sebastienrousseau/dotfiles/vibe

Vibe Orchestrator

Ported from pcx-wave/vibe-skill. This dotfiles edition deploys the skill files to ~/.claude/skills/vibe/ via chezmoi, plus a dot ai delegate / dot ai cost shim that lets non-Claude-Code users access the same delegator.

Why

Claude (Sonnet 4.6) costs $3/M input, $15/M output. Cheap delegates:

• Mistral Medium 3.5: $1.50/$7.50 per M tokens (or $0 within Le Chat Pro sub)
• DeepSeek V4 Flash: $0.14/$0.28 per M tokens

For coding tasks that do many file reads, the delegator pattern saves massively because Claude sees a single tool call + final diff instead of every intermediate read/edit.

How to use

From Claude Code (preferred)

/vibe rename every UserService method that starts with "get" to "fetch"

Claude writes the prompt → invokes tools/vibe-delegate → reads the streaming JSON output → reviews the resulting git diff. You see one compact report instead of 30 file reads.

Toggle auto-delegate mode with /vibeon / /vibeoff. While ON (~/.local/share/vibe-auto.flag exists), Claude is asked to send coding tasks through Vibe by default. The flag is advisory: no hook enforces it.

From the terminal (no Claude Code)

dot ai delegate "rename every UserService method that starts with get to fetch"
dot ai cost --since 7   # spend report

The CLI shim wraps the same delegator. Runs are logged to the SQLite database ~/.local/share/dotfiles-ai.db; an older ~/.local/share/delegate-runs.jsonl is imported once and kept as a fallback.

Model selection

Default: whatever your ~/.vibe/config.toml lists as default_model.

Override per session:

/vibe-model-pick deepseek-v4-flash

The override writes a flag at ~/.local/share/vibe-model.flag consumed by tools/vibe-delegate. Clear with /vibe-model-clear.

Reports

/vibe-report                # all-time overview
/vibe-report --since 7      # last 7 days
/vibe-report --project foo  # one project
/vibe-report --fails        # only failures + breakdown by failure type

Reports show: total runs, success rate, tokens delegated, delegate cost, Claude-equivalent cost (the would-have-been spend), savings ratio, and failure breakdown (timeouts, exit_err, syntax errors, search/replace fails, empty runs).