Phoenix CLI Development

Phoenix CLI Design Specification

The Phoenix CLI (px) is a command-line interface for the Phoenix AI observability platform. It serves two distinct audiences simultaneously: humans typing commands in a terminal and coding agents (Claude Code, Cursor, Codex, Gemini CLI) executing commands programmatically.

This specification uses RFC 2119 keywords (MUST, SHOULD, MAY, etc.) to indicate requirement strength.

Command Structure: Noun-Verb

All commands MUST follow a noun-verb pattern, modeled after the GitHub CLI (gh):

px <resource> <action> [arguments] [options]

Resource names MUST be singular — they name the type of thing you're acting on, not how many:

px project list              # not "px projects"
px project create            # not "px create-project"
px trace get <trace-id>
px dataset get <name-or-id>
px auth status

Standard verbs

Commands SHOULD use these verbs consistently across all resources:

Not every resource supports every verb — datasets MAY omit create via CLI if the primary flow is through the SDK. Commands SHALL only add verbs that make sense for the resource.

Additional verbs for specialized actions are RECOMMENDED when the standard set doesn't cover it:

• px auth login, px auth status
• px self update
• px docs fetch
• px api graphql <query>

Backward compatibility during migration

The CLI is evolving from a flat structure (px projects, px traces) toward full noun-verb. During the transition, both forms MAY coexist. When migrating an existing command:

1. The new noun-verb form MUST be created as the primary command
2. The old form SHOULD be kept as a hidden alias (Commander's .alias() or a hidden command) so existing scripts don't break
3. Only the noun-verb form SHALL be documented going forward

Dual-Audience Design

The CLI MUST be equally usable by a person at a terminal and by a coding agent. Every command that outputs data MUST support --format:

• pretty (default) — Human-readable tables and formatting
• json — Indented JSON for human inspection of structured data
• raw — Compact single-line JSON for piping into jq or agent consumption

Commands MAY support additional formats (e.g., --format text for prompts). The default MUST always be pretty.

Progress indicators MUST write to stderr. Agents SHOULD pass --no-progress to suppress them.

# Agent-friendly invocation
px trace list --format raw --no-progress | jq '...'

Semantic exit codes

Defined in src/exitCodes.ts. Commands MUST use the named constants and MUST NOT use bare numeric literals.

Interactive default with non-interactive mode

Commands MAY prompt interactively when a required value is missing or a confirmation is needed — this is the human-friendly default. The --no-input flag MUST suppress all prompts: both missing-value prompts and destructive-action confirmations. Non-interactive mode is also activated automatically when no TTY is attached (piped stdin).

In non-interactive mode, if a required value is missing, the command MUST exit immediately with ExitCode.INVALID_ARGUMENT and print the correct invocation. If a confirmation would have been shown, the command MUST proceed as if confirmed:

# Human: missing --name triggers interactive prompt
px project create

# Agent: all inputs as flags, no prompts
px project create --name my-project --format raw --no-input

# Human: gets "Are you sure?" prompt
px dataset delete my-dataset

# Agent: skips confirmation
px dataset delete my-dataset --no-input --format raw

Error: Missing required flag --name.
  px project create --name <project-name>

Idempotent commands

Commands SHOULD be idempotent where possible. Running the same command twice MUST NOT produce duplicate resources or unexpected errors:

• create commands SHOULD support --if-not-exists to return the existing resource instead of failing
• delete commands on a missing resource SHOULD exit with ExitCode.SUCCESS (not an error)

Return structured data on success

Mutating commands (create, update, delete) MUST return the affected resource in the selected --format on stdout. Commands MUST NOT print bare success messages like "Project created." — output the resource so agents can extract IDs, URLs, and other fields:

$ px project create --name foo --format raw
{"id":"proj_abc","name":"foo","createdAt":"2025-03-15T10:00:00Z"}

Fail fast with actionable errors

When a command fails due to invalid input, the error message MUST include the correct invocation syntax. Follow the pattern established by getConfigErrorMessage() in src/config.ts:

• Show what was wrong
• Show the correct command to fix it
• Suggest a related command when helpful (e.g., "Available projects: px project list")

When --format raw or --format json is active, errors SHOULD also be written as structured JSON to stderr so agents can parse them. Use the StructuredError shape:

interface StructuredError {
  error: string; // Human-readable error message
  code: string; // ExitCode constant name (e.g., "INVALID_ARGUMENT")
  hint?: string; // Suggested command to resolve the issue
}

{
  "error": "Project not found",
  "code": "FAILURE",
  "hint": "px project list --format raw"
}

Progressive help discovery

Every subcommand MUST include a --help with at least one concrete example. Use Commander's .addHelpText('after', ...) to append an Examples: block:

$ px project create --help
Usage: px project create [options]

Create a new Phoenix project.

Options:
  --name <name>       Project name (required)
  --description <d>   Project description
  --format <format>   Output format (pretty|json|raw) (default: "pretty")
  -h, --help          Display help

Examples:
  px project create --name my-project
  px project create --name my-project --format raw

Agents discover capabilities incrementally: px → px project → px project create --help. Each level MUST provide enough information to navigate deeper.

Adding a New Command

Options interface

Every handler MUST define a TypeScript interface for its options — the typed mirror of the flags its command registers. Field names MUST be descriptive; abbreviations MUST NOT be used.

The shapes that recur across commands live in src/commands/options.ts and MUST be reused rather than re-declared:

A command whose flags exactly match a shared shape MUST use that type directly instead of declaring an identical local one. Otherwise, extend the matching base and declare only the flags unique to the command:

interface SpanListOptions
  extends ProjectScopedOptions<SpanOutputFormat>, AnnotationInclusionOptions {
  /**
   * `--span-kind <kinds...>`: Filter by OpenInference span kind — `LLM`,
   * `CHAIN`, `TOOL`, …
   *
   * @example ["LLM", "TOOL"]
   */
  spanKind?: string[];
}

Every field MUST carry a TSDoc block that opens with the literal flag and its argument placeholder, states any default declared in the .option() call, and gives an @example where a concrete value clarifies usage. Negated flags (--no-progress, --no-input) MUST document their inverted reading: Commander defaults them to true and only sets false when the flag is passed.

Commands that prompt for input or confirmation MUST support non-interactive mode — --no-input for prompts (px setup), --yes via ConfirmationOptions for destructive confirmations.