rpi git:20260909.6f499fbA

Immutable. This exact content is served forever at /api/v1/blob/d98944ee84419e87.

---
name: rpi
description: 'Own an authorized outcome through implementation, checks and fresh final validation; load planning and memory only when useful. Triggers: "run rpi", "run one traversal", "execute this plan", orchestration or worker delegation that implements changes.'
practices:
- bdd-gherkin
- tdd
- design-by-contract
hexagonal_role: domain
consumes:
- plan
- implement
- validate
produces:
- rpi-report.v1
context_rel:
- kind: customer-of
  with: plan
- kind: customer-of
  with: implement
- kind: customer-of
  with: validate
skill_api_version: 1
user-invocable: true
metadata:
  graph_root: true
  tier: meta
  dependencies: [plan, implement, validate]
  capabilities: [own_authorized_outcome, report]
  effects: [dispatch_core_phases]
  canonical_status: canonical
  disposition: keep
output_contract: 'concise human-readable result; optional rpi-report.v1 when a caller or declared consumer requests machine-readable evidence'
---

# RPI

Own the authorized outcome through finish. Use the native coding agent and
shell; BD or the caller's tracker owns work/status/handoffs, Git owns content
and delivery history. Keep one authoritative work account. AgentOps adds a
small operating charter and fresh judgment, not a scheduler or a second queue.

## Operating charter

1. Keep the accepted outcome, scope and real bounds in view. Use the existing
   intent; a trivial clear change needs no Plan, Recall or Learn worksheet.
2. Take the smallest action that advances acceptance or resolves a consequential
   uncertainty. Load [Plan](../plan/SKILL.md) only when intent or approach needs
   shaping. Evidence may disprove an assumption: revise the approach within
   unchanged accepted outcome and scope. Acceptance changes need caller authority.
3. [Implement](../implement/SKILL.md) the change and repair ordinary known
   defects directly. A failed check with an understood cause is implementation
   work, not a reason for another plan, council or helper.
4. Use cheap discriminating checks during edits, then the required integration
   checks. Preserve valid exact-input receipts. Reserve finishing capacity for
   integration, fresh final judgment, repairs and a truthful handoff. Complete
   required checks and known repairs before dispatching final judgment, then
   keep that subject unchanged until the review returns.
5. Obtain [Validate](../validate/SKILL.md) in a fresh author-distinct context over
   the exact final subject. Default to the author's model family; cross-model
   review is opt-in. There is no fixed ten-minute cap. Required caller-selected
   reviewers remain required. Every necessary finding stays visible.
6. Repair actionable findings within authority and real remaining bounds, then
   obtain fresh judgment over the changed subject. Stop at completed acceptance,
   cancellation, an explicit refusal, a spent real bound, or a genuine causal
   stall that the bounded help below cannot resolve. Activity, saved pages,
   changed digests and repeated reviews are not completed capability.

## Causal stall and bounds

Unknown cause, recurrence, no progress or evidence of the wrong objective calls
for causal examination. A genuine causal stall admits **at most one bounded
fresh helper** for that incident, when authorized and within remaining bounds.
Give it the failed assumption, evidence and one discriminating question. Resume
only when the answer supplies a different testable approach; an unhelpful answer
ends the attempt with the unresolved facts. Do not create a helper chain or
rename the same incident to obtain another helper. Known failures get direct
repair. Cancellation, refusal and spent hard time/cost/quota skip help.

Respect actual caller/native limits, including an explicit repair-round bound
when supplied; no invocation, compaction, helper or new subject renews them.
A retry count alone is not a spent time or quota budget. Keep compact recovery
state only when interruption threatens evidence: accepted intent, exact current
subject, useful receipts, unresolved cause, bounds and helper use, in the native
handoff source. Prompt text is no proof of native enforcement.
[Outer-goal guidance](references/outer-goal.md) is optional and outside the core.

## On-demand tools

- Plan shapes missing intent or revises an approach falsified by evidence.
- Implement owns edits, direct repairs and factual checks.
- Validate owns fresh exact-subject judgment; the author cannot issue binding PASS.
- [Memory](../memory/SKILL.md) recalls applicable reviewed topic pages, or performs
  separately budgeted mining and curation when requested. It is never an entry
  or completion toll. No-match and no-change are valid.

Specialists, including anti-ceremony, premortem, council, research and runtime
adapters, remain optional. Risk increases evidence depth; it does not mandate a
specialist dispatch. Read [boundaries](references/boundaries.md) when a source,
review or delivery boundary matters; do not turn the charter into another packet.

## Evidence and report

Bind the accepted intent and exact subject for the fresh validator, derive
complete changed paths and factual check receipts, and disclose orphaned
acceptance evidence when the change affects it. Use the existing provenance
helpers described in Validate; new proof uses caller-selected protected external
non-Git storage. Preserve legacy `.agents/` evidence. Do not invent an identity
or treat a model's declared role as freshness. Missing freshness or necessary
evidence means NOT_PROVEN; proven acceptance failure means FAIL.

Return the caller-visible result, changed subject, strongest checks and material
unchecked acceptance. PASS requires all acceptance, exact identity and empty
`not_checked`; the report never hides remaining work. `NOT_PLANNED` and
`NOT_BUILT` describe progress, not semantic judgment. Do not append a next action
as a substitute for finishing authorized work. The interactive response is the
default; persist `verdict.v2` or `rpi-report.v1` only when the caller requests
machine-readable evidence or a declared consumer requires it. When no machine
artifact was requested, do not create a hidden one.

## Prompt

```text
Use rpi to finish this accepted change within its scope and remaining deadline.
Repair understood failures directly. If an assumption fails, revise the approach
without changing acceptance. Run required checks and obtain fresh final Validate.
Use Memory only if an applicable prior constraint would change the next action.
```

## It's working if

A clear small edit reaches checks without planning or memory paperwork; an
understood test failure is fixed directly; a disproved assumption changes the
approach; an unknown recurring failure gets no helper chain; the final exact
subject receives fresh author-distinct judgment and the report states any gap.

The grandfathered developer-only Python reference is an optional fixed-dispatch
adapter with explicit repair rounds, not the native charter's execution engine.
Its narrower [adapter contract](references/bounded-adapter.md) and tests remain
available without adding a runtime, command, scheduler or mandatory worksheet.