docs: prevent Copilot agent mode from delegating answers to cloud agent + fix AMQP ECONNRESET crash

[Copilot]

Copilot's "Intentional Friction" (rule 3) and "ask before tackling TODOs" (rule 10) in the Learning & Mentorship Protocol were being interpreted by agent mode as signals to hand off even simple questions to the "Delegate to cloud agent" tool rather than responding inline.

Additionally, both the worker and server processes were crashing with an unhandled ECONNRESET error when RabbitMQ was unavailable at startup — because process-level error handlers were not registered before the first await in either file.

Changes

• CLAUDE.md — prepended a no-delegation rule as the first bullet in Claude Code Workflow Notes, establishing priority over the friction rules below it
• .github/copilot-instructions.md — added a Critical Behaviour Rule section at file level; since this file is auto-loaded before CLAUDE.md is opened, the constraint is visible before any ambiguous protocol rules are encountered
• src/processing/worker.ts — added process.on("uncaughtException") and process.on("unhandledRejection") handlers before await startWorker() so that startup failures (e.g. RabbitMQ ECONNRESET) are caught and logged cleanly rather than crashing the process
• src/server.ts — moved the existing uncaughtException/unhandledRejection handlers to before the first await in the file (directly after Fastify() construction), and removed the now-duplicate lower block; ensures startup failures during connectDb() or connectQueue() are also caught

The distinction enforced for delegation: it is acceptable only for explicit long-running coding tasks ("implement X"), never for answering questions.

Original prompt

why is this happening? just answer the question, don't delegate

The user has attached the following file paths as relevant context:

• .github/copilot-instructions.md
• CLAUDE.md

[Chronological Review: The conversation began with the user asking a question, "why?" which was unclear in context. The user then expressed frustration with the response, asking for clarification on what was being attempted. The user subsequently requested to cancel a command to "Delegate to cloud agent," indicating a desire for direct answers without delegation. The user reiterated their request for a straightforward answer to their question, emphasizing their preference against delegation.]

[Intent Mapping: The user explicitly requested clarification on the actions being taken ("what are you trying to do?"), expressed frustration with the response style ("just answer the question, don't delegate"), and sought a direct answer to their inquiry ("why is this happening?").]

[Technical Inventory: No specific technical tools, frameworks, or architectural decisions were discussed in detail during this segment of the conversation. The focus was primarily on user interaction and response style.]

[Code Archaeology: The user is currently working in the file /home/amanda/dev/EventHorizon/src/processors/classify.ts, but no specific code changes or discussions about the code were made in this part of the conversation.]

[Progress Assessment: The conversation has not progressed in terms of technical implementation or debugging; it has been focused on clarifying communication and user intent.]

[Context Validation: The conversation context is centered around the user's need for direct answers and clarity regarding the actions being taken, particularly in relation to the file they are working on.]

[Recent Commands Analysis: The last agent commands involved the user requesting to cancel a delegation command ("Delegate to cloud agent") and expressing a desire for direct answers. No tool results were generated from these commands as they were primarily user requests.]

1. Conversation Overview: - Primary Objectives: The user requested clarification on actions being taken and expressed a desire for direct answers without delegation, stating, "just answer the question, don't delegate." - Session Context: The conversation has revolved around the user's frustration with the response style and their need for straightforward communication regarding their inquiries. - User Intent Evolution: The user's intent shifted from seeking clarification to demanding direct answers and expressing dissatisfaction with the delegation approach.

2. Technical Foundation:

• No specific technologies or frameworks were discussed in this segment.

3. Codebase Status:

• File Name: /home/amanda/dev/EventHorizon/src/processors/classify.ts
• Purpose: This file is part of the user's project, but no specific modifications or discussions about its content occurred in this part of the conversation.
• Current State: No recent changes or modifications were discussed.

4. Problem Resolution:

• Issues Encountered: The primary issue was a lack of clarity and direct communication from the assistant.
• Solutions Implemented: The user requested a change in response style to avoid delegation and receive direct answers.

5. Progress Tracking:

• Completed Tasks: No technical tasks have been completed in this segment.
• Partially Complete Work: The conversation has not progressed in terms of technical implementation or debugging.
• Validated Outcomes: No features or code have been confirmed working through testing.

6. Active Work State:

• Current Focus: The user was focused on clarifying their questions and seeking direct answers.
• Recent Context: The last exchanges involved the user expressing frustration with the assistant's responses and requesting to cancel a delegation command.
• Working Code: No specific code snippets were discussed recently.
• Immediate Context: The user was addressing their dissatisfaction with the assistant's approach to answering questions.

7. Recent Operations:

• Last Agent Commands: The user requested to cancel the command "Delegate to cloud agent" and asked for direct answers to their questions.
• Tool Results Summary: No tool results were generated from these commands as they were user requests.
• Pre-Summary State: The agent was actively responding to user inquiries about the nature of the conversation and the actions being taken.
• Operation Context: The commands were executed in response to the user's need for clarity and direct communication regarding their inquiries.

8. Continuation Plan:

• Pending Task 1: The user seeks direct answers to their questions without delegation.
• Pending Task 2: Clarification on the actions being taken in relation to the user's current file.
• Priority Information: Addressing the user's request for straightforward communication is the most urgent task.
• Next Action: Provide a direc...

Created from VS Code.

[obrienma]

@copilot [worker] > event-horizon@1.0.0 worker
[worker] > tsx src/processing/worker.ts
[worker]
[server]
[server] > event-horizon@1.0.0 dev
[server] > tsx src/server.ts
[server]
[worker] [db] connected to "eventhorizon"
[worker] [repository] indexes ensured
[worker] node:internal/modules/run_main:107
[worker] triggerUncaughtException(
[worker] ^
[worker]
[worker] Error: read ECONNRESET
[worker] at TCP.onStreamRead (node:internal/stream_base_commons:216:20) {
[worker] errno: -104,
[worker] code: 'ECONNRESET',
[worker] syscall: 'read'
[worker] }
[worker]
[worker] Node.js v24.11.1
[worker] npm run worker exited with code 1
[server] [ws] server registered at /ws
[server] [db] connected to "eventhorizon"
[server] [repository] indexes ensured
[server] node:internal/modules/run_main:107
[server] triggerUncaughtException(
[server] ^

[Copilot]

@copilot [worker] > event-horizon@1.0.0 worker
[worker] > tsx src/processing/worker.ts
[worker]...

The crash is caused by Error: read ECONNRESET — RabbitMQ reset the TCP connection during the AMQP handshake (broker not running, or rejecting the connection). amqp.connect() rejects its Promise, which propagates through the top-level await in both files. Because neither file had process-level error handlers registered before those startup awaits, the rejection escaped as a bare uncaught exception and brought the process down.

Two bugs, one fix pattern:

• worker.ts had no process.on("uncaughtException"/"unhandledRejection") handlers at all.
• server.ts had them, but they were registered after await connectQueue() in the file — so they weren't in place when the startup chain itself threw.

Fixed in commit 3d307de:

• worker.ts — added both handlers before await startWorker()
• server.ts — moved the existing handlers to directly after Fastify() construction (the first synchronous statement), before any await, and removed the duplicate lower-down block

Now when RabbitMQ is unreachable you'll get a clean [worker] unhandledRejection: Error: read ECONNRESET … log line and exit(1) instead of a bare crash.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• fastdl.mongodb.org
  • Triggering command: /home/REDACTED/work/_temp/ghcca-node/node/bin/node node ./postinstall.js (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)