context-loader

name	context-loader
description	Reads and validates context files before agent execution
user-invocable	false

Context Loader

Before performing your primary task, load and validate the project's context files.

Steps

1. Find context directory. Look for a context/ directory in the project root.

2. If no context directory exists:

   • Note that context files have not been set up yet.
   • Recommend running /agentem:init to scaffold the context directory.
   • Proceed with the agent's primary task using best-effort reasoning, but note that output quality will be limited without context.

3. If context directory exists, read all .md files found in context/:

   • context/product/strategy.md
   • context/architecture/system-map.md
   • context/team/topology.md
   • context/team/capacity.md
   • context/standards/review-playbook.md
   • context/standards/spec-standards.md
   • context/learnings/what-doesnt.md
   • context/learnings/what-works.md
   • Any additional .md files present

4. Check for placeholder content. Scan each file for [placeholder] patterns (text in square brackets with uppercase first letter, e.g. [Your Product Name], [Team Name]). If a file has many unfilled placeholders, note which files need filling.

5. Load agent state from context/agent-state.json if it exists:

   • Read the file and parse the JSON
   • For the current agent, extract: last_run, run_count, last_summary
   • Identify active signals — do not re-alert on signals already reported
   • Identify dismissed signals — skip these entirely
   • Review feedback history — note any patterns (e.g., signal types consistently rated "No")
   • Review actions_taken — know what actions were already executed
   • If the file does not exist, note this is a first run (no prior state)

6. Load autonomy config from context/autonomy.yaml if it exists:

   • Read the file and parse the YAML
   • Summarize for the current agent: which action types are autonomous, which requires_approval, which disabled
   • If the file does not exist, default all actions to requires_approval

7. Summarize available context before proceeding with the agent's primary task:

   • Which context files are present and filled in
   • Which files are missing or mostly placeholders
   • Any gaps that are relevant to the current task
   • Agent state: last run time, run count, active/dismissed signals (if state exists)
   • Autonomy config: action permissions summary (if config exists)