Authors and validates .nika.yaml DAG workflows with cost audits and tamper-evident run traces.
npx skills add https://github.com/davepoon/buildwithclaude --skill nika
Please help me install this Agent Skill. Skill: nika Source: https://github.com/davepoon/buildwithclaude/tree/main/plugins/all-skills/skills/nika Command: npx skills add https://github.com/davepoon/buildwithclaude --skill nika
Repeatable AI tasks often live as ad-hoc prompts with unknown cost, unverified structure, and no audit trail. This Skill captures a repeated task as a checked, budgeted .nika.yaml workflow file that is audited before a single token is spent.
nika examples or nika new --from <template>, then pick models and builtins from the embedded nika catalog and nika tools listings.nika check validates the DAG, prints a cost ceiling or floor before any token is spent, and --native-strict flags exec: tasks that a builtin could replace.nika test replays golden files offline with mock models, and every run writes a hash-chained journal to .nika/traces/ verifiable with nika trace verify..nika.yaml DAG with typed infer: outputs, a --max-cost-usd spend cap, and CI-pinned golden tests so the workflow runs identically every time.Ask the agent to capture your repeated AI task as a nika workflow file, run nika check until it passes, and hand you the validated .nika.yaml ready to run.
High-intent search queries and answers about installing and using this skill.
Start from a template with `nika examples list` and `nika new --from <template> <file>.nika.yaml`, never from scratch. The envelope is always `nika: v1` plus `workflow: <kebab-id>` plus `tasks:`, then validate with `nika check` until it exits clean.
Run `nika check <file>`, which prints a cost ceiling before any token is spent. A `≤ $X` value is a ceiling, while `≥ $X FLOOR` means a task is unbounded due to missing max_tokens, an uncataloged model, or expression fan-out.
Each task uses exactly one verb: `infer:` for LLM calls, `exec:` for shell commands as a last resort, `invoke:` for builtins or MCP tools, and `agent:` for bounded multi-turn loops with tool allowlists and turn limits.
Yes. Preview with `--model mock/echo` for offline runs, then pin the workflow with `nika test <file> --update` to write a golden file. Subsequent `nika test` runs replay and compare deterministically with zero keys required.
Only after confirming no `nika:*` builtin or MCP tool covers the task, such as build tools, git, or CLIs without an MCP surface. Every surviving `exec:` must be documented in the workflow's exec ledger header to pass `--native-strict` review.
Every run writes a hash-chained journal to `.nika/traces/`. Use `nika trace verify <trace>` to check the chain for tampering and `nika trace show <trace>` to read the run card, citing the trace rather than memory.
edri2orInject environment variable secrets into task context without exposing values.
ruvnetGenerates token usage and USD cost reports aggregated by agent, model, and tier.
zhouboyu-xrealConnect Hermes Agent to MCP servers for external tool integration.
EngineerWithAIGenerate Apache Airflow DAGs using TaskFlow API patterns.