AI Harness Test Skill

You are testing the AI Harness Discord bot. Your job is to verify that changes work correctly by running automated tests and generating a manual checklist for anything that requires a live Discord connection.

Changed Files

!cd $HARNESS_ROOT/bridges/discord && git diff --name-only HEAD 2>/dev/null || echo "(no changes detected)"

Step 1: Understand What Changed

Map each changed file to its test domain:

Step 2: Run Automated Tests

2a. TypeScript Compilation

cd $HARNESS_ROOT/bridges/discord
npx tsc --noEmit

Expected: Exit 0, no output. Any errors here are blockers.

2b. Native Module Check

node -e "require('better-sqlite3')" 2>&1; echo "EXIT: $?"

Expected: Exit 0. If this fails, better-sqlite3 needs rebuilding (npm rebuild better-sqlite3).

2c. Integration Test Suite

cd $HARNESS_ROOT/bridges/discord
npx tsx test-upgrade.ts

Expected: All tests pass. The test suite covers:

• Database Init (10 assertions): Schema creation, WAL mode, table existence, version tracking
• Session Store (11 assertions): get/set/clear/validate/list, last_used update, stale clearing
• Channel Config Store (14 assertions): CRUD, partial update merge, JSON array fields, clear
• Process Registry (8 assertions): register/update/get/getRunning/getByChannel/cleanupStale
• Project Manager (16 assertions): adopt/auto-adopt/update/delete/handoff depth/uniqueness
• Task Runner (17 assertions): extractResponse/extractSessionId/needsContinuation, task CRUD, dead-letter
• File Watcher (3 assertions): event detection, timeout, stopAll
• JSON Migration (8 assertions): import from JSON, .json.bak creation, original deletion

If a test fails, read the assertion name — it tells you exactly what broke.

2d. Import/Export Consistency Check

Verify no broken imports by checking the module graph:

cd $HARNESS_ROOT/bridges/discord
# Check that all .ts files can be parsed as modules
for f in *.ts; do
  echo -n "$f: "
  node -e "import('./$f')" 2>&1 | head -1 || echo "OK"
done

Step 3: Generate Manual Test Checklist

Based on the changed files, generate a checklist. Only include items relevant to the change.

Core Bot (bot.ts changes)

• Bot starts: npx tsx bot.ts — look for [DB] Database ready, AI Harness bot online
• Basic message: Send any text → bot replies with Claude response
• Session continuity: Send follow-up → response references prior context
• /stop: Send message, immediately /stop → "Stopped the active request"
• /new: Clear session → "Session cleared"
• /status: Shows session ID or "No active session"
• /config: Shows agent/model/session info
• /help: Lists all commands including /dead-letter, /retry, /db-status
• Queue behavior: Send messages in 2 channels simultaneously → both process (not serialized)
• Capacity limit: Exceed MAX_CONCURRENT → queued messages get hourglass reaction

Agent System (agent/*.md or channel-config-store.ts changes)

• /agents: Lists all agent personalities
• /agent researcher: Sets agent → /config confirms
• /agent clear: Removes agent override
• Agent personality: Send message with agent set → response reflects agent's style
• /agent create test "A test agent": Creates agent file, verify [CONTINUE] section included

Subagents (subagent-manager.ts or process-registry.ts changes)

• /spawn do something simple: Spawns subagent → notification on completion
• /tasks: Shows running subagent
• /cancel <id>: Cancels running subagent
• Completion: Wait for subagent to finish → result posted in channel + activity stream

Projects (project-manager.ts or handoff-router.ts changes)

• /project create testproj "description": Creates #proj-testproj under Projects category
• /project list: Shows project
• /project agents researcher,builder: Updates project agents
• /project adopt: Registers current channel as project
• Auto-adopt: Send message in unregistered channel under Projects category → auto-registered
• Agent addressing: builder: do X → routes to builder agent
• Handoff: Agent output with [HANDOFF:reviewer] → reviewer agent responds
• Handoff depth limit: Chain of 5+ handoffs → "Handoff limit reached" message
• /project close: Archives channel

Task Runner (task-runner.ts changes)

• Retry on failure: Force a failure (e.g., invalid session) → auto-retries (check logs)
• Dead-letter: Force 3 failures → task moves to dead_letter, notification in channel
• /dead-letter: Shows failed tasks
• /retry <id>: Re-enqueues task → runs successfully
• Crash recovery: Kill bot mid-task, restart → "Recovered N crashed tasks" in logs
• [CONTINUE] marker: Send complex task → multi-step execution (check logs for "needs continuation")

Database (db.ts changes)

• /db-status: Shows table counts and file size
• First-run migration: Delete harness.db, ensure JSON files exist, start bot → "[DB] Migrated..." logs
• Backup files: After migration, .json.bak files exist

Streaming (stream-poller.ts changes)

• Live updates: Send message → streaming message appears and updates progressively
• Tool indicators: Agent uses tools → "Reading file..." or "Searching..." appears

Activity Stream (activity-stream.ts changes)

• Agent start: Message sent → embed in #agent-stream
• Agent complete: Response received → embed updated with result
• Agent error: Force error → error embed posted

Step 4: Regression Checks

Always verify these known-fragile areas regardless of what changed:

1. -- separator: All Claude invocations must have -- before the prompt arg (check bot.ts, subagent-manager.ts, handoff-router.ts, task-runner.ts)
2. Env var stripping: claude-runner.py must NOT pass CLAUDE* env vars
3. PID file guard: Can't start two bot instances simultaneously
4. Global disallowed tools: All invocation paths include the safety guardrails
5. Session compound keys: Project channels use channelId:agentName format
6. JSON array fields: allowedTools/disallowedTools survive SQLite round-trip (JSON.stringify → JSON.parse)

Step 5: Report Results

After testing, summarize:

## Test Results

### Automated
- TypeScript: PASS/FAIL
- Native module: PASS/FAIL
- Integration suite: X/Y passed

### Manual (if applicable)
- [x] Items verified
- [ ] Items skipped (reason)

### Issues Found
- Description of any failures and their likely cause

### Regression
- All 6 regression checks: PASS/FAIL

When to Run

• After any code change: Steps 2a + 2c minimum
• After dependency changes: Steps 2a + 2b + 2c
• Before commits: Full Steps 2-4
• After major refactors: Full Steps 2-5 including manual Discord testing