Python

Agent-First Data (AFDATA) — Suffix-driven output formatting and protocol templates for AI agents.

The field name is the schema. Agents read latency_ms and know milliseconds, api_key_secret and know to redact, no external schema needed.

Installation

pip install agent-first-data

Quick Example

A backup tool invoked from the CLI — flags, env vars, and config all use the same suffixes:

API_KEY_SECRET=sk-1234 cloudback --timeout-s 30 --max-file-size-bytes 10737418240 /data/backup.tar.gz

For CLI diagnostics, enable log categories explicitly:

--log startup,request,progress,retry,redirect
--verbose   # shorthand for all categories

Without these flags, startup diagnostics should stay off by default.

The tool reads env vars, flags, and config — all with AFDATA suffixes — and can emit a startup diagnostic event:

from agent_first_data import *
import os

startup = build_json(
    "log",
    {
        "event": "startup",
        "config": {"timeout_s": 30, "max_file_size_bytes": 10737418240},
        "args": {"input_path": "/data/backup.tar.gz"},
        "env": {"API_KEY_SECRET": os.environ.get("API_KEY_SECRET")},
    },
    trace=None,
)

Three output formats, same data:

JSON:  {"code":"log","event":"startup","args":{"input_path":"/data/backup.tar.gz"},"config":{"max_file_size_bytes":10737418240,"timeout_s":30},"env":{"API_KEY_SECRET":"***"}}
YAML:  code: "log"
       event: "startup"
       args:
         input_path: "/data/backup.tar.gz"
       config:
         max_file_size: "10.0GB"
         timeout: "30s"
       env:
         API_KEY: "***"
Plain: args.input_path=/data/backup.tar.gz code=log event=startup config.max_file_size=10.0GB config.timeout=30s env.API_KEY=***

--timeout-s → timeout_s → timeout: 30s. API_KEY_SECRET → API_KEY: "***". The suffix is the schema.

API Reference

Total: 15 public APIs and 2 types + AFDATA logging (3 protocol builders + 2 redacted value helpers + 4 output functions + 1 internal + 1 utility + 4 CLI helpers + OutputFormat + RedactionPolicy)

Protocol Builders (returns dict)

Build AFDATA protocol structures. Return dict objects for transport payloads.

# Success (result)
build_json_ok(result: Any, trace: Any = None) -> dict

# Error (simple message, optional hint)
build_json_error(message: str, hint: str = None, trace: Any = None) -> dict

# Generic (any code + fields)
build_json(code: str, fields: Any, trace: Any = None) -> dict

Redacted Values (returns Any)

Use these before raw HTTP/MCP/SSE serializers that do not call output_json.

redacted_value(value: Any) -> Any
redacted_value_with(value: Any, redaction_policy: RedactionPolicy) -> Any

Use case: structured protocol payloads (frameworks automatically serialize)

Example:

from agent_first_data import *

# Startup
startup = build_json(
    "log",
    {
        "event": "startup",
        "config": {"api_key_secret": "sk-123", "timeout_s": 30},
        "args": {"config_path": "config.yml"},
        "env": {"RUST_LOG": "info"},
    },
    trace=None,
)

# Success (always include trace)
response = build_json_ok(
    {"user_id": 123},
    trace={"duration_ms": 150, "source": "db"},
)

# Error
err = build_json_error("user not found", trace={"duration_ms": 5})

# Error with hint
err_hint = build_json_error("wallet not found", hint="list wallets with: afpay wallet list", trace={"duration_ms": 5})

# Specific error code
not_found = build_json(
    "not_found",
    {"resource": "user", "id": 123},
    trace={"duration_ms": 8},
)

CLI/Log Output (returns str)

Format values for CLI output and logs. output_json uses full _secret redaction by default. output_json_with supports explicit scoped policies. YAML and Plain always redact _secret and apply human-readable formatting.

output_json(value: Any) -> str   # Single-line JSON, original keys, for programs/logs
output_json_with(value: Any, redaction_policy: RedactionPolicy) -> str
output_yaml(value: Any) -> str   # Multi-line YAML, keys stripped, values formatted
output_plain(value: Any) -> str  # Single-line logfmt, keys stripped, values formatted

class RedactionPolicy(enum.Enum):
    RedactionTraceOnly = "RedactionTraceOnly"
    RedactionNone = "RedactionNone"
    RedactionStrict = "RedactionStrict"

Example:

from agent_first_data import *

data = {
    "user_id": 123,
    "api_key_secret": "sk-1234567890abcdef",
    "created_at_epoch_ms": 1738886400000,
    "file_size_bytes": 5242880,
}

# JSON (secrets redacted, original keys, raw values)
print(output_json(data))
# {"api_key_secret":"***","created_at_epoch_ms":1738886400000,"file_size_bytes":5242880,"user_id":123}

# YAML (keys stripped, values formatted, secrets redacted)
print(output_yaml(data))
# ---
# api_key: "***"
# created_at: "2025-02-07T00:00:00.000Z"
# file_size: "5.0MB"
# user_id: 123

# Plain logfmt (keys stripped, values formatted, secrets redacted)
print(output_plain(data))
# api_key=*** created_at=2025-02-07T00:00:00.000Z file_size=5.0MB user_id=123

Internal Tools

internal_redact_secrets(value: Any) -> None  # Manually redact secrets in-place

Most users don’t need this. Output functions automatically protect secrets.

Utility Functions

parse_size(s: str) -> int | None  # Parse "10M" → bytes

Returns None for invalid, negative, or overflow input.

Example:

from agent_first_data import *

assert parse_size("10M") == 10485760
assert parse_size("1.5K") == 1536
assert parse_size("512") == 512

CLI Helpers (for tools built on AFDATA)

Shared helpers that prevent flag-parsing drift between CLI tools. Use these instead of reimplementing --output and --log handling in each tool.

class OutputFormat(enum.Enum):  # JSON="json", YAML="yaml", PLAIN="plain"

cli_parse_output(s: str) -> OutputFormat         # Parse --output flag; raises ValueError on unknown
cli_parse_log_filters(entries: list[str]) -> list[str]  # Normalize --log: trim, lowercase, dedup, remove empty
cli_output(value: Any, format: OutputFormat) -> str     # Dispatch to output_json/yaml/plain
build_cli_error(message: str, hint: str = None) -> dict  # {code:"error", error_code:"invalid_request", hint?, retryable:False, trace:{duration_ms:0}}

Canonical pattern — parse all flags before doing work, emit JSONL errors to stdout:

import sys
from agent_first_data import (
    OutputFormat, cli_parse_output, cli_parse_log_filters,
    cli_output, build_cli_error, output_json,
)

try:
    fmt = cli_parse_output(args.output)
except ValueError as e:
    print(output_json(build_cli_error(str(e))))
    sys.exit(2)

log = cli_parse_log_filters(args.log.split(",") if args.log else [])
# ... do work ...
print(cli_output(result, fmt))

See examples/agent_cli.py for the complete working example (pytest examples/agent_cli.py).

Usage Examples

Example 1: REST API

from agent_first_data import *
from fastapi import FastAPI

app = FastAPI()

@app.get("/users/{user_id}")
async def get_user(user_id: int):
    response = build_json_ok(
        {"user_id": user_id, "name": "alice"},
        trace={"duration_ms": 150, "source": "db"},
    )
    # API returns raw JSON — no output processing, no key stripping
    return response

Example 2: CLI Tool (Complete Lifecycle)

from agent_first_data import *

# 1. Startup
startup = build_json(
    "log",
    {
        "event": "startup",
        "config": {"api_key_secret": "sk-sensitive-key", "timeout_s": 30},
        "args": {"input_path": "data.json"},
        "env": {"RUST_LOG": "info"},
    },
    trace=None,
)
print(output_yaml(startup))
# ---
# code: "log"
# event: "startup"
# args:
#   input_path: "data.json"
# config:
#   api_key: "***"
#   timeout: "30s"
# env:
#   RUST_LOG: "info"

# 2. Progress
progress = build_json(
    "progress",
    {"current": 3, "total": 10, "message": "processing"},
    trace={"duration_ms": 1500},
)
print(output_plain(progress))
# code=progress current=3 message=processing total=10 trace.duration=1.5s

# 3. Result
result = build_json_ok(
    {
        "records_processed": 10,
        "file_size_bytes": 5242880,
        "created_at_epoch_ms": 1738886400000,
    },
    trace={"duration_ms": 3500, "source": "file"},
)
print(output_yaml(result))
# ---
# code: "ok"
# result:
#   created_at: "2025-02-07T00:00:00.000Z"
#   file_size: "5.0MB"
#   records_processed: 10
# trace:
#   duration: "3.5s"
#   source: "file"

Example 3: JSONL Output

from agent_first_data import *

result = build_json_ok(
    {"status": "success"},
    trace={"duration_ms": 250, "api_key_secret": "sk-123"},
)

# Print JSONL to stdout (secrets redacted, one JSON object per line)
# Channel policy: machine-readable protocol/log events must not use stderr.
print(output_json(result))
# {"code":"ok","result":{"status":"success"},"trace":{"api_key_secret":"***","duration_ms":250}}

Complete Suffix Example

from agent_first_data import *

data = {
    "created_at_epoch_ms": 1738886400000,
    "request_timeout_ms": 5000,
    "cache_ttl_s": 3600,
    "file_size_bytes": 5242880,
    "payment_msats": 50000000,
    "price_usd_cents": 9999,
    "success_rate_percent": 95.5,
    "api_key_secret": "sk-1234567890abcdef",
    "user_name": "alice",
    "count": 42,
}

# YAML output (keys stripped, values formatted, secrets redacted)
print(output_yaml(data))
# ---
# api_key: "***"
# cache_ttl: "3600s"
# count: 42
# created_at: "2025-02-07T00:00:00.000Z"
# file_size: "5.0MB"
# payment: "50000000msats"
# price: "$99.99"
# request_timeout: "5.0s"
# success_rate: "95.5%"
# user_name: "alice"

# Plain logfmt output (same transformations, single line)
print(output_plain(data))
# api_key=*** cache_ttl=3600s count=42 created_at=2025-02-07T00:00:00.000Z file_size=5.0MB payment=50000000msats price=$99.99 request_timeout=5.0s success_rate=95.5% user_name=alice

AFDATA Logging

AFDATA-compliant structured logging via Python’s logging module. Every log line is formatted using the library’s own output_json/output_plain/output_yaml functions. Span fields are carried via contextvars (async-safe), automatically flattened into each log line.

API

from agent_first_data import init_logging_json, init_logging_plain, init_logging_yaml
from agent_first_data.afdata_logging import AfdataHandler, get_logger, span

# Convenience initializers — set up the root logger with AFDATA output to stdout
init_logging_json(level="INFO")    # Single-line JSONL (secrets redacted, original keys)
init_logging_plain(level="INFO")   # Single-line logfmt (keys stripped, values formatted)
init_logging_yaml(level="INFO")    # Multi-line YAML (keys stripped, values formatted)

# Low-level — create a handler for custom logger stacks
AfdataHandler(format="json")  # format: "json" | "plain" | "yaml"

# Logger with default fields (returns logging.LoggerAdapter)
get_logger(name, **fields)

# Span context manager — adds fields to all log events within the block
span(**fields)

Setup

from agent_first_data import init_logging_json, init_logging_plain, init_logging_yaml

# JSON output for production (one JSONL line per event, secrets redacted)
init_logging_json("INFO")

# Plain logfmt for development (keys stripped, values formatted)
init_logging_plain("DEBUG")

# YAML for detailed inspection (multi-line, keys stripped, values formatted)
init_logging_yaml("DEBUG")

Log Output

Standard logging calls work unchanged. Output format depends on the init function used.

import logging
logger = logging.getLogger("myapp")

logger.info("Server started")
# JSON:  {"timestamp_epoch_ms":1739000000000,"message":"Server started","target":"myapp","code":"info"}
# Plain: code=info message="Server started" target=myapp timestamp_epoch_ms=1739000000000
# YAML:  ---
#        code: "info"
#        message: "Server started"
#        target: "myapp"
#        timestamp_epoch_ms: 1739000000000

logger.warning("DNS lookup failed")
# JSON:  {"timestamp_epoch_ms":...,"message":"DNS lookup failed","target":"myapp","code":"warn"}

Span Support

Use the span context manager to add fields to all log events within the block. Spans nest and work with both sync and async code.

from agent_first_data import span

with span(request_id="abc-123"):
    logger.info("Processing")
    # {"timestamp_epoch_ms":...,"message":"Processing","target":"myapp","request_id":"abc-123","code":"info"}

    with span(step="validate"):
        logger.info("Validating input")
        # {"timestamp_epoch_ms":...,"message":"Validating input","target":"myapp","request_id":"abc-123","step":"validate","code":"info"}

Logger with Default Fields

Use get_logger for per-component fields that appear on every log line:

from agent_first_data import get_logger

logger = get_logger("myapp.auth", component="auth")
logger.info("Token verified")
# {"timestamp_epoch_ms":...,"message":"Token verified","target":"myapp.auth","component":"auth","code":"info"}

Custom Code Override

The code field defaults to the log level. Override with an explicit field:

from agent_first_data import get_logger

logger = get_logger("myapp")
logger.info("Server ready", extra={"code": "log", "event": "startup"})
# {"timestamp_epoch_ms":...,"message":"Server ready","target":"myapp","code":"log","event":"startup"}

Output Fields

Every log line contains:

Log Output Formats

All three formats use the library’s own output functions, so AFDATA suffix processing applies to log fields too:

All formats automatically redact _secret fields in log output.

Output Formats

Three output formats for different use cases:

All formats automatically redact _secret fields.

Supported Suffixes

• Duration: _ms, _s, _ns, _us, _minutes, _hours, _days
• Timestamps: _epoch_ms, _epoch_s, _epoch_ns, _rfc3339
• Size: _bytes (auto-scales to KB/MB/GB/TB), _size (config input, pass through)
• Currency: _msats, _sats, _btc, _usd_cents, _eur_cents, _jpy, _{code}_cents
• Other: _percent, _secret (auto-redacted in all formats)

Repository

This package is part of the agent-first-data repository, which also contains:

• spec/ — Full AFDATA specification with suffix definitions, protocol format rules, and cross-language test fixtures
• skills/ — AI coding agent skill for working with AFDATA conventions

To run tests, clone the full repository (tests use shared cross-language fixtures from spec/fixtures/):

git clone https://github.com/agentfirstkit/agent-first-data
cd agent-first-data/python
python -m pytest

License

MIT