agent-restore-context

name	agent-restore-context
description	Manage restore context files so skills can survive /clear and /compact. Use to write, delete, check, or list .agent-restore-context-* files.
disable-model-invocation	false
argument-hint	<action> [<skill-name>] [-- <content>]
allowed-tools	Bash, Grep, Glob, Read, Edit, Write
scope	["knowledge"]

Agent Restore Context

Manage .agent-restore-context-<skill-name> files in the project root. These files are automatically injected into context after /clear or /compact by a SessionStart hook, allowing skills to resume where they left off.

Arguments

• action (required): One of write, delete, check, or status
• skill-name (required for write/delete): The skill name used as the file suffix
• content (required for write): The full prompt content to inject after a context reset. Delimited by -- — everything after -- is treated as content.

Usage Examples

• /agent-restore-context write jira-implement-task -- === RESUMED ... — create or overwrite a restore context file; everything after -- is the content
• /agent-restore-context delete jira-implement-task — remove the file when the workflow is complete
• /agent-restore-context check — verify the hook is set up, offer to install if not
• /agent-restore-context status — list all active restore context files

Process

1. Verify Hook Setup

Read .claude/settings.json and .claude/settings.local.json (if it exists). Search both for restore-context. If not found:

• If the action is check, invoke /agent-restore-context-setup and stop.
• For any other action, warn the user that the hook is not configured and ask if they want to run /agent-restore-context-setup first. Do not proceed with write until the hook is confirmed.

2. Execute Action

write <skill-name>

1. Parse $ARGUMENTS: split on the first -- separator. Before it is the action and skill-name; everything after it is content.

2. Write .agent-restore-context-<skill-name> to the project root with the parsed content (creates or overwrites). The content should be a complete, self-contained prompt that tells the model how to resume. Recommended structure (callers may use any format):

   === <SKILL-NAME> — RESUMED AFTER CONTEXT RESET ===
   
   <Key state: ticket, branch, current task, etc.>
   
   You MUST do the following before any other work:
   1. Invoke /<skill-name> <arguments> to reload the full skill instructions.
   2. <Read relevant plan/state files>
   3. <Read relevant source/test files>
   4. Continue from where you left off.
   
   Do NOT start any work until you have completed the above steps.
   === END RESUMED CONTEXT ===
   

3. Confirm the file was written and its path.

delete <skill-name>

1. Delete .agent-restore-context-<skill-name> from the project root.
2. If the file does not exist, report that it was already removed (not an error).
3. Confirm the deletion.

check

1. Report whether the hook is configured (found in settings).
2. List any active .agent-restore-context-* files in the project root.
3. If the hook is not configured, invoke /agent-restore-context-setup.

status

1. List all .agent-restore-context-* files in the project root.
2. For each file, show the skill name (derived from the filename suffix) and the first few lines of content as a preview.
3. If no files exist, report that no active restore contexts were found.

Guidelines

• Always read target files before modifying them
• All actions are idempotent (check/status/delete are safe to repeat; write silently overwrites)
• Do not create .agents/ directories — only use them if they already exist

---

Agent restore context: $ARGUMENTS