SuperAPI AI Directory

quack

Agents & models

View repository

Updated 10-16 01:57

About this skill

Skill: quack

Explicit route control 🦆. Alias-first auto-route, else one targeted disambiguation question.

Purpose

Provide explicit user-controlled routing for workflow-like requests while preserving safety and approval constraints.

Philosophy Guardrails (skill-local)

Inherit shared guardrails from references/GUARDRAILS.md.

  • ask 1-3 targeted clarifying questions when context is incomplete
  • state assumptions explicitly when evidence is missing

Activation

Use only when user explicitly invokes quack; do not auto-activate from inferred intent.

Method

  1. Verify explicit quack invocation.
  2. Normalize intent: strip separator after quack (:, -, —), strip outer matching quotes, strip trailing punctuation (?!.;:).
  3. Bare quack (heartbeat path):
    • If normalized intent is empty or whitespace-only:
      • Read assets/heartbeat.md and select one random heartbeat line from the list
      • Read and emit full assets/quick-help.md content verbatim (no summarization)
      • Emit closing prompt: What would you like to route?
      • Stop (do not generate ad-hoc quips, do not enter disambiguation)
  4. Parse optional override: use <subagent> or with <subagent> or via <subagent>.
  5. Validate override (if present) against platform subagent list; if invalid, ask one correction question: Need one detail: unknown subagent "<x>". Use duckling or general? and stop.
  6. Determine effective subagent: if override parsed in step 4, use that; otherwise default to duckling.
  7. Determine route execution policy:
    • inline-default: duck-design, duck-teach, duck-debug, duck-debt
    • delegated-default: duck-patch, duck-refactor, duck-risk, duck-review, duck-triage, duck-simplify
    • user override forces delegation regardless of default policy
  8. If explicit full skill name in input, resolve that skill and execute route (see step 11).
  9. Otherwise, load assets/route-aliases.json and attempt case-insensitive match.
  10. Apply keyword-based precedence (when multiple skills match):
    • Scan intent for precedence keywords:
      • Risk signals (rollback, breaking, compatibility, compat, failure, impact) -> prioritize duck-risk
      • Complexity signals (overengineered, complexity, simpler, too many, bloated) -> prioritize duck-simplify
      • Learning signals (why, how, what does, explain, teach me) -> prioritize duck-teach over duck-debug
      • Test signals (test, coverage, should I test, before PR) -> prioritize duck-triage
      • Design signals (choose, tradeoff, approach, option) -> prioritize duck-design
    • Reorder matched skills with precedence-matched skill first
    • Still present alternatives (precedence aids ranking, doesn't eliminate choice)
  11. If multiple aliases match after precedence, apply tie-break: exact match > longest alias > ask one disambiguation question.
  12. Route execution (explicit skill name or alias match):
    • Load role instructions from assets/subagent-runbook.md for resolved skill
    • If inline-default policy AND no user override:
      • execute skill inline first
      • if execution fails, ask one corrective question and stop
      • emit Routing: <skill>. only after inline execution step completes
    • Otherwise (delegated-default OR user override):
      • delegate to subagent using harness-native dispatch tool with subagent=<effective_subagent> (determined in step 6, defaults to duckling) first
      • pass skill_name=<resolved_skill> parameter
      • if dispatch fails, ask one corrective question and stop
      • emit Routing: <skill> via <subagent>. if user override supplied, else Routing: <skill>. only after dispatch succeeds
    • Do not stop at routing text alone when execution path is available.
  13. Alias miss (disambiguation):
    • Detect intent fragment and ask one targeted question:
      • debug-ish (error/fail/trace/stack/broken): Need one detail: is this debug, trace, or review?
      • rollout/risk (rollout/migration/compat/rollback): Need one detail: is this risk review or design tradeoff?
      • code-change (fix/change/refactor/clean up): Need one detail: do you want review, patch, refactor, or simplify?
      • tech-debt (todo/defer/fixme/debt): Need one detail: is this debt audit or simplify?
      • unknown: Need one detail: which route fits—review, debug, design, teach, triage, risk, simplify, patch, refactor, or debt?
    • Wait for user clarification, then retry alias resolution.

Boundaries

  • Preserve user decision ownership.
  • If a change would weaken trust-boundary validation, security controls, data-loss prevention, accessibility requirements, or explicit user requirements, refuse it and offer only a safe alternative preserving the constraint.
  • No edits/mutating commands/subagent delegation that changes workspace state without explicit bounded approval.