codify

# Codify <!-- tkstack:execution:start --> Tool: `.tkstack/tkstack skills codify -- <command>`. Omit the command for deterministic man-page help. | Executable operation | Contract | | --- | --- | | Default/help and man(1) export | [Reserved help](commands/help.md) | | Declared package version | [Reserved version](commands/version.md) | | `model` | [Typed parameters](commands/model.md) | | `verify` | [Typed parameters](commands/verify.md) | | `native check` | [Typed parameters](commands/native_check.md) | | `native scaffold` | [Typed parameters](commands/native_scaffold.md) | | `native compile` | [Typed parameters](commands/native_compile.md) | | `native run` | [Typed parameters](commands/native_run.md) | | Evidence, effects, native execution, and semantic boundaries | [Execution contract](references/execution.md) | `model` reads the explicit contract; `verify` checks retained facts/evidence. Native commands execute existing code. Residual semantic decisions use the focused procedures below; a recorded review is not an independently authenticated judgment. <!-- tkstack:execution:end --> Translate the procedure once into native code, verify its contract, and publish an executable `*.lock.json`. Subsequent runs execute the code frozen in that artifact. Markdown is authoring input and tool discovery; avoid reconstructing the algorithm on every invocation. Tool: `python3 <codify-directory>/scripts/codify.py <command> ...`. Commands: `scaffold`, `compile`, `check`, `run`, `commands`, `invoke`. `commands` validates/exports Markdown tool signatures; `invoke` dispatches CLI or OpenAI tool-call arguments into the same verified lock. Every command package has reserved help/version discovery. Default help renders a generated manual with fixed console rules; `help --format man` exports man(1) source. Version reads declared package metadata. Discovery never invokes a lock. | Work | Read before acting | | --- | --- | | Define/invoke subcommands, aliases, typed flags, or OpenAI tools | [Markdown command contract](references/command-definitions.md) | | Lower a prompt/Markdown/skill into actual code | [Compilation procedure](references/compilation.md) | | Invoke the compiler or execute its artifact | [Commands, parameters, and contracts](references/commands.md) | | Understand frozen code, provenance, invalidation, and verification | [Artifact contract](references/artifact.md) | | Configure the residual model, Decisions, or harness boundary | [Inference profiles](references/inference.md) | | Reuse this engine from another authoring tool | [Reuse and ownership](references/reuse.md) | | Compare with GitHub Agentic Workflows | [Sources and differences](references/sources.md) | | Try a complete deterministic conversion | [Example source](examples/normalize-labels/PROMPT.md), [cases](examples/normalize-labels/cases.json) | Use existing code/assets before generating new implementations. Keep workflow control flow, facts, gates, retries, and bounds in code; declare only residual inference. The compiler verifies supplied implementation and cases; it cannot infer arbitrary natural-language meaning. The authoring agent completes that one-time translation. A starter, prompt-forwarding wrapper, or passing example is insufficient evidence of semantic equivalence. Keep authored workflow definitions and lock artifacts in their requested/native paths (a skill's `scripts/`). Keep private records, staging, and locks under `.tkstack/`; read [local state storage](references/storage.md). Compilation runs contract code; `check` and `compile --dry-run` are read-only. Running an artifact executes trusted code with the caller's authority; it is not a sandbox.