Harness permissions

Who may call which tool is decided in one place: Claude Code's own permission rules in .claude/settings.json. WUWEI does not keep a second allow list.

wuwei init merges three sets of rules into the project settings and leaves your own untouched:

Rules	Contents	Why
allow	git, gh, reading, writing and searching files	everyday work should not prompt you
deny	force pushes, admin merges, repository deletion, edits to the ledger	the doors, held by the harness as well as the hook
ask	terraform, tofu, pulumi, kubectl and helm writes	infrastructure is your decision

Deny rules and WUWEI's hook refusals take precedence over allow rules, so allowing broadly is safe: the doors stay held.

Allowing a tool

To let the team post to Slack or update Linear, allow the tool the way you always would: answer "Yes, don't ask again" on the prompt, or add mcp__slack (a whole server) or mcp__linear__save_issue (one tool) to permissions.allow. WUWEI still scans every outward call for secrets and the workspace canary, whatever the rules allow.

Fewer prompts

If you still see prompts you always approve, Claude Code's /fewer-permission-prompts skill reads your transcripts and proposes allow rules for them.

Why: a permission system you maintain twice drifts, and a plugin that re-asks what the harness already allows is friction for nothing.