5.7 -- Update integrations

Context

APIs change. Tools evolve. An endpoint that worked yesterday returns a 404 error tomorrow. Your agent uses integrations (Telegram, Vault, PostgreSQL, Docker, GitHub) and each one can break independently.

The challenge: adapt without breaking everything.

When an API changes

Warning signals

• The agent receives 4xx/5xx errors on an endpoint it was using.
• An email from the vendor announces a change (deprecation notice).
• A CLI command no longer works after an update.
• Results have changed format (JSON with different fields).

Update process

1. Identify: which endpoint/command has changed.
2. Read the docs: changelog, migration guide, new API.
3. Test in isolation: call the new version manually before integrating it.
4. Update the workflow: modify WORKFLOWS.md with the new command/endpoint.
5. Test the complete workflow: dry run with the agent.
6. Deploy: update in production.

Update template

## Update integration: [name]

Date: YYYY-MM-DD
Reason: [deprecation / bug / new feature]

### Before
- Endpoint: [old]
- Command: [old]
- Format: [old]

### After
- Endpoint: [new]
- Command: [new]
- Format: [new]

### Impact
- Affected workflows: [list]
- Scripts to modify: [list]

### Test
- [ ] Manual call OK
- [ ] Agent dry run OK
- [ ] Production OK

Adapt without breaking

Rule 1: never in production first

Always test in isolation. A manual curl, a test script, a staging environment. Never directly in the production workflow.

Rule 2: keep the old one as fallback

Don't remove the old integration until the new one is validated in production for at least one week.

# Example: new Telegram endpoint
NEW_URL="https://api.telegram.org/bot${TOKEN}/sendMessage"
OLD_URL="https://api.telegram.org/bot${TOKEN}/sendMessage"  # same in this case

# Test the new one
if ! curl -s -f "$NEW_URL" -d "chat_id=$CHAT_ID&text=test" > /dev/null; then
    echo "New endpoint failed, falling back to old one"
    curl -s "$OLD_URL" -d "chat_id=$CHAT_ID&text=test"
fi

Rule 3: one change at a time

Don't update Telegram, Docker, and the GitHub API on the same day. If something breaks, you won't know what. One change, one test, one validation.

Skills up to date

If your agent uses skills (MCP, plugins, extensions):

Check available skills

List all your skills/tools available.
For each one, tell me:
- Name
- Last used
- Is it still working? (test it)

Update a skill

1. Check the current version.
2. Read the changelog for the new version.
3. Update in a test environment.
4. Test the features you use.
5. Deploy in production.

Remove unused skills

Each skill is a maintenance point. If you haven't used it for 2 months, disable it. This reduces the attack surface and simplifies debugging.

Common mistakes

Ignoring deprecation notices. "It still works, I'll see later." Then the API cuts off the old endpoint on a Sunday evening.

Updating blindly. pip install --upgrade or npm update without reading the changelog. Breaking change surprise.

No fallback. The old one is removed, the new one doesn't work. You're stuck.

Forgetting to update workflows. The integration is updated but WORKFLOWS.md still references the old command. The agent uses the old workflow and it crashes.

Steps

1. List all integrations of your agent.
2. For each one, note the current version and date of last verification.
3. Subscribe to changelogs/release notes for critical tools.
4. When an update is necessary, follow the process above.
5. Update WORKFLOWS.md after each integration change.

Verification

• [ ] The list of integrations is documented with versions.
• [ ] Each integration has been tested in the last month.
• [ ] Deprecation notices are being tracked.
• [ ] WORKFLOWS.md is up to date after each integration change.
• [ ] A fallback exists for critical integrations.