vibe

VBW

Kernel decides next step; you carry it out. Even for a one-line fix: no project file changes until the contract is approved and a build runs.

"${CLAUDE_PLUGIN_ROOT}/bin/vbw" next --json 2>&1 || true
"${CLAUDE_PLUGIN_ROOT}/bin/vbw" config models 2>&1 || true
"${CLAUDE_PLUGIN_ROOT}/bin/vbw" workflows on 2>&1 || true
"${CLAUDE_PLUGIN_ROOT}/bin/vbw" config autonomy 2>&1 || true

User said: $ARGUMENTS

Every new request, now or later, goes to vbw:triage first; "what's next" or "suggest next" follows vbw:whats-next.

No JSON above (skill shell execution off): run VBW_SESSION_ID=${CLAUDE_SESSION_ID} vbw next --json, vbw config models, vbw config autonomy with Bash first. "not a VBW project": set it up yourself (vbw init, vbw statusline on; say what was set up in one line), then VBW_SESSION_ID=${CLAUDE_SESSION_ID} vbw next --json again.

Setting stated in words

Before any step, read the user's words (now, or later in the session). Words naming guided, balanced or hands-off, or a plain phrase clearly meaning one ("ask me before each step" = guided, "do not ask, just go" = hands-off), run vbw config set autonomy <value> first. Say in one line what was set, then follow the Autonomy section with that value, not the line loaded above.

• Same as the current setting: it writes nothing and says nothing about it.
• Words that do not clearly name one setting change nothing: never guess.
• Stated later in the session: record it at once; it applies from the next step. A switch to guided while the autonomous run is armed runs vbw auto off ${CLAUDE_SESSION_ID}; a switch to balanced or hands-off runs vbw auto on ${CLAUDE_SESSION_ID} once.
• --auto runs alone this time only and never changes the setting.

Autonomy

Last line above, or the value just set, = how much VBW does alone. Approval, accept and shipping always stop for user.

• balanced, hands-off: run vbw auto on ${CLAUDE_SESSION_ID} once; VBW keeps going until a step needs user.
• hands-off: at needs_decisions, take the recommended option (as decide and tell me below).
• guided: before each step not needing user, say what and why, ask (AskUserQuestion): "Go ahead", "Explain more", "Stop here". --auto runs alone this time.