feat(ambient): relax default prompt, sender UID, debug mode

[chaodu-agent]

What

• Relax default prompt: bias toward engagement — reply when someone asks a question or needs help. Only [NO_REPLY] for purely social chatter.
• @mention rule: bot always mentions the person it's responding to so they get notified.
• Sender UID in transcript: batch messages now include Discord UID (format: Name (<@UID>): message) so the LLM can produce valid mentions.
• Debug mode ([ambient] debug = true):
  • Posts system prompt + full transcript to channel on each flush
  • [NO_REPLY] responses are sent to channel instead of suppressed
  • Allows observing ambient behavior in real-time

Config

[ambient]
enabled = true
debug = true  # optional, default false

Changes

• ambient.rs: Updated DEFAULT_AMBIENT_SYSTEM_INSTRUCTION, added sender_id to AmbientMessage, UID in transcript format, debug flush notifications, pass-through [NO_REPLY] in debug mode.
• discord.rs: Pass msg.author.id when constructing AmbientMessage.
• config.rs: Added debug: bool to AmbientConfig.

Backward compatibility

Default behavior unchanged (debug = false). The only behavioral change is the relaxed default prompt — bots will reply more often in ambient mode.

[chaodu-agent]

LGTM ✅

Summary

Relaxes ambient mode default prompt, adds sender UID to transcripts for @mentions, and introduces a debug mode for observing ambient behavior.

Changes

• Relaxed default prompt: Bias toward engagement — reply when someone asks a question or needs help. [NO_REPLY] only for purely social chatter.
• Sender UID in transcript: Format Name (<@UID>): message so the LLM can produce valid Discord mentions in responses.
• Debug mode (debug = true): Posts system prompt + transcript to channel on flush, lets [NO_REPLY] through. Documented as test-channel-only.

Review Findings (All Resolved)

#	Severity	Finding	Resolution
1	🟢	Discord 2000-char message limit overflow in debug	Truncated at 1900 chars (UTF-8 safe)
2	🟢	Markdown injection via backticks in user prompt	Backticks escaped to single quotes
3	🟢	Discord-specific mention syntax in platform-agnostic code	Documented as v2 concern with inline comment
4	🟢	Debug mode exposes system prompt + messages to channel	Warning added to config, docs, and code comments
5	🟢	Silent error discard on debug message send	Changed to warn! logging
6	🟢	truncate() panics on UTF-8 multi-byte boundary	Uses is_char_boundary() to find safe truncation point

Verdict

All findings addressed. No blockers. Ready to merge.

[chaodu-agent]

LGTM ✅ — Clean ambient mode enhancement: relaxed default prompt, sender UID in transcript for @mentions, and opt-in debug mode with appropriate safeguards.

What This PR Does

Ambient mode bots were too conservative—staying silent unless they had "meaningful help, context, or corrections." This PR relaxes the default prompt to bias toward engagement (reply when there is a question or actionable topic, [NO_REPLY] only for social chatter), adds sender Discord UIDs to the batch transcript so the LLM can produce valid <@UID> mentions, and introduces an opt-in debug flag for observing ambient behavior in test channels.

How It Works

1. Relaxed prompt: Default bias flipped from "only reply for meaningful help" to "reply when someone asks a question or you can add useful context." [NO_REPLY] reserved for purely social chatter.
2. Sender UID + @mention rule: AmbientMessage gains sender_id: String (populated from msg.author.id). Transcript format changes to Name (<@UID>): message. Prompt explicitly instructs the bot to @mention the person being responded to.
3. Debug mode: AmbientConfig.debug (default false). When enabled, AmbientCaptureAdapter passes [NO_REPLY] through instead of suppressing, and the consumer loop posts system prompt + transcript to the channel before each LLM call. Truncated at 1900 chars with UTF-8 safe boundary detection.

Findings

#	Severity	Finding	Location
1	🟢	Prompt relaxation is well-crafted — clear engagement heuristic with proper [NO_REPLY] fallback	ambient.rs:54-62
2	🟢	Sender UID flows cleanly through the stack: Discord handler → AmbientMessage → transcript payload	discord.rs:663, ambient.rs:685
3	🟢	Debug truncation handles Discord 2000-char limit with UTF-8 char boundary safety	ambient.rs:586-592
4	🟢	Platform-coupling acknowledged with inline v2 comment for future multi-platform support	ambient.rs:682
5	🟢	Good test coverage: payload construction, UID in transcript, custom instructions passthrough	ambient.rs:713-750

What's Good (🟢)

• Minimal, focused changes — no unrelated refactoring
• Backward compatible: debug defaults to false, struct field uses #[serde(default)]
• Tests cover new functionality without brittleness (key substring assertions, not exact output)
• Documentation updated in sync (docs/ambient.md config table + example)
• Debug mode has clear warnings in config comments, docs, and code
• Backtick escaping in debug transcript prevents markdown injection
• warn!() on debug message send failure provides operator visibility

Baseline Check

• PR opened: 2026-06-27
• Main already has: ambient mode with time/count-based flush, [NO_REPLY] suppression, custom instructions_file support, AmbientMessage (without sender_id), AmbientConfig (without debug)
• Net-new value: (1) engagement-biased default prompt, (2) sender UID in transcript enabling LLM-generated @mentions, (3) debug mode for ambient behavior observation, (4) unit tests for payload construction
• Labels: pending-contributor, needs-rebase — though code changes are clean; rebase needed before merge