fya

fya is a Claude Code print-mode wrapper backed by an interactive PTY session.

It provides practical parity with claude -p for callers that rely on print-mode text, JSON, or stream-json output. Internally, it starts interactive claude in a hidden PTY, types the prompt into the terminal, tails Claude Code transcript logs, and emits Claude-compatible output.

fya mirrors the one-shot claude -p contract: one process invocation accepts one prompt, streams one answer, emits one final result, and exits after cleanup.

Requirements

• Go 1.26 to build from source
• Claude Code installed as claude
• Unix/macOS for PTY support

Windows returns an unsupported PTY error in v1.

Installation

Homebrew:

brew install umputun/apps/fya

Binary releases: download from GitHub Releases for linux/darwin amd64/arm64.

From source:

git clone https://github.com/umputun/fya.git
cd fya
make build

The binary is written to .bin/fya.

Usage

fya accepts the print-mode shape expected by tools that invoke Claude Code:

printf 'say hello\n' | fya --print --output-format=stream-json

A positional prompt also works without stdin and takes precedence over stdin:

fya --print "say hello"

Supported consumed compatibility flags:

• -p, --print
• --output-format=text|json|stream-json
• --input-format=text|stream-json
• --json-schema=SCHEMA
• --replay-user-messages

Wrapper controls:

• --cwd=PATH - working directory for the interactive Claude session, default .
• --idle-timeout=DURATION - transcript idle duration before completion, default 2s
• --turn-timeout=DURATION - maximum wall-clock duration for one turn, default 30m
• --gate - unattended gate profile; aborts a turn after 5m without transcript activity, measured from the last transcript write rather than from turn start
• fya turn-timeout and no-activity failures include FYA_TRANSIENT_TIMEOUT (with terminal_reason: "fya_turn_timeout" or "fya_no_activity_timeout") so orchestration tools can retry them as transient Claude continuation stalls
• --typing-wpm=N - prompt typing speed in words per minute, default 100
• --typing-jitter=FLOAT - per-character delay jitter ratio, default 0.20 (0 disables jitter)
• --max-wpm-size=N - paste the prompt in one write instead of typing it when the prompt is longer than N words, default 100. 0 always types rune-by-rune. Pasting avoids the multi-minute typing latency of large prompts; typing keeps shorter prompts arriving as individual keystrokes.
• --readiness-timeout=DURATION - maximum wait for Claude input readiness, default 30s
• --silent - accepted for compatibility; fya does not emit synthetic tool-progress text
• --dbg - enable fya debug logging. Named --dbg so it does not collide with Claude's own --debug flag, which is forwarded to Claude.

Recognized Claude launch flags are forwarded to interactive claude, including --dangerously-skip-permissions, --verbose, --model, --effort, permission/tool flags, MCP/config flags, and related interactive Claude flags. Unknown flags fail fast instead of being forwarded.

Structured JSON Output

--json-schema=SCHEMA is for callers that expect native Claude's structured-output JSON envelope. In v1 it is intentionally narrow:

• requires --output-format=json
• requires text input (--input-format=text, the default)
• does not support stream-json schema mode
• does not retry or ask Claude to repair invalid output

SCHEMA is the JSON Schema text. fya compiles it before starting Claude, appends a schema-output instruction to the prompt, then validates the final assistant text as exactly one JSON value.

On success, the final JSON result keeps result as Claude's raw assistant text and adds top-level structured_output as a JSON value, not a quoted JSON string:

{"type":"result","subtype":"success","is_error":false,"result":"{\"summary\":\"done\"}","structured_output":{"summary":"done"},"session_id":"...","num_turns":1,"terminal_reason":"end_turn"}

If Claude's final answer is not valid JSON or does not match the schema, fya emits a valid JSON error result, exits non-zero, and does not emit structured_output. The validation error result uses terminal_reason: "fya_structured_output_invalid". Existing startup, timeout, cancellation, transcript, and Claude-exit error results keep their original failure reason and bypass structured-output validation.

Environment

• FYA_CLAUDE_DIR - override Claude's config/transcript root for fya transcript discovery. Defaults to ~/.claude. This is not forwarded to child Claude.
• DEBUG - alias for --dbg, enables fya debug logging when set to any non-empty value. This is not forwarded to child Claude.
• ANTHROPIC_API_KEY - preserved for the child Claude process when present.
• CLAUDECODE - stripped from the child process to avoid nested Claude Code session errors.
• FYA, FYA_*, and shell _ - stripped from the child process so wrapper-specific environment and command-path details are not leaked through normal environment inspection.

Ralphex

fya can be used in Ralphex as the claude_command while keeping the usual Claude-compatible arguments:

claude_command = /path/to/fya
claude_args = --dangerously-skip-permissions --output-format stream-json --verbose

Ralphex passes the prompt on stdin and appends --print. fya consumes print/output flags itself, forwards interactive Claude launch flags such as --dangerously-skip-permissions, --verbose, --model, and --effort, and writes JSONL to stdout.

Invoke fya by absolute path in consumers that also need the real Claude binary. Do not replace claude on PATH with fya or put a fya shim named claude before the real binary; fya starts its child process by running claude from PATH.

Optional Parity Check

The live parity script compares claude -p stream-json output with fya. Dry-run mode does not call Claude:

scripts/compare-claude-p.sh --dry-run

Live mode consumes Claude quota:

make build
scripts/compare-claude-p.sh

Live downstream-integration smoke tests are also optional and manual because they consume Claude quota. For those checks, configure the consuming tool to invoke fya by absolute path while leaving the real claude child binary on PATH, then run a disposable job.

Architecture

• app/ - executable composition root
• scripts/ - optional local validation helpers
• docs/plans/completed/ - completed implementation plans

See ARCHITECTURE.md for the PTY flow, transcript tailing, readiness detection, cleanup, and output contract.

Known Limitations

• v1 supports Unix/macOS PTYs only. Windows returns an unsupported PTY error.
• Transcript parsing follows the current Claude Code JSONL shapes used by the implementation tests. It is not byte-for-byte parity with every Claude stream event.
• stream-json emits Claude-style assistant/user message events from the transcript plus one final result containing the accumulated assistant answer. fya relays native message-shaped events and does not synthesize tool: text progress.
• Multi-turn --input-format=stream-json history is outside the one-shot design. Exactly one user message is accepted.
• --json-schema is v1-only structured-output compatibility for --output-format=json with text input; it has no stream-json schema mode and no automatic retry.
• Live parity and downstream-integration smoke checks are manual because they call Claude and consume quota.

Development

make test
make lint
make build