preInstructions get Cody promps flagged

[vdavid]

Context:
We have an option to add a preInstruction text to Cody in the VS Code extension. Cody adds the preInstruction to the prompt
here.

Problem:
It seems like we are banning users who are using the preInstruction config. It seems quite widespread anecdotally: we’ve seen 3 in the last 24h.

Solution ideas:

1. Disable preinstructions and ship a stable patch release. If this is an experimental feature (need to check!) then this is OK, otherwise a bad idea.
2. Move the preInstruction into a separate message.
3. Check our logic in Cody Gateway to confirm why prompts with the preInstruction are being flagged.

[chrsmith]

Are you sure this is the problem? Cody Gateway just checks that the prompt contains a specific string. So adding anything in addition to the prompt prefix won't cause any problems. (It doesn't even need to be a prefix.)

[chrsmith]

Yes, this seems highly unlikely to cause any problems.

1. We only append the preInstruction to the intro sent in the LLM prefix.

    const intro = ps`You are Cody, an AI coding assistant from Sourcegraph. ${
        preInstruction ?? ''
    }`.trim()

2. We then look for existing prompt strings here:
   https://github.com/sourcegraph/sourcegraph/blob/f716d7aa11bad3bf2d2be74b1d91cd3e535942d0/cmd/cody-gateway/internal/httpapi/completions/flagging.go#L80

3. With the actual configuration data here:
   https://github.com/sourcegraph/infrastructure/blob/main/cody-gateway/envs/prod/cloudrun/main.tf#L54

So using this wouldn't cause the unknown_prompt flag to trigger. However, the request could get flagged if there were "forbidden phrases" in the preInstruction. Or if it were large enough to influence the "token size of the request".