Nika Workflow

Nika Workflow

Nika is an open-source Rust engine that captures a repeatable AI task as one plain-text *.nika.yaml file. The file is audited before a single token is spent, runs against local or cloud providers under a spend ceiling, and records a hash-chained journal you can verify afterwards.

Use it for the work that repeats. Nika is not a coding agent and does not compete with one: OpenClaude writes the code, Nika runs the chore that comes back every week and hands you a receipt for it.

Use this skill when

• The user describes a prompt or prompt chain they run more than once — a weekly digest, an inbox triage pass, a release-note draft — and wants it captured as something re-runnable.
• A shell or Python script in the repo calls a model API, or wraps HTTP, file and JSON plumbing around model output.
• The user asks for a hard spend ceiling on an AI task, or asks what a task will cost before it runs.
• The user needs a record of what an AI run actually did: a receipt, an audit trail, something a reviewer can check independently.
• A *.nika.yaml file is open or being written, or nika check reported a NIKA-XXXX finding that needs fixing.

Do NOT use this skill when

• The user wants a one-off answer or a single tool call — answer directly, a workflow file is overhead.
• The user wants application code written, refactored or reviewed — that belongs to a coding skill.
• The task needs conversational back-and-forth mid-execution — workflows are non-interactive by design; a confirm gate asks one question, it does not hold a discussion.

Procedure

1. Confirm the engine is present: nika --version. If it is missing, say so and stop — installing is the user's move (brew install supernovae-st/tap/nika, or the paths listed at nika.sh). Never improvise workflow YAML from memory.
2. Start from a skeleton, never a blank file. nika new --from '?' lists the embedded templates, nika new --from <template> <file>.nika.yaml writes one, and bare nika try lists complete runnable lessons.
3. Fill it in. The envelope opens with nika: <kebab-case-name> — that one key carries both the mark and the file's name, and there is no workflow: envelope key — plus a tasks: map keyed by task id. Exactly one verb per task: infer (a model call) · exec (a subprocess, whose command: is an argv list) · invoke (a builtin or MCP tool) · agent (a bounded multi-turn loop). Every value the file depends on is declared in one of three authorities: inputs: · const: · secrets: — a deployment-supplied value is an inputs: entry with required: false and a default:.
4. Declare the boundary. permits: states what the workflow may touch, and an ABSENT block means zero authority — any effect without a grant is refused at check time. nika check <file> --infer-permits prints the tightest block the workflow actually needs; paste it in.
5. Audit before anything runs: nika check <file>. Exit 0 is clean, exit 2 carries findings. Each finding names its task and the fix it wants; nika check <file> --fix applies the mechanical repairs. Decode an unfamiliar code with nika explain NIKA-XXXX. Loop until clean.
6. Report cost from the check output, never from a guess. ≤ $X is a ceiling; ≥ $X FLOOR means at least one task is unbounded, and you name why. A model running locally is unpriced compute — say unpriced, never free.
7. Hand the run to the user. Typing nika run <file> is their move; give them the line, with --max-cost-usd <n> when spend matters. An offline rehearsal costs nothing: --model mock/echo.
8. After a run that mattered, prove it. nika trace verify <trace> checks the hash-chained journal under .nika/traces/, and nika trace evidence <trace> exports a pack a reviewer can check without trusting you. Cite the trace, never a memory of the run.

Examples

In scope: "every Monday I paste a competitor changelog into a model and ask what changed" → capture it as a workflow with a fetch task, a bounded infer, and a written report; audit it clean; hand over the run line.

In scope: "this deploy script calls a model to draft the release note" → port the model call to infer:, the file writes to the nika:write builtin, declare the permits, pin the behavior with nika test <file> --update, and leave the old script for the user to retire.

Out of scope: "explain what this regular expression does" → answer directly; no file, no workflow.

Out of scope: "refactor this module" → a coding skill owns that. Nika does not write application code.

Self-check before responding

• Did nika check exit 0 on the exact file being handed over? If it did not, the file is not ready to hand over.
• Is the cost reported the way the audit reported it — a ceiling, or a floor with a named reason — with no local model described as free?
• Does the file declare a permits: block, or is the body genuinely pure compute and saying so with an empty one?
• Was the run proposed as a command line rather than executed?
• Is every credential referenced through the declared secrets: block, with no literal value written anywhere in the file?