Skip to content

What to expect from Archcore

Archcore makes recorded project context available to your agent. You can ask it to read, create, update, and connect documents; supported lifecycle hooks also supply context during a session.

Automatic behavior depends on your agent’s capabilities and configuration. MCP access alone does not provide lifecycle hooks or slash commands.

The agent can use your recorded decisions when changing code. It has less to reconstruct from the implementation or ask you to explain again.

SituationObservable sign that context is being used
You start a task in a familiar moduleThe agent reads and names the relevant rules or decisions before proposing changes
A request conflicts with an accepted decisionThe agent explains the conflict and asks you to resolve it before changing that decision
You record a new constraintA later task retrieves the record and explains how it affects the proposed code
You continue in a new sessionThe agent reads saved decisions and plans, distinguishes drafts from accepted records, and identifies unfinished work

These are behaviors to look for, not guaranteed outcomes. A document on disk or a delivered hook message does not prove the agent read or applied it. Ask it to cite the relevant record and explain the consequence for the task; review the resulting code as well.

The long-term aim is to retain useful knowledge between tasks. Outdated or contradictory context still needs correction; adding documents alone does not establish better code or fewer mistakes.

With a working session-start hook, the agent receives a bounded recap: document counts, the current branch, work in progress, recent accepted documents, tags, and relation counts.

The recap tells the agent what context exists. It does not load every document’s complete text. To answer a specific project question, the agent can search and read the relevant documents through MCP.

On hosts that support pre-edit context injection, Archcore can supply the rules and specifications associated with the file about to change. That context is advisory: it informs the agent but does not prove that the resulting code follows it.

GitHub Copilot CLI’s pre-write event accepts a permission decision but cannot receive this context. On Codex CLI, hooks need its feature flag and a trusted project configuration; Windows hooks are unavailable in the documented integration.

If automatic injection is unavailable, ask the agent to read the relevant context explicitly.

Ask for the outcome you need, such as recording a decision or planning a feature. On plugin hosts, the Archcore commands provide shortcuts for these requests.

Document changes go through MCP tools. With a working write guard, direct edits to .archcore/ documents are blocked so the agent uses those tools. MCP mutations validate document structure; post-write hooks can also report precision findings and documents affected by a relation cascade.

MechanismEffect
Write guardCan block a direct document write
MCP validationCan reject an invalid document mutation
Context injectionSupplies information before an edit
Post-write findingsReport issues for review; do not undo the write

The connection guide explains the mechanisms and host configuration.

Review proposed documents before accepting them. A draft is work in progress; accepted records your approval. A template or a successful validation check does not establish that the content is true.

Archcore does not resolve conflicting decisions automatically. Relations can record a contradiction or replacement, but you and the agent still need to read the evidence and record the resolution.

Your host supplies execution tools and permissions. Archcore supplies project context; a planning command does not replace your host’s controls or establish the correctness of generated code.

Ask the agent to read a known document and cite its path. Inspect the tool activity for the retrieval. If the record is missing or a hook does not run, use Troubleshooting.