RUMBO Agent Reliability

View saved version →

---
name: audit-final-state
description: Independently audit an agent task before accepting its final state.
---

# Audit Final State

Audit:
1. Workspace identity.
2. Canonical authority.
3. Exact current state/head.
4. Scope and permissions.
5. Tool, skill, and dependency provenance.
6. Human approval for consequential actions.
7. Execution receipt.
8. Expected postcondition.
9. Unexpected side effects.
10. Rollback or recovery state.
11. Cross-agent handoff.
12. Drift since start.

Verdict:
- GO: all required controls evidenced.
- CONDITIONAL: blocker has owner and next action/date.
- NO-GO: acceptance would exceed evidence or authorization.

A passing process is not proof of a passing outcome.

View saved version →

---
name: canonical-state-recovery
description: Recover authoritative current state before an agent changes code, data, configuration, or external state.
---

# Canonical State Recovery

Use before consequential work.

1. Identify the authoritative system and exact scope.
2. Record repository/project, branch, exact head/version, files, and environment.
3. Separate CURRENT, HISTORICAL, CANDIDATE, and ARCHIVED state.
4. Record decisions, blockers, negative memory, and next gate.
5. Prefer fresh direct receipts over narratives or stale documents.
6. Label missing evidence NOT_PROVEN or UNRESOLVED.
7. Stop when authority, scope, or current state is ambiguous.

Required output: Objective, Authority, Current state, Decisions, Blockers, Next gate, Unresolved.

Never promote prepared, historical, or candidate material to CURRENT without evidence.

Truth rules:
prepared != executed
executed != verified
candidate != production
historical != current

View saved version →

---
name: deep-research-reconcile
description: Reconcile conflicting evidence across sources while preserving provenance and uncertainty.
---

# Deep Research & Reconcile

Use when multiple sources disagree or current state is unclear.

1. Gather primary evidence first.
2. Rank evidence by authority and freshness.
3. Preserve contradictions instead of overwriting them.
4. Classify claims PROVEN, NOT_PROVEN, UNRESOLVED, HISTORICAL, CANDIDATE, or FAIL_CLOSED.
5. Re-run only checks that can change a decision.
6. Produce one reconciled state and identify its evidence.

Evidence precedence:
1. Fresh authoritative receipt.
2. Exact current-head/object verification.
3. Durable same-scope artifact.
4. Explicit correction with provenance.
5. Historical narrative.

Never use a model assertion, UI state, old PASS, or score as proof of current execution.

View saved version →

---
name: execute-verify-close
description: Execute an authorized bounded action, verify the actual effect, and close only with evidence.
---

# Execute Verify Close

Use only when exact target, intended effect, authorization, rollback/no-effect path, and verification criterion are known.

State machine:
READY -> AUTHORIZED -> EXECUTED -> VERIFIED -> CLOSED

Failure path:
EXECUTED -> FAILED -> RECONCILE

Never infer success from process exit, queue acceptance, UI feedback, model output, or a prepared artifact.

Verify:
- exact target
- expected postcondition
- unintended changes
- execution/effect receipt
- final authoritative state
- rollback/no-effect status

Use NOT_PROVEN when the effect cannot be independently evidenced. Only mark CLOSED when required evidence exists.

View saved version →

---
name: goal-loop-controller
description: Keep agent work aligned to one measurable goal with explicit gates and stop conditions.
---

# Goal Loop Controller

Run: GOAL -> STATE -> OPTIONS -> NEXT ACTION -> VERIFY -> RECONCILE -> GATE.

For each iteration:
1. State one concrete goal.
2. Read current state before acting.
3. Pick the smallest action that can materially advance the goal.
4. Stay inside authorized scope.
5. Verify independently when possible.
6. Update state and evidence.
7. Stop when the next action needs missing authority, credentials, external approval, or unproven evidence.

Do not repeat unchanged diagnostics, create parallel systems, reopen closed paths without new evidence, or expand scope merely because a path is blocked.

Stop on missing authority, ambiguous state, destructive remediation without explicit approval, missing receipt, or unsatisfied production gate.