: v124workflow:25  id: hello2627model: ollama/qwen3.5:4b   # local · zero key · swap for any provider in the catalog2829# The declared zero. `permits:` is default-deny once present, so an empty30# block is a statement, not a formality: this workflow opens no file, reaches31# no host, runs no program. Leaving the block out says something different —32# that no boundary was declared — and both forms refuse an effecting tool at33# run, with different verdicts: an absent block reports NIKA-AUTH-006 (no34# boundary to judge against), `permits: {}` reports NIKA-SEC-004 (outside the35# boundary you drew). Declaring the zero is what puts the checker's PERMITS36# and TRIFECTA judges on duty — under-declaring quietly stands them down.37⛨permits: {}3839tasks:40  greet:41    ◇infer:42      prompt: "Say hello in French, in one short sentence."43      # The ceiling. On a priced model this is what turns the cost report from44      # UNBOUNDED into a hard worst-case number.45      #46      # 2048 for two words looks absurd until you measure it. `max_tokens` caps47      # think + answer, and qwen3.5 is a reasoning model that deliberates before48      # it speaks — on this exact prompt: 64 → "", 256 → "", 512 → "", 1024 →49      # "Bonjour.", 2048 → "Bonjour !". Under the threshold the run is GREEN and50      # the output is EMPTY, because the budget ran out mid-thought. Size the51      # cap for the thinking, or pick a non-thinking model (llama3.2:3b) when52      # you want a tight one.53      max_tokens: 20485455# The return value. It does NOT appear on the live run — that surface shows56# progress. `--output json` is the export contract: the typed `outputs:` as57# ONE JSON object on stdout, which is how a caller reads a workflow's result58# (`exec: nika run sub.nika.yaml --output json` + `capture: stdout`).59#60#   nika run examples/01-hello.nika.yaml --output json61#   → {"greeting":"Bonjour !"}62outputs:63  greeting: ${{ tasks.greet.output }}64

nika-spec@6ac29351a · sha256 f857a9ec2910dd47…