Skip to content

agentyper

Agent-first Python CLI library — Typer-compatible, built on argparse + pydantic GitHub: https://github.com/romamo/agentyper Version: 0.1.4 (Alpha) | License: MIT | Language: Python 3.10+ Author: Roman Medvedev | First release: 2026-03-03

Overview

agentyper is an early-stage Python CLI framework explicitly designed for agent-first CLI ergonomics. It is API-compatible with Typer (drop-in replacement via import agentyper as typer) but built on argparse + pydantic instead of Click. Its central thesis: existing CLI frameworks were designed for human operators; agentyper is designed for the era when AI agents are the primary CLI consumer.

Key design decision: agent ergonomics are injected automatically at the framework level — command authors write normal Python functions, and the framework handles --schema, --format, --yes, --answers, structured errors, and exit code taxonomy without any per-command work.

Architecture & Design

Stack: - Parser: argparse (stdlib — no Click dependency) - Validation: pydantic v2 (type coercion, field-level errors, schema export) - Output: rich (human display) + structured JSON (agent display) - TTY detection: isatty() for automatic format selection

Core components:

src/agentyper/
  _internal/
    _app.py       — Agentyper app, command registration, routing
    _errors.py    — Exit codes, structured error formatting
    _interactive.py — confirm(), prompt(), edit() with agent bypass
    _output.py    — output(), echo(), render_output()
    _params.py    — Option, Argument (Typer-compatible)
    _schema.py    — JSON Schema generation from function signatures
  __init__.py     — Public API
  cli.py          — CLI entry point
  testing.py      — Test helpers

Auto-injected global flags (every command gets these for free): - --format {table,json,csv} — structured output, auto-selected via isatty() - --schema — JSON Schema for command params or full app - --yes / --no — bypass confirm() calls non-interactively - --answers '{"confirms":[...],"prompts":[...]}' — pre-supply all interactive answers - -v / -vv — INFO/DEBUG logging

Exit code taxonomy (3-code system):

EXIT_SUCCESS    = 0  — success
EXIT_VALIDATION = 1  — bad input, agent should retry with correction
EXIT_SYSTEM     = 2  — system error, agent should abort

Typer migration path:

# Before:
import typer
# After (one line change):
import agentyper as typer  # everything else stays identical

Agent Compatibility Assessment

What it handles natively

  • Schema discoverability — --schema on every command and the full app, exports JSON Schema derived from Python type hints and docstrings
  • Output format — --format json/csv/table, auto-detected via isatty(); non-TTY defaults to machine-readable
  • Interactivity bypass — confirm(), prompt(), edit() all resolve without blocking in non-TTY/agent mode via --yes and --answers
  • Structured errors — exit_error() emits structured JSON error objects; pydantic validation errors are formatted into field-level structured output automatically
  • Exit code taxonomy — 3-code system (0/1/2) with named constants
  • Validation before execution — pydantic validates all arguments before the command function body executes

What it handles partially

  • Exit codes — 3-code system (0/1/2) is simpler than the 9-code table in the requirements (no distinct codes for NOT_FOUND, TIMEOUT, CONFLICT, RATE_LIMITED, etc.)
  • Output format — JSON output exists but no ok/data/error/meta envelope schema; no pagination metadata; no warnings[] array
  • ANSI suppression — uses rich which has NO_COLOR support, but not all edge cases of color leakage are documented as handled
  • Stderr/stdout discipline — echo() vs output() separation exists but enforcement is opt-in by command authors

What it does not handle

  • Timeouts (--timeout flag, timeout exit code 7)
  • Idempotency keys (--idempotency-key)
  • Partial failure / multi-step step manifests
  • Pagination (--limit, --cursor, pagination metadata in responses)
  • Signal handling (SIGTERM handler, partial JSON on kill)
  • Child process tracking
  • Config file shadowing / --show-config
  • Retry hints (retryable, retry_after_ms in errors)
  • Observability / audit trail / request IDs
  • Dry-run for destructive operations
  • Network proxy awareness
  • Binary/encoding safety
  • Working directory sensitivity
  • Undeclared filesystem side effects

Challenge Coverage Table

# Challenge Rating Reason
1 Exit Codes & Status Signaling ~ 3-code taxonomy (0/1/2) — partial; missing NOT_FOUND, TIMEOUT, etc.
2 Output Format & Parseability ~ --format json automatic; no standard envelope, no pagination meta
3 Stderr vs Stdout Discipline ~ output() vs echo() split exists; not strictly enforced on authors
4 Verbosity & Token Cost ~ --format csv is 4× cheaper than table; no --fields selector
5 Pagination & Large Output ✗ No --limit, --cursor, or pagination metadata
6 Command Composition & Piping ✗ No --output id mode, no stdin pipe support
7 Output Non-Determinism ✗ No stable sorting guarantee, no data/meta separation
8 ANSI & Color Code Leakage ~ rich respects NO_COLOR; isatty() detection; not exhaustively tested
9 Binary & Encoding Safety ✗ No explicit binary/encoding handling
10 Interactivity & TTY Requirements ✓ --yes, --answers, isatty() auto-detection — core feature
11 Timeouts & Hanging Processes ✗ No built-in timeout mechanism
12 Idempotency & Safe Retries ✗ No idempotency key support
13 Partial Failure & Atomicity ✗ No step manifests or partial failure reporting
14 Argument Validation Before Side Effects ✓ Pydantic validates all args before function body runs
15 Race Conditions & Concurrency ✗ No session isolation or locking primitives
16 Signal Handling & Graceful Cancellation ✗ No SIGTERM/SIGPIPE handlers installed
17 Child Process Leakage ✗ No child process tracking
18 Error Message Quality ✓ Structured JSON errors with field-level pydantic details
19 Retry Hints in Error Responses ✗ No retryable or retry_after_ms fields
20 Environment & Dependency Discovery ✗ No doctor command or preflight checks
21 Schema & Help Discoverability ✓ --schema on every command and full app — core feature
22 Schema Versioning & Output Stability ✗ No schema_version in responses, no deprecation warnings
23 Side Effects & Destructive Operations ✗ No danger_level, no --dry-run framework support
24 Authentication & Secret Handling ✗ No secret redaction or --secret-from-file
25 Prompt Injection via Output ✗ No external data tagging or sanitization
26 Stateful Commands & Session Management ✗ No --config isolation or state inspection
27 Platform & Shell Portability ~ Python stdlib base is portable; no explicit platform checks
28 Config File Shadowing & Precedence ✗ No config precedence tracking or --no-config
29 Working Directory Sensitivity ✗ No --cwd flag, no absolute path enforcement
30 Undeclared Filesystem Side Effects ✗ No side effect declaration or cleanup command
31 Network Proxy Unawareness ✗ No proxy handling (network is out of scope)
32 Self-Update & Auto-Upgrade Behavior ✓ Early-stage tool; no auto-update behavior implemented
33 Observability & Audit Trail ✗ No request IDs, trace IDs, or audit log

Summary: ✓ 5 / ~ 6 / ✗ 22

Strengths for Agent Use

  1. Drop-in Typer replacement — lowest migration cost of any framework; massive Typer ecosystem becomes agent-compatible with one import change
  2. Schema-first — --schema is automatic on every command, not an afterthought
  3. Interactivity solved — --answers JSON payload is a uniquely powerful pattern for pre-supplying all interactive responses
  4. Pydantic validation — field-level structured errors are immediately useful to agents for self-correction
  5. isatty() auto-detection — format selection without agent configuration
  6. Active agent focus — the only Python CLI framework explicitly designed around agent ergonomics

Weaknesses for Agent Use

  1. Alpha quality (v0.1.4, March 2026) — not production-ready; API may change
  2. 3-code exit taxonomy — too coarse; agents can't distinguish NOT_FOUND from TIMEOUT from RATE_LIMITED
  3. No output envelope — JSON output lacks ok/data/error/meta structure; agents must handle ad-hoc schemas per command
  4. Execution gaps — no timeouts, signals, partial failure, pagination, idempotency
  5. Security gaps — no secret redaction, no prompt injection protection
  6. No observability — no request IDs, no audit trail
  7. Rich dependency — pulls in rich even for agent use where all color output should be suppressed

Verdict

agentyper is the most conceptually aligned framework in this review — it explicitly targets agent ergonomics as its primary design goal, not a retrofit. Its auto-injected --schema, --format, --yes/--answers, and structured errors address the highest-frequency agent pain points. However, at v0.1.4 it covers only ~17% of the 33 challenges, leaving critical execution reliability (timeouts, signals, idempotency), security (secrets, prompt injection), and observability entirely unaddressed. As a foundation for a more complete agent-compatible CLI framework, it is the strongest starting point in the Python ecosystem — but it requires significant extension before it meets the full requirements.