add --with-review and --verbose flags

[lfmundim]

• Added an optional --with-review N parameter
  • when passed, Claude will take a pass over the changes implemented by Mistral as a reviewer and ask for changes if anything is detected. It will iterate N times doing this. Helps offset the manual review needed at the expense of Claude tokens, therefore its optional

• Added an optional --verbose parameter
  • YouTube's "stats for nerds", prints out cost per model once finished. Mainly so you can easily verify if it delegated, how much did it run etc
  • Input and Output lines mean I/O with the selected coder model (e.g. Mistral 3.5 Medium)
  • Total is the sum of I/O
  • Claude line is the projected Claude Token cost if Claude was used through the entire process
  • The big relevant number is the Total line, telling you how many Claude Tokens were "spared" or used on Mistral instead

• Added an optional --allow-ghostwriting parameter
  • Allows Claude to write code instead of Mistral

[pcx-wave]

@lfmundim Thanks for this!

May i pls request a few points of review :

1. Main concern: the hard STOP on a Vibe failure, Wrote nothing happens regularly so i'm afraid it will break the flow too much. Today a failure relaunches and retries up to 3× (SKILL.md:254, Step 6) with claude 'manual finish' so the task still completes (but indeed its not ideal). Halting on the first failure leaves the run stranded. The halt message says "Fix prompt and retry," but Claude wrote that prompt, not the user, so the default should be Claude auto-retries a couple times, then asks (retry / --allow-ghostwriting / abort). Keep "never ghostwrite" but it shouldn't mean "never complete." . Can you pls address this ?

2. Did you test the no-flag path to run smoothly end-to-end ? For easier use it is better to keep no flag.

3. SKILL.md repeats the same review-fields list in 3 places (§8.4, §8.6 JSON, bottom table) and the WROTE_NOTHING rule in 3 (table + dedicated section + orchestration rule). Pls collapse each to one canonical spot, would reduce skill changes ~25 lines with no behavior change, and save context tokens.

Nice work overall 👍

[lfmundim]

Thanks for the feedback!

I had Claude address items 1 and 3, I agree, I was a bit overzealous, I think because Claude was attempting to ghostwrite too much.

Regarding item 2, I'm assuming you'e asking about if I ran /vibe without --with-review and --verbose? I did, it works fine, but without the review pass or the stats at the end