Supergoal

About

One objective -> smallest correct change -> verified against ground truth. If invoked, run this skill's contract instead of downgrading to an inline shortcut. SKILL.md is the router; reference/ carries procedure. Unless explicitly invoked, pure brainstorming and user-driven step-by-step work use normal direct collaboration.

Standing rules (read first, every mode). Before classifying the mode, read .supergoal/rules/RULES.md if present. Honor it across phases as top-priority preference, but rules never weaken safety gates. Create/edit it only when the user explicitly asks (reference/rules.md). Check only the repo root (.supergoal/rules/RULES.md), and skip the check entirely in an ephemeral single-task workspace (auto-LIGHT tier; see role-loop ## Tier selection).

Core principles

• Ground truth beats proxy: re-run REAL tests, re-read request/docs, and do not optimize to self-grading.
• Exact proof beats review; implementation is delegated to a builder subagent.
• LIGHT tier substitutes in-context equivalents for vault artifacts, gate scripts, the Z file, and the builder dispatch; its approval rule and commit bar live in reference/role-loop.md ## Tier selection.
• Smallest correct change; match surrounding code. Scope-minimalism governs code surface area, not UI quality: polished user-facing UI is baseline correctness.
• GREENFIELD / DEBUG / LEGACY code changes use Before/After Eval before Build: prove before, target after, and delta with trusted commands (reference/delivery-gate.md).
• Ask only when genuinely ambiguous; resolve code-answerable questions by reading the code.
• Report for humans: outcome first, Simplified Technical English prose, the project's own vocabulary (reference/reporting.md); machine-checked markers stay verbatim.
• Docs language: for persistent repo docs (docs/**, run vaults, .domain-agent/**, ADR/spec/changelog), match the target repo's dominant prose language; mixed/none -> the user's language. Keep identifiers, paths, commands, and machine-checked anchors in canonical English so checks keep matching.
• DEBUG done-bar: a green reported-repro is NOT done. Before commit, print three literal lines and satisfy them - GATE.owner= the invariant-owning frame (for a raised exception: the frame that RAISES it via the traceback, not where it surfaces; for recursion: inside the enumerated cycle), GATE.alt_repro= a structurally different second repro now passing, GATE.conformance= raw-literal returns reachable through the patched path (and its symmetric sibling methods) converted to the module's canonical forms (S.One over 1). Full rules: role-loop "DEBUG hidden-contract gate".
• Hard stops: a destructive or irreversible step (drop data, force-push, external publish) needs explicit consent; if the real tests cannot pass, report it - never fake a pass.

Run isolation (GREENFIELD / DEBUG / LEGACY that edits code)

After mode detection, resolve the source/base branch and target/integration branch (repo policy, else ask). Verify both refs before mutating files, then create a run worktree from the source/base branch. Do all code work there. Do not mutate the original checkout. Commit or merge only into the verified target/integration branch after verification and user acceptance. Commit is hard-gated by the Commit gate (reference/delivery-gate.md, backstop templates/commit-gate.sh): non-green means fix/ask, never commit on assumption. Full contract: reference/role-loop.md.

Mode (classify, state it in one line)

Tier (code modes; state it with the mode line). GREENFIELD / DEBUG / LEGACY also declare LIGHT / STANDARD / DEEP per reference/role-loop.md ## Tier selection. User words "quick"/"light" or "thorough"/"deep" override detection; upgrade is one-way, never downgrade.

The no-code/utility/planning modes - QA-ONLY, REVIEW-ONLY, ARCHITECTURE, WAYFINDER, PROTOTYPE, TEACH, LEARN-DOMAIN, HARNESS-EVAL, SKILL-MINE - write no product code by default and confirm before installing anything. PROTOTYPE may write throwaway sandbox code; it cannot ship until routed back through delivery.

UI/UX overlay (any mode shipping user-facing UI). Load reference/ui-ux.md at Frame; apply the Expressive/polished baseline by default (reference/taste-skill-v2.md is the authority for ALL user-facing UI), through Build and Verify. GREENFIELD frontend: always; LEGACY: only new UI (else reuse the existing design system); non-visual work (lib, API, backend, CLI): skip.

Board overlay (optional). If the live dashboard is enabled, the conductor calls sg-emit at each phase transition; it observes only, never gates (reference/observability.md).

Default loop (GREENFIELD / DEBUG / LEGACY) - five gates, fresh context per gate

Load and follow reference/role-loop.md; it is the sole detailed authority for run setup, vault lifecycle, role inputs/outputs, retries, verification, and finalization. Invoking supergoal for these modes is explicit authorization to use its fresh-context subagents; ask again only for normal safety or permission gates. Red-green evidence is required, plus DB evidence when persisted data is load-bearing.

Mandatory core: Frame -> Plan approval -> Build -> Exact Verify/QA -> Finalize. Use one builder + one auditor verifier per iteration; browser/CLI proof adds one evidence-only qa-tester before the auditor. Only a named, recorded escalation trigger permits the conditional plan attack. Frame writes GOAL.md first and freezes a self-sufficient PLAN.md; Build starts only after approval and runs in a separate fresh-context builder from that plan; qa-tester captures the promised browser/CLI evidence, then a fresh adversarial verifier (qa-auditor) reruns REAL tests, audits the promised E2E/live/API/browser proof, and owns the final verdict, GOAL ticks, and R-LOOP. Finalize requires every criterion green, the completion marker, user acceptance, and the commit gate. Exact verification outranks review.

Roles -> personas: builder/improver=agents/executor.md, evidence-only browser/CLI tester= agents/qa-tester.md, final verifier for every default-loop path=agents/qa-auditor.md, escalation reviewer=agents/code-reviewer.md, security=agents/security-reviewer.md (others in agents/<role>.md).

Reference map (load only what the current phase needs)

Done = mode stated; smallest diff; Before/After Eval complete for code-mode changes; REAL tests + request/docs green (not proxy); runtime MUST proven by real behavior; code-mode runs use red-green test + DB evidence if data load-bearing; neighbor snapshots re-run with unnamed drift resolved; every GOAL.md Success Criterion checked, with no orphan scope; Z-<date>.md written with run branch + completion timestamp; DEBUG prod issue has reproduction fidelity and, if non-exact, residual risk + post-deploy confirmation plan; user-facing UI at the Expressive baseline; destructive steps consented; commit/merge only after the commit gate passes (reference/delivery-gate.md); verified commands reported.