superteam-non-interactive

Deprecated: this skill remains installable for historical CI compatibility, but it is no longer the recommended Patina Project workflow for new automation.

superteam-non-interactive is the CI-safe Superteam entry point. It reuses the superteam teammate contracts, but removes human-in-the-loop pauses so GitHub Actions can keep running without waiting for chat approval, confirmation, or follow-up choices.

Quick start

Use this skill when the operator invokes /superteam-non-interactive, asks for Superteam in GitHub Actions, or requests a headless one-shot Superteam run.

Load the sibling Superteam files from skills/superteam/: SKILL.md, pre-flight.md, routing-table.md, and project-deltas.md.
Resolve the issue from an explicit #<number>, SUPERTEAM_ISSUE, or the current branch name. If none exists, halt.
Run the normal Superteam pre-flight with interaction_mode=non-interactive.
Route to the owning teammate using the normal Superteam routing table.
Before delegation, load the routed teammate contract from skills/superteam/agents/ or skills/superteam/.claude/agents/.
Before any place where superteam would ask the operator, use the non-interactive autonomy rules below. Do not stop solely to wait for a human approval, confirmation, or policy choice.

Non-Interactive Rules

Never ask a question, wait for approval, present choices, or rely on a future chat reply.
Never infer missing required context from vibes. Use only invocation inputs, environment variables, repository state, host capabilities, and the autonomy defaults documented here.
Never halt just because an interactive Superteam run would have asked a human to approve, choose, confirm, or continue. In non-interactive mode, convert those points into deterministic action, explicit safe defaults, or a machine-actionable blocker.
Use normal Superteam branch switching, dirty-worktree refusal, project delta handling, model selection, execution-mode binding, feedback routing, and latest-head shutdown.
Treat ambiguous operator text or conflicting durable state as a blocker, not as a request for clarification.
Prefer explicit invocation inputs over environment variables; prefer environment variables over repository state.
Output concise status lines that CI logs can preserve: current phase, issue, branch, PR if any, blocker if any, and verification evidence.

Autonomy Defaults

Unset policy inputs must not create a human approval stop. Apply these defaults unless an explicit input overrides them:

Gate 1: auto-advance a clean or fully dispositioned design. The normal Superteam evidence remains required, including a real adversarial review and no approval-blocking finding.
Design findings: route fixable design findings back to Brainstormer in the same run. Halt only when a finding remains blocked after the routed teammate has acted or the blocker is not resolvable by the workflow.
Publishing: allow Finisher to push, create, or update the PR when the host token and branch permissions allow it. SUPERTEAM_ALLOW_PUBLISH=0 is the opt-out for dry-run or diagnosis jobs.
PR shape: create draft PRs by default. SUPERTEAM_DRAFT_PR=0 requests a ready-for-review PR after local review passes.
CI and feedback: if checks are pending, queued, or temporarily unavailable, report monitoring with latest-head evidence and exit without asking for a human. A later run resumes the latest-head gate.
Merge: do not merge by default. Merging requires an explicit repository automation outside this skill or a future documented input.

Inputs

At minimum, one issue source must be present:

Prompt token: #71
Environment: SUPERTEAM_ISSUE=71
Branch: 71-kebab-title

Optional CI policy inputs:

SUPERTEAM_APPROVE_CLEAN_DESIGN=0: disables automatic Gate 1 advancement and turns a clean design into a blocker for diagnosis. The default is 1.
SUPERTEAM_ALLOW_PUBLISH=0: prevents Finisher from pushing or opening/updating a PR. The default is 1.
SUPERTEAM_DRAFT_PR=0: asks Finisher to create a ready-for-review PR when publishing. The default is 1 (draft).

Unset optional policy inputs use the autonomy defaults above. Do not prompt to enable or confirm them.

Gate Behavior

Gate 1 remains real. In non-interactive mode:

No design exists: Brainstormer may write and commit one.
Design exists but no plan exists: Team Lead advances to Planner when the normal Gate 1 evidence is clean or dispositioned. This is the non-interactive replacement for a human approve token.
Gate 1 evidence is missing: generate or rerun the missing evidence when the workflow can do so. Halt only if the evidence cannot be produced by the current run.
Gate 1 evidence is blocked or ambiguous: route fixable issues back to Brainstormer; halt only for unresolved material blockers with superteam halted at Gate 1: <reason>.
Plan exists: execution may continue from the committed plan without a chat approval prompt.

Publish behavior remains Finisher-owned. Finisher may push or open a PR only when local review has passed and host permissions allow the operation. If SUPERTEAM_ALLOW_PUBLISH=0, halt with the exact publish action that would have been taken.

Halt Format

Every non-interactive blocker must be explicit and machine-actionable:

superteam halted at <teammate or gate>: non-interactive blocker: <reason>

For CI readability, include a final next_run: line when a later run could resume automatically, or next_input: only when a concrete machine-provided input (not a human approval) would unblock the run.

Red Flags

Prompting the operator in a GitHub Actions run.
Silently treating missing approval as approval.
Halting for SUPERTEAM_APPROVE_CLEAN_DESIGN or SUPERTEAM_ALLOW_PUBLISH being unset.
Skipping adversarial design review because the run is headless.
Treating a blocked or ambiguous Gate 1 review as approved.
Publishing after SUPERTEAM_ALLOW_PUBLISH=0.
Reporting complete before the latest-head PR completion gate passes.