Copy Connections Credentials

Copy Connections — Credentials (post-execute)

This skill runs after execute has created the connection shells. It helps the user get credentials attached to their connections so the schema skill can run next.

Fivetran UI link format

The direct link to a connection's setup page is:

https://fivetran.com/dashboard/connections/<connection_id>

Where <connection_id> is the target connection ID returned by create_connection during execute (recorded in .copy-connections/results.md).

Core approach

Be flexible and user-friendly. Different users will want to handle credentials differently — some will hand you a file with everything, some will want to go connection-by-connection, some will just want links to the Fivetran UI. Adapt to the user rather than forcing a specific flow.

What you know

During scope, get_connection_details returned the full config for each connection. Credential fields came back masked as "********". The plan records which fields are masked and classifies each connection's auth style:

• OAuth — has OAuth-related masked fields (refresh_token, access_token, oauth_*). These cannot be patched via the API — the user must complete OAuth in the Fivetran UI via a browser redirect. Always provide the UI link for these.
• Credential-based — has standard masked fields (password, api_key, api_secret, etc.). These can be patched via the API using modify_connection.
• None — rare; some connectors don't need auth.

Use this information to tell the user what each connection needs and what their options are.

Guidelines

1. Start by showing the user what's needed. For each connection, explain what kind of auth it uses and what fields need to be filled. Group by auth type so the picture is clear.

2. For OAuth connections, always provide the Fivetran UI link. There's no API path for OAuth — be upfront about this. Make it easy: give them the direct link to each connection's setup page.

3. For credential-based connections, let the user choose how to provide them. Options include:

   • Providing a file with all credentials at once
   • Giving them to you inline (all at once or one-by-one)
   • Doing it themselves in the Fivetran UI (give links)
   • Any combination — some here, some in the UI

4. If they provide a file or bulk credentials, check for gaps. If any connections are missing, ask about the remaining ones rather than silently skipping them.

5. Don't be rigid about the collection method. If they start giving you credentials inline, don't stop them to ask about a mode. If they say "here's a file," read it. If they say "just give me the links," give links. Match their energy.

6. Patch credentials via modify_connection with the credential field values. Always use run_setup_tests: false — setup tests are the schema skill's job, and only run_connection_setup_tests (not modify_connection with run_setup_tests: true) triggers schema discovery. Process and report results.

7. After patching, report what succeeded and what failed. If something fails, let them know they can fix it in the UI or retry.

What this skill does NOT do

• It does not run setup tests. That happens in the schema skill.
• It does not apply schema config. That happens in the schema skill.
• It does not unpause connections.

Required checkpoints

1. Confirm execute completed first. If .copy-connections/results.md doesn't exist, tell the user to run execute first.
2. Never log, echo, or persist credential values. Security rules below.
3. Update results.md with per-connection credential outcomes.

After credentials are handled

Prompt the user about the next step: running setup tests and applying schema config via the copy-connections-schema skill. If some connections still need credentials (OAuth not done yet, some skipped), mention those and suggest coming back after they're handled.

Security rules

These are non-negotiable.

1. Never log credential values. Refer to the field name (e.g., "attached password for postgres_app") not the value.
2. Never echo credential values back to the user. Say "got password" not "got ".
3. Never write credential values to any artifact. Results.md, logs, plan — none of these contain credential values.
4. Never send credential values anywhere but Fivetran. The only valid destination is the Fivetran API.
5. If the user provides a credentials file, check for .gitignore protection. If in a git repo and credentials.yaml isn't in .gitignore, add it and tell the user. Prompt for cleanup when done — don't auto-delete.

Error handling

• Credential patch fails for a connection. Report the error and move on. The user can retry or fix it in the UI.
• User provides partial credentials. Handle what's there, ask about the rest.
• User abandons mid-flow. Fine. What's done is done. Results.md reflects the current state. They can re-run later.

Re-running this skill

Users may run this multiple times. On re-run, check results.md to see what still needs credentials and focus on those. Offer to re-patch already-done ones too (e.g., rotated key).

Key product knowledge

Credentials don't make a connection exist; they make it connectable. The connection shells are already real connections in Fivetran — just paused and untested.

OAuth truly cannot be done from the CLI. It requires a browser redirect. Provide the link and move on.

Setup tests are NOT part of this skill. They happen in the schema skill after credentials are in place.

Anti-patterns to avoid

• Don't force a specific collection method. Let the user drive.
• Don't echo credential values.
• Don't persist credential values to any artifact.
• Don't attempt OAuth from this skill.
• Don't run setup tests. That's the schema skill's job. Never pass run_setup_tests: true on modify_connection.
• Don't apply schema config. That's the schema skill's job.