rdleclerc Ubergoal
Core rule
ubergoal: thin lifecycle wrapper and goal owner; create or bind the goal, classify risk, bind
$uberplan or a Tier 0/1 micro-intent by path/revision, execute that authority,
then send diff/evidence to $uberaccept. Benefit >> cost.
Routing
| Need | Route |
|---|---|
| Goal launch | platform goal primitive; references/goal-objective.md; validate with scripts/validate_goal_objective.py --target-chars 3400 --strict-target |
| Tier 0/1 plan | ${UBER_GUIDE_ROOT:-~/repos/agentic-architecture-guide}/docs/coding-agent-work-contract.md |
| misunderstanding-prevention | $uberplan Task Understanding Review |
| Rigorous plan or user expectation / surprise assessment | $uberplan |
| Everyday defect; repeated incident | ../references/debug-loop.md; repeated -> $uberrca |
| Architecture-shaped failure | route to $uberarchitect before $uberplan |
| policy-adherence / final proof | $uberaccept |
| Complexity/dead code | $ubersimplify |
| Source or artifact adoption research | $uberassess |
| Post-run learning | $uberskillevolver |
| Dispatch/runtime states | ../references/dispatch-and-sessions.md; ../references/operational-states.md |
Refactor campaign / ubercampaign |
references/refactor-campaign-profile.md; references/campaign-profile.md |
For material unexpected test failures, stop before five consecutive failures,
run $uberrca, revise through $uberplan, and merge any changed scope. For Recurring loop mode, set
loop_mode, read ../references/loop-engineering.md, and do not create uberloop.
Use templates/uber-run-receipt.md only for an activated receipt.
Approved-plan execution boundary
Before editing, record plan identity, scope, non-goals, proof,
candidate_kind, change_budget, and mechanism_inventory. Read sources,
then owner, consumer, and proof. Contradiction requires replanning.
- Execute planned work, repository-mandatory gates, and evidence-activated safety protections.
fix_within_scopepermits only a bounded correction already authorized.replanapplies when evidence changes the implementation contract without changing operator authority.user_decisionapplies before changing operator-owned product scope, ownership, architecture/dependency commitments, public behavior, acceptance, or external-action authority; internal contract changes stayreplan.- Planning never authorizes an external action. Preserve exact target, authorization, idempotency, rollback, receipt, and authoritative readback.
- Do not add generic docs, refactors, broad tests/evals, reviewers, subagents, or lifecycle artifacts unless the plan, repository contract, or observed risk activates them.
- Compare actual files, additions/deletions (churn), and new mechanisms with
approved budget after each implementation step. An overrun is
replan, oruser_decisionwhen it changes an operator-owned cap or commitment; never self-waive it.
After each step, record proof and one state: continue, fix_within_scope,
ready_for_acceptance, replan, user_decision, or blocked. continue
requires a runnable approved step and no missing gate. Missing operator authority
is user_decision; missing proof is blocked. Return the literal state, barrier,
and safe next actions.
A proved no_change candidate uses states: complete the partial-fix check,
attach evidence, return ready_for_acceptance, and let $uberaccept decide.
Do not invent a lifecycle state or cosmetic diff.
ubergoal never issues accepted, complete, or SHIP. No reviewer verdict
is UberAccept. After plan, proof, and tier review pass, return
ready_for_acceptance and invoke $uberaccept.
Review ladder
| Tier | Work class | Required review |
|---|---|---|
| 0 | typo/cosmetic only — nothing that fixes observed misbehavior | none; commit carries a tier0: trailer |
| 1 | contained single-surface change, clear tests | one exact-diff review pass by a capable lane, including a one-line scope echo against the operator-original ask |
| 2 | cross-repo doctrine/pointer edits; behavior surfaces (prompts/skills/evals); medium-risk code | exact-diff review + independent adversarial lane (different vendor or fresh context) + scope-fidelity verdict |
| 3 | production/runtime services; live-injected context surfaces; provider routing; security/data-subject surfaces; major refactor / mass deletion | full 4-phase ladder (plan review, exact-diff, adversarial, acceptance) using fresh independent reviewer contexts at the highest model and effort allowed by active project policy + review-board lanes |
Tier 2 uses one independent risk-specific lane. Tier 3 keeps risk-activated review-board lanes. Effort scales with tier; active project policy binds model/effort. Repository ladder is canonical; if uncertain, take the higher tier.
Micro-intent fast path
For Tier 0/1 low-risk work without runtime/provider/security/data-subject or
cross-boundary ambiguity, use 2-3 sentences of scope / intent, criteria,
non-goals, verification, candidate_kind, change_budget, and
mechanism_inventory. Include
failure_case_id | case_updated | not_applicable_with_reason.
Scope fidelity and completion
Preserve the operator original instruction, agent-interpreted scope, proposed narrowed scope, explicit deferrals/non-goals, approval evidence, and Scope fidelity verdict. Use a separate ledger only for requested long-running, resumable, multi-owner, or handoff work.
Do not call update_goal(status="complete") until $uberaccept returns
accepted, each touched repo is committed/reverted/approved-dirty, and the
handoff reports git state and gaps. Other states remain non-complete. File
failure intake only when the terminal state requires it.
Optional Claude adversary
Use only when the operator explicitly requests Claude by name.
A generic cross-model request does not authorize selecting Claude.
Do not invoke Claude or alternate reviewer from task similarity.
Contract (subprocess reference-following proven): ../references/claude-adversary.md.
Ask: Load-bearing goal? Skip test. Testable decomposition.