Verified against Semitexa Ultimate 2026.09.19.1020

Prompt CLI

The package ships console commands to inspect and operate the catalog. Each supports --json for machine-readable output.

How it works

# List the catalog (optionally filtered by channel)
bin/semitexa prompt:list --channel=llm

# Show one prompt: metadata, raw template, variables, partials
bin/semitexa prompt:show --id=os.persona

# Render the effective prompt (tenant override -> catalog) exactly as the
# model would receive it; --var supplies the getter values by name
bin/semitexa prompt:render --id=os.persona --var=assistantName=Semi --var=userName=Taras

# Manage per-tenant overrides
bin/semitexa prompt:override set --id=os.persona --system="..."
bin/semitexa prompt:override list
bin/semitexa prompt:override history --id=os.persona
bin/semitexa prompt:override revert --id=os.persona --rev=2
bin/semitexa prompt:override remove --id=os.persona

# Record who changed an override and why (kept on the version)
bin/semitexa prompt:override set --id=os.persona --system="..." \
    --author="ops lead" --reason="the room said the answers ramble"

# Append attributable guidance instead of rewriting a body
bin/semitexa prompt:guidance add --id=os.persona --text="Keep replies to two sentences." \
    --author="ops lead" --reason="the room said the answers ramble"
bin/semitexa prompt:guidance list --id=os.persona
# The row id comes from `prompt:guidance list` (the "Row" column)
bin/semitexa prompt:guidance disable --row=01a0a11f-b676-76a2-a307-81861ca79800
bin/semitexa prompt:guidance remove --row=01a0a11f-b676-76a2-a307-81861ca79800

prompt:render shows the resolved text with overrides and guidance applied, so it is the fastest way to confirm what a prompt actually expands to. prompt:eval (from semitexa/llm) goes one step further and sends the rendered prompt to the live LLM, optionally comparing the tenant override against the catalog default.

Why this matters

Prompts are debuggable like any other addressable resource. You can list what exists, see exactly how a prompt renders for a given tenant, and manage overrides — all without reading source or redeploying.