菜单
Use when checking harness health, setting up observability cadences, understanding snapshot formats, configuring telemetry export, or verifying that the harness's own observability is working — covers all four layers of harness observability
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Habitat-Thinking/ai-literacy-superpowers --skill harness-observability命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use when acting as the adversarial spec reviewer — raises steel-manned objections across six categories before plan approval, requires evidence per objection, and discloses what was not challenged
Use when designing, reviewing, or classifying an agent that guards the human — defines the sentinel category, the three-part signature (S1 read-only, S2 advisory-to-human, S3 explicit honesty rule), the near-miss gallery that explains why read-only-plus-advisory is not sufficient, the honesty-rule-before-detection-logic discipline, and the three anti-patterns that eject an agent from the category
This skill should be used when the user asks to "assess AI literacy", "run an assessment", "check literacy level", "evaluate our AI collaboration", "where are we on the framework", or wants to determine their team's AI literacy level using the ALCI instrument.
基于 SOC 职业分类
正在显示 SKILL.md
| name | harness-observability |
| description | Use when checking harness health, setting up observability cadences, understanding snapshot formats, configuring telemetry export, or verifying that the harness's own observability is working — covers all four layers of harness observability |
Harness observability is the measurement layer that makes context engineering, architectural constraints, and garbage collection evidence-based rather than intuition-based. Without it, a harness is infrastructure you hope works. With it, you know.
This skill covers four layers of observability, each answering a different question at a different timescale.
This skill does not cover CI workflow configuration (middle loop), individual constraint design, or garbage collection rule authoring — those belong to their respective skills. It covers how to observe whether those mechanisms are operating effectively.
For the snapshot schema and field definitions, consult
references/snapshot-format.md. The snapshot includes a Diaboli panel
(after Session Quality, before Operational Cadence) that surfaces
advocatus-diaboli activity as descriptive stats across all objection records.
| Layer | Question | Timescale | Consumer |
|---|---|---|---|
| Operational Cadence | "Is the harness running?" | Session / weekly / quarterly | Developer in session |
| Trend Visibility | "How has the harness changed?" | Weekly / quarterly | Team over time |
| Telemetry Export | "Can I visualise this externally?" | Continuous | External dashboards |
| Meta-Observability | "Is the observability itself working?" | Quarterly | Agents + team |
The harness has mechanisms — audit, GC, reflection, mutation testing — but mechanisms only work if they run. Operational cadence ensures they actually fire on schedule.
The primary tool is /harness-health. Running it:
Recommended cadences:
| Mechanism | Cadence | Command |
|---|---|---|
| Health snapshot | Monthly | /harness-health |
| Full audit | Quarterly | /harness-audit |
| Assessment | Quarterly | /assess |
| Reflection review | Quarterly | Review REFLECTION_LOG.md |
| Mutation trend check | Monthly | Download CI artifacts |
A Stop hook nudges the developer when the last snapshot is older than 30 days. This turns a declared cadence into an actual practice.
Snapshots are point-in-time. Trends are the delta between snapshots. No external tooling required — trends are computed by diffing markdown files.
When /harness-health runs and a previous snapshot exists, it:
For quarterly reviews, /harness-health --trends reads all snapshots
and produces a multi-period trend view.
What agents do with trends: The orchestrator reads the latest
snapshot's Trends section before planning work. Declining mutation
scores might trigger test quality prioritisation. A stalled learning
flow might prompt a /reflect nudge.
For teams that want Grafana, Langfuse, or other external dashboards, the snapshot data can be exported as OpenTelemetry metrics.
This is optional. The file-based approach (Layers 1-2) works without any external tooling. Telemetry export adds continuous visibility for teams operating at scale.
For OTel metric names, semantic conventions, and a reference export
script, consult references/telemetry-export.md.
The self-referential layer: the harness checking whether its own observability is working.
"A harness that cannot verify its own operation is a harness you hope works."
The harness-auditor agent runs five meta-checks:
| Check | Signal |
|---|---|
| Snapshot currency | Stale = outer loop not running |
| Cadence compliance | Overdue = practice not followed |
| Learning flow | Stalled = compound learning broken |
| GC effectiveness | Silent = rules may be misconfigured |
| Trend direction | Unacknowledged decline = drift unnoticed |
For thresholds and detailed check definitions, consult
references/meta-observability-checks.md.
| Situation | Action |
|---|---|
| Starting a new quarter | Run /harness-health to baseline |
| After a major feature lands | Run /harness-health to check impact |
| Harness feels stale | Run /harness-health --deep for authoritative check |
| Quarterly review | Run /harness-health --trends for multi-period view |
| Setting up a new project | Configure cadences in CLAUDE.md |
| Want external dashboards | Follow references/telemetry-export.md |
/harness-health maintains two README elements:
Health badge — shields.io, colour-coded:
| Status | Condition | Colour |
|---|---|---|
| Healthy | All layers operating, no overdue cadences | Green |
| Attention | One layer degraded or outer loop overdue | Amber |
| Degraded | Multiple layers degraded or no snapshot in 30+ days | Red |
Health icon — a custom SVG linked to the latest snapshot.
The badge colour factors in meta-health. Stale snapshots or overdue cadence trigger amber even if enforcement ratios look fine.
