add-compact

name	add-compact
description	Add /compact command for manual context compaction. Solves context rot in long sessions by forwarding the SDK's built-in /compact slash command. Main-group or trusted sender only.

Add /compact Command

Adds a /compact session command that compacts conversation history to fight context rot in long-running sessions. Uses the Claude Agent SDK's built-in /compact slash command — no synthetic system prompts.

Session contract: /compact keeps the same logical session alive. The SDK returns a new session ID after compaction (via the init system message), which the agent-runner forwards to the orchestrator as newSessionId. No destructive reset occurs — the agent retains summarized context.

Phase 1: Pre-flight

Check if src/session-commands.ts exists:

test -f src/session-commands.ts && echo "Already applied" || echo "Not applied"

If already applied, skip to Phase 3 (Verify).

Phase 2: Apply Code Changes

Merge the skill branch:

git fetch upstream skill/compact
git merge upstream/skill/compact

Note: upstream is the remote pointing to sbusso/claudeclaw. If using a different remote name, substitute accordingly.

This adds:

• src/session-commands.ts (extract and authorize session commands)
• src/session-commands.test.ts (unit tests for command parsing and auth)
• Session command interception in src/index.ts (both processGroupMessages and startMessageLoop)
• Slash command handling in agent/runner/src/index.ts

Validate

npm test
npm run build

Rebuild container

./src/runtimes/docker/build.sh

Service name: Derived from the directory name: com.claudeclaw.<dirname> (macOS) / claudeclaw-<dirname> (Linux). For example, if cwd is my-assistant, the service is com.claudeclaw.my-assistant. Determine the correct service name before running service commands below.

Restart service

launchctl kickstart -k gui/$(id -u)/com.claudeclaw  # macOS
# Linux: systemctl --user restart claudeclaw

Phase 3: Verify

Integration Test

1. Start ClaudeClaw in dev mode: npm run dev
2. From the main group (self-chat), send exactly: /compact
3. Verify:
   • The agent acknowledges compaction (e.g., "Conversation compacted.")
   • The session continues — send a follow-up message and verify the agent responds coherently
   • A conversation archive is written to groups/{folder}/conversations/ (by the PreCompact hook)
   • Container logs show Compact boundary observed (confirms SDK actually compacted)
   • If compact_boundary was NOT observed, the response says "compact_boundary was not observed"
4. From a non-main group as a non-admin user, send: @<assistant> /compact
5. Verify:
   • The bot responds with "Session commands require admin access."
   • No compaction occurs, no container is spawned for the command
6. From a non-main group as the admin (device owner / is_from_me), send: @<assistant> /compact
7. Verify:
   • Compaction proceeds normally (same behavior as main group)
8. While an active container is running for the main group, send /compact
9. Verify:
   • The active container is signaled to close (authorized senders only — untrusted senders cannot kill in-flight work)
   • Compaction proceeds via a new container once the active one exits
   • The command is not dropped (no cursor race)
10. Send a normal message, then /compact, then another normal message in quick succession (same polling batch):
11. Verify:
    • Pre-compact messages are sent to the agent first (check container logs for two runAgent calls)
    • Compaction proceeds after pre-compact messages are processed
    • Messages after /compact in the batch are preserved (cursor advances to /compact's timestamp only) and processed on the next poll cycle
12. From a non-main group as a non-admin user, send @<assistant> /compact:
13. Verify:
    • Denial message is sent ("Session commands require admin access.")
    • The /compact is consumed (cursor advanced) — it does NOT replay on future polls

Validation on Fresh Clone

git clone <your-fork> /tmp/claudeclaw-test
cd /tmp/claudeclaw-test
claude  # then run /add-compact
npm run build
npm test
./src/runtimes/docker/build.sh
# Manual: send /compact from main group, verify compaction + continuation
# Manual: send @<assistant> /compact from non-main as non-admin, verify denial
# Manual: send @<assistant> /compact from non-main as admin, verify allowed
# Manual: verify no auto-compaction behavior

Security Constraints

• Main-group or trusted/admin sender only. The main group is the user's private self-chat and is trusted (see docs/SECURITY.md). Non-main groups are untrusted — a careless or malicious user could wipe the agent's short-term memory. However, the device owner (is_from_me) is always trusted and can compact from any group.
• No auto-compaction. This skill implements manual compaction only. Automatic threshold-based compaction is a separate concern and should be a separate skill.
• No config file. ClaudeClaw's philosophy is customization through code changes, not configuration sprawl.
• Transcript archived before compaction. The existing PreCompact hook in the agent-runner archives the full transcript to conversations/ before the SDK compacts it.
• Session continues after compaction. This is not a destructive reset. The conversation continues with summarized context.

What This Does NOT Do

• No automatic compaction threshold (add separately if desired)
• No /clear command (separate skill, separate semantics — /clear is a destructive reset)
• No cross-group compaction (each group's session is isolated)
• No changes to the container image, Dockerfile, or build script

Troubleshooting

• "Session commands require admin access": Only the device owner (is_from_me) or main-group senders can use /compact. Other users are denied.
• No compact_boundary in logs: The SDK may not emit this event in all versions. Check the agent-runner logs for the warning message. Compaction may still have succeeded.
• Pre-compact failure: If messages before /compact fail to process, the error message says "Failed to process messages before /compact." The cursor advances past sent output to prevent duplicates; /compact remains pending for the next attempt.