feat(amp-send,amp-reply): --body-file and --body-stdin for rich content

[joceqo]

Problem

amp-send and amp-reply take the message body as a positional shell
argument. This breaks for rich content even when the caller carefully
escapes things:

• Backticks (interpreted as command substitution)
• Fenced code blocks with embedded $var or \ sequences
• Box-drawing characters (occasional Unicode handling issues)
• Multi-line bodies with embedded quotes

Real-world failure mode

Observed in a multi-agent orchestration session: an AI agent was asked
to reply with a CLI UI mockup containing fenced code blocks with
backticks and box-drawing chars. The agent's
amp-reply <msg-id> "<rich content>" call entered a stuck loop in zsh.
The agent attempted multiple workarounds (writing the body to a temp
file, escaping backticks individually, switching quoting styles) and
never recovered. Other agents in the same swarm — sending plain text —
delivered fine. The friction is specifically rich content + positional
shell arg.

Fix

Both scripts now accept the message body via three mutually-exclusive
sources:

Source	Flag	Use case
Positional argument	(default)	Plain text, short messages
--body-file PATH	new	Pre-composed rich content (Markdown, mockups, code)
--body-stdin	new	Piped from another tool or heredoc

If more than one source is provided, the script errors with a clear
message. If none is provided, it falls through to the existing
"missing body" path with help shown.

Existing positional usage is unchanged — these options are purely
additive. No breaking changes.

Verified end-to-end

Rich content (backticks, fenced code blocks, box-drawing chars ┌──┐,
Unicode ★, and $var-looking sequences) preserved byte-for-byte at
the recipient via both --body-file and --body-stdin. Plain
positional usage continues to work.

amp-send alice "Subject" --body-file report.md
cat reply.md | amp-send alice "Subject" --body-stdin
amp-reply msg_xyz --body-file rich-reply.md

ShellCheck

Both patched scripts pass shellcheck. Only pre-existing info-level
warnings remain (SC1091 about following the helper file, SC2181
in amp-send.sh:293 about indirect $? check) — both unrelated to
this change.

Why this matters for AI-driven AMP

Multi-agent orchestrations (especially LLM agents writing bash) are
exactly the case where positional shell args fail unpredictably. A
file-based body is far easier for an agent to compose reliably:
write to /tmp/reply-$$.md, then call amp-reply <msg-id> --body-file /tmp/reply-$$.md.
This makes amp-reply / amp-send safe to use as a target API for
agent code generation.

🤖 Generated with Claude Code