context-instructions

Generate or update .github/copilot-instructions.md with enterprise context based on everything configured so far.

Important: Include ALL configured MCP servers

List every MCP server from .mcp.json — both newly installed and pre-existing ones. For each, document:

• What services it covers
• Key tools and when to use them
• Example: workiq covers M365 — "When user asks about emails, meetings, SharePoint docs, or Teams messages, use the workiq MCP server's ask_work_iq tool"

What to generate

Add an ## Enterprise Context section (or update it if it exists) with:

Per-tool blocks

For each configured MCP server/tool:

• MCP server name
• Key tools it provides (list the actual tool names agents should use)
• When to use this tool vs. alternatives
• Team conventions (project keys, space names, URLs)
• Prefer MCP tools over CLI equivalents when available

Skills and agents

Reference any relevant skills or agents:

• "Use the azure-deploy skill for deployments"
• "The context-wizard agent can reconfigure this setup"

Team Intent & Constraints

Retrieve distilled intent from the session store:

SELECT value FROM session_state WHERE key = 'distilled_intent';

Fall back to conversation context if not available. If neither exists, skip this section and note: "No distilled intent available — run /context-distill to generate."

Include the distilled output from Phase 8 (Distill). This is the Intent layer of ISEE — codified so every agent inherits it.

Intent statements

Express what the team values as clear directives:

• "This team prioritises reliability over feature velocity"
• "Security reviews are required for all external-facing changes"

Constraints

Hard rules that agents must always respect:

• "Production deployments require manual approval from a team lead"
• "All PRs must reference a Jira ticket"
• "Secrets must be stored in Azure Key Vault, never in environment variables"

Autonomy boundaries

Render the classified autonomy boundaries from Phase 8 as a structured table:

Before performing any action listed below, check its autonomy level. PROCEED actions can be taken without confirmation. ALWAYS ASK actions require explicit user approval via ask_user. NEVER actions must not be attempted.

The table above is an example — populate it with the actual classifications confirmed by the user in Phase 8 (Distill). Use the icons consistently:

• ✅ for PROCEED
• 🛑 for ALWAYS ASK
• ⛔ for NEVER

Team topology (if available)

Ownership and routing:

• "Team Alpha owns the payments service"
• "Security team must approve changes to auth modules"

Intent changelog reference

After writing the intent and constraints sections, add a footer linking to the changelog:

> Intent changelog: [.github/intent-changelog.md](intent-changelog.md) | Last updated: {date}

This makes the changelog discoverable from the instructions file and provides an audit trail for when and why guardrails changed.

Cross-tool workflows

Based on the tools configured AND the processes discovered in Phase 3/8, describe common workflows:

• Bug triage: issue tracker → monitoring → fix → PR
• Deployment: PR merge → CI/CD → cloud deploy → monitoring verify
• Security: scanning tool → issue tracker ticket → fix → rescan
• Documentation: code change → update docs → review
• Add any team-specific workflows discovered during distillation (e.g., incident response, change advisory, release ceremonies)

Documentation sources

Reference the locations identified in Phase 3, using the content tags:

• [intent] sources: "Team strategy and priorities are in SharePoint — search via workiq"
• [constraint] sources: "Security policies are in the compliance repo"
• [process] sources: "Deployment process is documented in Confluence space ENG"
• [reference] sources: "API specs are in the repo under docs/api/"
• "When asked about 'how we do X', search documentation sources tagged [process] first"

Important

• Read the existing file first — preserve non-enterprise sections
• Use the edit tool to add/update the Enterprise Context section
• Keep instructions actionable — tell Copilot WHEN to use each tool
• Reference actual tool names that Copilot can invoke
• Intent statements and constraints should be written as directives, not descriptions — they tell agents what to do, not what the team thinks about

Merge algorithm for existing files

When updating copilot-instructions.md:

1. File doesn't exist → create it with the full ## Enterprise Context section
2. File exists, no ## Enterprise Context heading → append the new section at the end of the file
3. File exists with ## Enterprise Context → replace everything from ## Enterprise Context up to (but not including) the next ## heading at the same level, or end of file if no next heading
4. Preserve user content — all sections outside ## Enterprise Context are never modified. Custom instructions, notes, and other headings remain untouched
5. Manual edit detection — if the existing Enterprise Context section contains content that doesn't match wizard-generated patterns (e.g., hand-written tool descriptions, custom workflows), warn the user via ask_user: "The Enterprise Context section appears to have manual edits. Regenerating will replace them. Proceed?"

Then mark this phase done:

UPDATE todos SET status = 'done' WHERE id = 'ctx-instructions';