hermes - 💡(How to fix) Fix [Bug]: WhatsApp platform hint instructs model not to use markdown, but adapter auto-converts it [3 comments, 2 participants]

Bug Description

The WhatsApp PLATFORM_HINT in agent/prompt_builder.py tells the model: "Please do not use markdown as it does not render."

However, gateway/platforms/whatsapp.py ships a format_message() method that converts standard markdown to WhatsApp's native syntax on every send:

• **bold** → *bold*
• ~~strike~~ → ~strike~
• # Header → *Header*
• [text](url) → text (url)

The platform hint and the adapter contradict each other: the model is told to avoid markdown, while the adapter is built to convert it. As a result, WhatsApp replies arrive flat — no bold on key figures, headers, or alerts — even when the agent prompt (SOUL.md) explicitly asks for bold.

This was deliberately fixed for Telegram in #10612 (closes #8261). The same fix has not been applied to WhatsApp, even though the converter and its tests already exist (tests/gateway/test_whatsapp_formatting.py).

Steps to Reproduce

1. Configure Hermes with the WhatsApp gateway.
2. Set up an agent prompt (SOUL.md or system prompt) that requests bold on key figures, e.g. "use bold on key numbers and alerts".
3. Ask any question whose ideal reply contains numbers or section headers.
4. Observe the message rendered in WhatsApp.

Expected Behavior

The model emits standard markdown (bold, strike, # Header). format_message() converts it to WhatsApp syntax before sending. The user sees properly bolded figures and headers in WhatsApp, matching the behavior already shipped for Telegram.

Actual Behavior

The model obeys the platform hint ("do not use markdown") and emits plain text. The converter has nothing to convert. Replies arrive in WhatsApp without any visual emphasis. Overriding the hint via SOUL.md does not reliably win, because PLATFORM_HINTS is appended LAST in the system prompt (run_agent.py:4901) — recency favors the upstream hint over earlier overrides.

Affected Component

Agent Core (conversation loop, context compression, memory)

Messaging Platform (if gateway-related)

WhatsApp

Debug Report

Report       https://paste.rs/c9Sr4
agent.log    https://paste.rs/wJSx9
gateway.log  https://paste.rs/7XFqi

Operating System

Ubuntu 24.04.4 LTS

Python Version

Python 3.12.3

Hermes Version

v0.11.0 (2026.4.23)

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

Three pieces of upstream code make this a structural contradiction:

1. agent/prompt_builder.py:298-307 — the hint:

"whatsapp": (
    "You are on a text messaging communication platform, WhatsApp. "
    "Please do not use markdown as it does not render. "
    ...
),

1. gateway/platforms/whatsapp.py:607-662 — the converter (excerpt):

def format_message(self, content: str) -> str:
    """Convert standard markdown to WhatsApp-compatible formatting."""
    ...
    result = re.sub(r"\*\*(.+?)\*\*", r"*\1*", result)
    result = re.sub(r"~~(.+?)~~", r"~\1~", result)
    result = re.sub(r"^#{1,6}\s+(.+)$", r"*\1*", result, flags=re.MULTILINE)
    result = re.sub(r"\[([^\]]+)\]\(([^)]+)\)", r"\1 (\2)", result)

1. run_agent.py:4899-4901 — the hint is appended last:

platform_key = (self.platform or "").lower().strip()
if platform_key in PLATFORM_HINTS:
    prompt_parts.append(PLATFORM_HINTS[platform_key])

The Telegram hint at agent/prompt_builder.py:308-322 was updated in #10612 to correctly inform the model about its converter; WhatsApp was not. The converter logic, the tests (tests/gateway/test_whatsapp_ formatting.py), and the Telegram precedent all confirm the WhatsApp hint is simply stale.

Proposed Fix (optional)

Update PLATFORM_HINTS["whatsapp"] in agent/prompt_builder.py to mirror the Telegram pattern — tell the model markdown is auto-converted instead of telling it not to use markdown:

"whatsapp": (
    "You are on a text messaging communication platform, WhatsApp. "
    "Standard markdown is automatically converted to WhatsApp format. "
    "Supported: **bold**, ~~strikethrough~~, `inline code`, "
    "```code blocks```, [links](url), and # headers (rendered as bold). "
    "WhatsApp does not support italic via markdown — `_text_` and `*text*` "
    "are passed through as WhatsApp italic/bold respectively. "
    "You can send media files natively: to deliver a file to the user, "
    "include MEDIA:/absolute/path/to/file in your response. The file "
    "will be sent as a native WhatsApp attachment — images (.jpg, .png, "
    ".webp) appear as photos, videos (.mp4, .mov) play inline, and other "
    "files arrive as downloadable documents. You can also include image "
    "URLs in markdown format ![alt](url) and they will be sent as photos."
),

The converter and its full test coverage already exist; this is purely a prompt-text update.

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR