# Nika > Runs repeatable AI work with checks, budgets and traces. - Skill: `supernovae-st/nika-2` (Agent Skill) - Install (CLI): `npx skillmds@latest add supernovae-st/nika-2` - Raw SKILL.md: https://api.skillmd.com/api/skills/supernovae-st/nika-2/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: MIT - Author: supernovae-st (https://skillmd.com/u/supernovae-st) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/supernovae-st/nika-2 --- # Nika Skill Use [Nika](https://nika.sh) through the Hermes `terminal` tool to capture repeatable AI work in checked `*.nika.yaml` files. Nika is an AGPL-3.0-or-later workflow engine; this MIT-licensed skill teaches the Hermes handoff to public Nika 0.118.7. Hermes owns the user's intent and authorization; Nika checks the file, admits execution under its boundaries, and produces outputs and evidence when execution records them. ## When to Use - Author, check, run, or diagnose an existing `*.nika.yaml` workflow. - Capture repeated model, file, HTTP, or command work in a reusable artifact. - Inspect costs, declared authority, outputs, or evidence before repeating a job. For a one-off answer or tool call, use the relevant Hermes tool directly. For interactive coding work, use the existing coding skill; a workflow becomes useful when a repeatable part has a clear input, output, and failure policy. ## Prerequisites Use `terminal(command="nika --version")` to identify the installed release. Installation and client wiring are separate effects: carry out already authorized setup within scope, or report the missing prerequisite. Installation routes are at https://nika.sh; do not assume an installed command matches this skill's release. Use `terminal(command="nika doctor")` for setup diagnostics and `nika catalog` for available model names. Keep credentials in the configured stores or environment; never put secret values in workflow files, tool arguments, logs, or a report. A local model still needs its runtime and compute resources. `mock/echo` simulates inference; it does not make arbitrary workflows offline or harmless. ## How to Run Use `search_files` to find an existing workflow, `read_file` to inspect it, and `write_file` or `patch` to author or repair it. Keep the artifact in the user's project and pass that project as the `terminal` working directory. Discover a template and create a file without overwriting an existing owner: ``` terminal(command="nika new '?'") terminal(command="nika new chain flow.nika.yaml", workdir="~/project") ``` Read the generated file. For language details, use the installed `nika --help`, `nika catalog --tools`, and the matching release's [authoring guide](https://github.com/supernovae-st/nika/blob/v0.118.7/.agents/plugins/nika/skills/nika-authoring/SKILL.md). If the read-only `nika mcp` oracle is already wired, use `nika_schema`, `nika_examples`, `nika_template`, `nika_check`, and `nika_explain` as needed. The oracle proposes or diagnoses; workflow execution remains a separate action. ### Preflight and authority Check the exact file after every edit, with the model override intended for the run if there is one: ``` terminal(command="nika check flow.nika.yaml --json --native-strict", workdir="~/project") terminal(command="nika explain flow.nika.yaml", workdir="~/project") ``` Read the exit code AND the JSON. `clean`, `native_strict_clean`, and `paid_ready` answer different questions. Require native-strict cleanliness before handing over a run; before paid inference, require `paid_ready: true` and inspect the access, capacity, model, and risk findings. Neither a green check nor a ready access lane is permission from the user, a live connectivity test, or proof that effects ran. Use `nika explain NIKA-XXXX` for a diagnostic. Fix its cause and check again. Preserve existing authorization and budgets. Continue work already authorized; resolve only a genuinely missing business decision, authority, or spend limit. Do not widen `permits:` or secret `egress:` merely to silence the checker. `nika check flow.nika.yaml --infer-permits` proposes a boundary to review against the intended effects; it does not authorize that boundary. Missing permits grant zero authority. Keep secrets store-backed and admit only intended sinks. ### A small native example Save this as `hash-abc.nika.yaml`. It needs no model or external service and calls a native builtin; check it before considering any run: ```yaml nika: hash-abc permits: tools: [nika:hash] tasks: probe: invoke: tool: nika:hash args: { content: abc, algo: sha256 } outputs: result: ${{ tasks.probe.output }} ``` Use exactly one of four verbs per task: `infer` for model output, `exec` for an external command, `invoke` for a callable tool or workflow, and `agent` for a bounded adaptive loop. Prefer an existing native tool before writing shell glue. A model can propose facts; deterministic tools enforce business rules. Bound `infer` with `max_tokens` and `agent` with `max_turns` and `max_tokens_total`. Caller values belong in `inputs:`, fixed values in `const:`, credentials in `secrets:`. Bind another task's data through `with:`; `after:` orders tasks when no data flows. Consult the schema rather than inventing fields or arguments. `nika:compose` checks a draft supplied through an agent's tool access; it does not execute it. `invoke: { workflow: ... }` calls a child through normal engine admission. The child must fit the parent's authority; neither drafting nor checking grants new execution rights. ### Execution and recovery Once the exact workflow, inputs, model, effects, and budget are ready and already authorized, execute through `terminal`. This is a command shape: replace every placeholder with the checked value; retain the user's configured cap. ``` terminal(command="nika run flow.nika.yaml --model --var = --max-cost-usd ", workdir="~/project") ``` For long work, use the Hermes terminal's supported background/session mechanism and retain its returned identifier. Inspect that session before starting another copy. A timeout or lost terminal response does not prove the process stopped. Before retrying after interruption, reconcile the process, trace, outputs, and external effects. Preserve negative results and never retry publication blindly. For a paused `nika:prompt`, use `nika run flow.nika.yaml --resume --answer =` to supply the authorized answer; it is execution, not inspection. Read `nika run --help` and preserve the original inputs, model, cap, and scope. Transient errors may use bounded `retry:`; expected errors may use a deliberate `on_error:` recovery. `after: { producer: unwind }` is best-effort cleanup for a producer that started, including cancellation or timeout. It does not run for an unstarted producer, and process death can prevent cleanup. ## Quick Reference | Command | Use | |---|---| | `nika new '?'` / `nika new