mi-orchestration

MI orchestration

The basic task is read-only audit or resumption planning. The attached flowchart is design input, not operational results or executable code. This skill does not provide the original st-pipeline.py, vendor collector, Slack integration, or /mi-decide. Re-confirm example numbers, dates, and call volumes against the account and run, and do not overwrite decisions with a data rebuild.

When to use this skill

Design and audit the raw → core → mart → health → signals → render/inbox lineage of the weekly MI report.
Find the latest evidence for an interrupted run and the smallest resumable boundary.
Distinguish weekly from backfill, missing from zero, estimates from actuals, and candidates from human decisions.
Review coverage/flip hold, quarantine, right-censoring, report slots, and human-only ledgers.

Simple analysis is handled by data-analysis, general monitoring design by monitoring-observability, barretlee/agent-pulse operations by agent-pulse. Leave new persistent knowledge stores to llm-wikiand keep only the linkage contract with MI evidence in this skill.

Instructions

1. Lock one batch of work

design, audit-run, resume-plan, decision-review . Record the exact checkout/root, run and data period, mode, original code revision, config, retained raw, desired output, external side effects and budget, and approver. Do not guess at mixed dates, numeric contradictions, or unverified schema/permissions; leave them as separate gaps. attached analysis, but verify project files and commands against the actual checkout.

2. Separate collection authorization first

vendor boundary. The existence of an official API, account entitlement, the user's approval to run this particular iteration, and rights to use and share results are separate. Do not add Sensor Tower Call Credits and request counts in the same unit. The current Sensor Tower terms prohibit automated access via bots/crawlers/browser extensions. The old Playliner terms are historical material only; check the current contract and approved interfaces. Before entering Service Data into AI, confirm Permitted AI Tool's written data processing agreement, administrative controls, and non-consumer/non-free/non-shared environment. Training, fine-tuning, and distillation are prohibited. API=0also does not exempt web UI automation, internal endpoints, seats, or data/AI usage rights. Unverified authorization means blocked; do not automatically fall back to another account/vendor/web session.

3. Bind the actual data lineage

data flow contract. For each output, link run/period, platform, entity namespace, country, unit, source interval and observation time, input bytes hash, code/config/schema version, and transformations. Record the design DAG and the execution receipt separately . Retain raw per run within the period and environment permitted by contract. Do not circumvent expiry/deletion obligations through immutability, and do not claim reproducibility from .prevor mtime alone. core is derived data regenerated from the source, and the human ledger is a separate append-only store. Do not turn missing values into zero or fill numbers with an LLM. Mark estimated metrics as estimate.

4. Uphold integrity, backfill, and decision gates

Read the coverage denominator and the previous-week flip definition and threshold from config and measurement evidence. Do not use examples such as the attached 80% or 10 as common criteria. On failure, hold, or quarantine, prohibit normal signals/render and leave only an inbox containing partial collection, cause, and next actions. Backfill may exclude only specific weekly-window checks. Data validity, permissions, lineage, and empty inputs are not ignored. Only when observation of a defined follow-up outcome is truncated mark it as right-censored with candidate and reason. Partial-month aggregation or ordinary missingness is incomplete/missing + reason, and a completed historical backfill is not turned into censored. Alert retries are at most once within the specified round budget, and authentication, prohibition, and quota exhaustion are not retried. operational and human gatesapply. Candidate proposal, peer review, readiness, and the actual accept/kill/hold/closed records are separate.

5. Create an audit and resume plan with read-only tools

local audit contractand inspect the pre-existing manifest. After execution, do not create a manual manifest or touch files to report that a past run was verified. The bundle format is not an adapter that directly reads the original MI files. Even when separately transforming actual project receipts, record the source and limitations of that transformation.

# Installed skill directory 기준. 이 예시는 합성 fixture이며 벤더 호출이 없다.
node scripts/mi-audit.mjs check --root assets/example --manifest manifest.json
node scripts/mi-audit.mjs plan --root assets/example --manifest manifest.json
node --test scripts/mi-audit.test.mjs

planis not executed. Passing hashes, input fingerprints, periods, and structure does not certify statistical validity, truthfulness of actual measurements, content/browser QA, collection permission, or human approval. Report downstream as stale from the boundary where source/config changed, and block resumption without the latest parent. Actual deterministic rebuild or paid re-collection is done only after separate approval and confirmation of the actual project command.

6. Hand off including the verification scope

Return one result as follows:

mode, run/period, root/revision, and the scope of evidence read;
per-step verified / stale / hold / fail / blockedand the first broken lineage;
evidence links from source→metric→signal→slot→human judgment, quarantine/censoring;
usage with API requests and credits/cost separated, or not-executed / unknown;
verification command and actual exit code, artifact integrity and the distinction between analysis/content/permission checks;
the smallest resume boundary, approval/measurement/file gaps, and explicit next actions.

Do not promote execution success, schedule active, JSON pass, or agent review that is not by a human into a human decision.

Examples

Weekly hold: coverage is below the configured threshold. Preserve valid partial results and quarantine, and skip signals/render. Leave the observation scope, cause, and next actions in the inbox, and obtain approval for additional collection.

Resume of a stale report: raw or config bytes have changed. Ignore the latest mtime of the previous HTML, find resume candidates starting from the first input fingerprint mismatch, and do not execute before parent verification.

Backfill candidate review: The one-month aggregation is incomplete. Block empty data and leave the observable denominator, period, and incomplete/missing reason. Only mark as right-censored when the defined follow-up outcome is unobserved at the observation end time. candidateDo not auto-promote based on human ledger accept.

Best practices

Preserve the immutable snapshot, versions of data/code/config, and calculation definitions together.
bluf, wave, insight Slots undergo a structural check, then the actual render and figures/descriptions are reviewed separately.
Human ledger, product execution, and external notifications are not tied to the automatic completion conditions of collection/analysis.
A synthetic fixture PASS is not evidence of actual vendor collection, market performance, or operational reproducibility.

About this skill