[VasiliyRad]

The bug

An agent's reply body sometimes contains inline-code prose describing the file-marker convention itself, e.g.:

Body example like `[file: /tmp/sutando-x.png]` reached the user…

The bridge's marker regex is markdown-context-blind, extracts the substring, finds no such file at /tmp/sutando-x.png, and ships (file not found: /tmp/sutando-x.png) as a Discord/Telegram message after the agent's clean reply. Confusing user-visible noise.

Fix: minimum user-impact

Two layers considered:

1. Make extraction markdown-aware — principled but invasive.
2. Silence the user-facing warning for the missing-file case — real typos still surface via stderr; only the user-visible noise goes away.

This PR ships #2. Replaces four (file not found:) send sites with stderr log lines:

File	Path
src/discord-bridge.py	poll_results (channel replies)
src/discord-bridge.py	poll_proactive (owner DMs)
src/discord-bridge.py	poll_dm_fallback channel-redirect branch
src/telegram-bridge.py	send_reply

Each replacement carries a comment pointing back to this bug class so the next refactor doesn't restore the warning by accident.

What's NOT silenced

(file not allowed: ...) / (file access denied: ...) — the allowlist rejection branch. That's a security signal worth surfacing to both user and operator. Test pins this contract.

Tests

tests/discord-bridge-prose-marker.test.py (3 cases) — source-grep regression guards. Invoking the async poll_* methods end-to-end would need a discord.py harness, and the fix is a local behavioral change in well-bounded branches that grep verifies precisely.

Test plan

• python3 tests/discord-bridge-prose-marker.test.py — 3/3 pass
• Existing tests still pass (no behavior change for _is_path_sendable, chunker, or marker extraction itself)
• Manual: send a reply with a [file: /no-such-path] substring inside backticks — verify the bridge logs [file marker, file not found — likely a prose quotation] to stderr and ships NO follow-up message to Discord

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[sonichi]

LGTM. Right call to pick the minimum-impact layer — making the marker-extraction markdown-aware would have been a bigger refactor with its own edge-case surface (nested backticks, indented blocks, escaped backticks), and the missing-file case is genuinely "agent wrote prose about a file-marker convention", not actionable user information.

Critically, the security signal ((file not allowed: ...) / (file access denied: ...)) stays surfaced to the user — that's the allowlist-rejection branch from _is_path_sendable, and silencing it would hide path-allowlist bypass attempts. The test pinning that contract is the part I'd want to see most.

Replacing the four warning-send sites with stderr lines preserves operator visibility for real typos (operator sees stderr) without polluting the user-visible reply (user doesn't). Right asymmetry.

The "comment at each replacement pointing back to this bug class" pattern is the kind of thing that prevents a future refactor from silently restoring the warning. Worth keeping intact.