discord-ai-dispatcher（Discord AI dispatcher 控制）

Generic Webhook Notifier: Advanced Skill Guide

The Webhook Notifier skill provides a robust architecture that combines hard heuristics with local AI reasoning tools to ensure 100% routing accuracy and high-density visual formatting.

🚀 The High-Consistency Architecture: Deterministic Routing

Instead of relying solely on AI categorization, this dispatcher implements a Deterministic Prefix Engine to ensure 100% routing accuracy:

1. Prefix Identification: An updateId (e.g., systemevent00100) directly dictates the target channel.

- systemevent -> Infrastructure Events - appalert -> Application Monitoring - userupdate -> User Activity

1. Zero-Guessing: This removes "Classification Drift" where an AI might misinterpret an update, ensuring data always reaches the correct stakeholders.

🔢 Deterministic Sequential Deduplication

To handle out-of-order API deliveries or network retries, the dispatcher uses a Sequential State Tracker:

• Logic: It extracts a numeric suffix from the ID (the XXXXX in prefixXXXXX).
• State: The highest seen ID for each category is stored locally.
• Rule: Use Numeric Comparison (100 > 99) instead of Alphabetical (100 < 99). Regex-extract digits [^\d] and cast to [long] to ensure the sequence advances even as ID lengths change.
• Result: You never see a stale update overwrite new data, and duplicate notifications are prevented.

🔒 Global Pipeline Locking (Self-Protection)

To prevent race conditions and file corruption when the AI takes longer than the polling interval:

1. Lock File: Create a dispatcher.lock at the start of the orchestrator.
2. Abort Logic: If a lock exists and is "fresh" (< 30 mins), the next pulse must exit immediately.
3. Atomic Recovery: Always use try/finally blocks to ensure the lock is released even if the script crashes.

🏗️ Self-Healing JSON Recovery

State files can be corrupted by power loss or interrupted writes.

1. Integrity Checks: Wrap every ConvertFrom-Json call in a try/catch.
2. Auto-Reset: If parsing fails, log the corruption, delete the broken file, and re-initialize the state from the API source.

• Benefit: The system self-recovers from disk errors without manual developer intervention.

🩹 The "Healing Cleaner" Pattern

Instead of a "Negative Filter" (which deletes bad formatting), this skill uses a Positive Healing logic to repair inconsistent AI output:

1. Label Extraction: It scans for colons and automatically wraps the text before them in **Bold Labels**.
2. Normalization: It forcefully standardizes fluctuating terms to established vocabulary.
3. Noise Blacklist: It aggressively strips AI conversational filler.
4. Artifact Stripping: It cleans up redundant formatting artifacts (** **), escaped characters, and unwanted markdown blocks.

• Result: You get a professional, uniform feed even if the AI model gets "creative".

📊 Dashboard Mode (Clean Chat Strategy)

Prevent chat clutter by updating existing messages instead of posting new ones.

1. Logic: Use PATCH (or corresponding update paradigm) instead of POST for webhooks.
2. State: Store the created message ID in a local state file.

• Advantage: Provides a single "Living" source of truth for your system status without generating unnecessary notification spam.

🪙 Token & Cost Optimization

...

• Payload Pruning: Strip all non-essential JSON keys before the AI sees them.
• Compressed Instructions: Use strict list-based prompts instead of paragraphs.
• Low Effort Fallback: If the AI is offline, use a hardcoded template for high availability.

🧠 The AI Improvement Layer: Learning From Gaps

The Webhook Notifier is built to be a self-improving system. When a gap is found, the agent must shift from Monitor to Diagnostician:

1. The RCA Feed (Log Diagnostics)

Instead of guessing, the agent must scan the dispatcher.log for the "Missing IDs":

• Skipped (duplicate) but not Delivered: Indicates the deduplication logic was too aggressive (Historical Logic Flaw).
• Timeout: Indicates the Local LLM needs a higher timeout or more GPU resources.
• No Log Record: Indicates a failure at the Poller or API fetch stage.

2. Strategic Improvement Loop

When the user feeds you an RCA report, you should automatically suggest (or implement) one of the following:

1. State Realignment: Resetting last-seen-ids.json or polling-state.json to the actual last-delivered marker to ensure the "Gaps" are re-sent.
2. Logic Tuning: If sequential IDs are conflicting with timestamp IDs, adjust the Test-IsSequentialDuplicate helper to be more specific.
3. High-Precision Recovery: Creating a one-time "Healing Queue" to force-inject missed notifications into Discord.

🛡️ Recovery Best Practices:

• Bypass Deduplication: When in "Heal Mode", always bypass the Test-IsSequentialDuplicate filter to ensure recovery happens.
• Audit History: Perform a "Log Audit" before making changes. Never assume a file is missing just because it's not in the DB—always check the logs to confirm if the Dispatcher ever saw it.

🎨 Best Practices for Professional Formatting

To maintain aesthetics and readability:

• Mandatory Bolding: Always wrap labels in visual emphasis (e.g., Key: Value).
• Horizontal Dividers: Every notification should logically separate from the next.
• Tool-Calling First: Calculate your metrics in code; use the AI *only* for the final beautification.

⚙️ Automation & Version Compatibility

• Scheduling Reliability: Use Headless Discrete Execution (Windows Task Scheduler / Cron). Avoid infinite while($true) loops to prevent memory leaks and handle system reboots gracefully.
• PowerShell 5.1 Compatibility:

- Hashtables: Modern PS 7.0 -AsHashtable fails in 5.1. Use manual property-looping to convert PSCustomObjects into Hashtables for reliable $state[$key] access. - Encoding: Always specify -Encoding UTF8 on Set-Content to prevent character corruption in mixed-version environments.

• Webhook First: Always provide users the option to setup webhooks immediately after configuring the skill for instant "out-of-box" operation.

*Designed for extensible and robust notification pipelines.*