---
name: reflective-skill-updater
description: "Review real task executions, errors, corrections and successful alternatives. Apply global learning: identify the canonical owner, propose narrow improvements and revalidate consumers after approval. Use after a completed or failed OLLIE run, including fresh-window replay failures. Route tenant facts separately; delegate approved skill creation/editing to Skill Creator."
---

# Reflective Skill Updater

**Version:** 2.2.0-rc.1. Upgraded from the supplied original skill and approved learning design.

## Review what actually happened

Use this same working chat while evidence is available. Read original instructions and the current
canonical SKILL.md plus affected references; do not judge by its name alone. Prefer available run
transcripts/tool receipts; otherwise use the user's summary with an explicit evidence limitation.
Do not invent transcript tools, file access or successful actions. Review failed runs too, but never
turn an incomplete result into a golden successful fixture.

Map objective -> intended procedure -> actual calls -> retries/errors -> user corrections -> final
result/delivery -> verification. Classify useful events as Worked, Friction, Failure or Discovery.
Preserve what already works. Promote only actionable, evidence-supported, recurrent knowledge.
A single incident can justify a narrow safety regression; it does not justify universal restrictions.

## One canonical owner per reusable rule

Use these primary destinations, not a sibling-copy rule:

| Class | Owner |
|---|---|
| OLLIE_RUNTIME | authority, access, email, bootstrap/execution/recovery owner |
| GLOBAL_BUSINESS_SKILL | reusable business outcome, formulas, validation |
| GLOBAL_PLATFORM_SKILL | reusable platform extraction/navigation/recovery |
| WORKSPACE_CONTEXT | live client definitions, rates, filters, targets, recipients, approvals |
| TENANT_BINDING | immutable resource/account/worker/profile/host-location binding |
| PRODUCT_CAPABILITY_REQUEST | missing host/Outloop primitive or managed contract change |
| LEARNING_RECEIPT_ONLY | non-reusable or insufficiently supported event |

Inspect the current authority, access, email and bootstrapper owners when relevant; do not copy
their doctrine here. The apparent failing report may be only a consumer of a platform defect.
Generated tenant Dispatcher/Operator files are evidence/consumers, not normal sources for global
learning. Search the catalog by outcome/platform before proposing a new skill.

Tenant-only findings return `NO_GLOBAL_SKILL_EDIT`; use the approved Context/binding update path.
Outloop-managed instructions are not patched manually. A runtime/product change returns
`PRODUCT_OR_RUNTIME_CHANGE_REQUIRED`; ordinary skill-update approval is not an implicit authorization
to alter the runtime contract. An explicit scoped runtime-change request may authorize a staged
canonical implementation, but never overrides managed policy or live deployment gates.

## Improvement Report before editing

Report: execution evidence and limits; what worked/failed; successful path; candidate learning;
category; canonical file/section; smallest proposed edit; why this owner; affected consumers;
regression fixture; tenant data removed; proof level; delivery/install impact; rollback; not promoted.
Resolve one owner first, then inspect actual dependencies/siblings for revalidation or reference
updates. Never infer family ownership from tenant suffixes or copy the same rule across clients.

Use `VERIFIED_IN_REAL_WORKFLOW`, `OBSERVED`, `ASSUMED`, `UNKNOWN`, `NOT_PROVEN` accurately.
Structural checks and simulated choices remain static evidence, not real unattended execution.
Read the FDE runbook when available; if missing, state that and use this durable contract instead
of inventing its contents. Do not stop useful analysis merely because optional docs are absent.

## Approved implementation

Wait for the specific approval or an existing standing approval that covers the exact change.
After it, use installed Skill Creator for structural edit, validation and packaging. Preserve all
supporting files. Stage in a working copy; never quietly overwrite another editor's newer version.
Validate business fixtures, negatives, discovery, dependencies, policy non-bypass and leak safety.
Package a complete skill in the host-supported format. Do not promise a Save button without an
actual tool-produced installation card. Version/hash, install/read-back and rollback are separate
from creating files. Do not distribute rollback packages as if they were the current install card.

## Close the learning loop

Install the canonical version once and refresh only affected consumers via their approved generator.
Read back installed bytes. Ask this chat for the shortest new-chat business prompt with only variable
inputs. Run it in a fresh, correctly connected workspace context. Verify equivalent method, source
coverage, calculations and output quality, not identical dynamic numbers. Do not make replay work
by pasting the old conversation or a hidden implementation manual into the prompt.

If replay fails, classify missing method vs missing client fact vs unavailable host permission and
fix the correct owner. Report readiness only for tested scope. Never globally leak names, IDs,
credentials, full transcripts or raw client fixtures. No automatic reflection schedule is enabled.