Deprecate attempt_completion tool in favor for a manual action triggered by the user

[M-Gonzalo]

Problem: right now, and no matter what model I use, I end up fighting against countless attempts by the LLM of telling me that the job is complete and I can start another task now. The model will fixate on a particular comment I made earlier, even if it was a simple nudge to take it in the right direction, and will give up after doing that particular subtask, when the real work is far from done. This happens multiple times per session.

Why is undesirable: every mistaken attempt at "completing" the task is expensive in several ways:

• it pollutes the context with a summary of what's done so far
• this leads to having fewer tokens available in the context window for the actual data needed for the job
• it takes the focus of the model out of the original task and into a never-ending race to "be done"
• the summaries generate a lot of output tokens, which are usually an order of magnitude more expensive (in actual money) than input tokens
• the most important one is that it takes up a lot of attention from the developer driving the session, who has to spend a lot of their time getting the model back on task instead of doing whatever they set out to do

Proposal: Completely remove the feature as a callable tool exposed to the models, and replace it with an inconspicuous button that can be manually pressed by the developer when the task is actually done, and then maybe generate the summary, a commit message, or whatever fits the use case.

[elianiva]

highly agree with this, it's so annoying when used in ask mode

although, orchestrator subtask flow may fail if we completely remove this since it's needed to pass context back to the parent

ideally we should be able to disable this, but keep it working for subtask since it's still needed for that use case

[deathtothenig]

I requested this 6 months ago: #406 If you just change the prompts by removing all mention of <attempt_completion> it won't use that tool. That's what I'm currently doing. I just change the prompt to...
"When you think you're done," ask the user if they have any questions or need any further assistance. And that's what Roo does now when they think they're done instead of the green wall of text.

[M-Gonzalo]

Yeah I'm trying out a version of this now. Instead of just deleting it, I appended this to my system prompt:

IMPORTANT: the `attempt_completion` tool, although avalible to you, is ONLY to be used at the behest of the user, when they say some variation of the phrase "Task is now complete". If you think you have completed ALL that was asked of you, including subtasks and revisions, you can ask the user if they think we're done. Nothing more than a question. Otherwise keep working and fullfilling your instructions.

It's working pretty good so far