菜单
Survey a corpus, classify, and dispatch outputs. Three modes: retro (transcript review → issues), trend (longitudinal performance analysis), sweep (GitHub issue triage → fix-epics). Delegates execution to pauli (retro/trend) or jr (sweep) to keep main context clean.
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/nicsuzor/aops --skill survey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。
正在显示 SKILL.md
基于 SOC 职业分类
| name | survey |
| type | skill |
| category | instruction |
| description | Survey a corpus, classify, and dispatch outputs. Three modes: retro (transcript review → issues), trend (longitudinal performance analysis), sweep (GitHub issue triage → fix-epics). Delegates execution to pauli (retro/trend) or jr (sweep) to keep main context clean. |
| triggers | ["survey","retro","transcript review","session review","trend review","performance trends","issue sweep","triage issues"] |
| modifies_files | true |
| needs_task | false |
| mode | orchestration |
| domain | ["framework","quality-assurance","operations"] |
| allowed-tools | Agent, Bash, Read, Grep, Glob, Edit, Write, Skill, AskUserQuestion, mcp__pkb__list_tasks, mcp__pkb__get_task, mcp__pkb__create_task, mcp__pkb__update_task, mcp__pkb__append, mcp__pkb__task_search |
| owner | junior |
| version | 1.0.0 |
| tags | ["retro","trend","sweep","quality","consolidation"] |
Survey a corpus, classify findings, and dispatch outputs according to the selected mode.
| Mode | Corpus | Primary output |
|---|---|---|
retro | Session transcripts (one at a time) | GitHub issues filed via gh |
trend | Many sessions / audit files | Trend report + recommendations |
sweep | Open GitHub issues | PKB tasks, fix-epics, closures |
Privacy Rule: Anonymize all findings. Do not expose real names, emails, student details, or raw session dumps.
This skill delegates execution to keep the main context clean:
retro / trend mode: Dispatch pauli with access to PKB and system tools.sweep mode: Dispatch jr to handle interactive triage and confirmation gates.Perform a critical, forensic review of a single session transcript, apply immediate fixes where appropriate, and file the tracking GitHub issues.
pauli dispatched within the session) is explicitly allowed and structurally sound because the subagent executes in a clean, detached context.$AOPS_SESSIONS is set and $AOPS_SESSIONS/transcripts exists. If not, stop and ask the user.$AOPS_SESSIONS/transcripts/YYYY-MM/*-${SID}-*-claude-full.md. Use -abridged.md only as a fallback.$AOPS_SESSIONS/transcripts/YYYY-MM/-full.md file..jsonl — a forensic review on a degraded transcript yields false findings; proceed on raw JSONL only with explicit user confirmation./craft defects: compliance framing, missing artifact chain, etc.)./learn that last task should have been xyz: When Nic invokes this style with a description of what should have happened, treat it as a directive to perform a dual action:
When the transcript shows an artifact whose premise a sharp principal would have bounced — "was this a good idea?" answered no; good, working, well-tested work done for a bad idea (canonical instance: a deterministic rig — regex/threshold/NLP/checklist — built for a call a smart agent should just make, judgment-non-delegable) — that nonetheless passed review, classify it as a bad-premise approval and score the miss against the reviewer who approved it, not only the author:
/verify / rbg / pauli). Each of those carries a forced step-0 Premise Test (canonical definition: [[premise-test.md]]); an approval means that forcing function was skipped or rationalised past — a reviewer failure, with test-passing as its expected surface, never an excuse.This makes the reviewer's miss visible and attributed — the compounding fix for floor-optimisation (#1585): a slipped-through bad premise becomes a logged, attributed miss instead of an invisible one, so the cost lands on the surface that should have caught it.
Produce a review in this exact format. Keep text concise:
## Transcript Review: <filename>
**Session**: <session_id> **Date**: <date> **Project**: <project>
**Verdict**: [EXCELLENT | GOOD | ADEQUATE | POOR | FAILING]
### Findings
[Concise description of failures/successes, grouped by structural cause.]
### Patterns (Optional)
[Upstream/structural root cause spanning multiple findings.]
gh issue list and gh pr list to avoid duplication.gh issue edit.Bug: <brief-slug>.reviewed_by block.uv run pytest) to verify no regressions were introduced.| Anti-pattern | What to do instead |
|---|---|
| Same-context self-grading (the same agent reviewing their own work within the same conversation/turn context without a fresh reviewer boundary) | Review by a fresh subagent (like pauli dispatched within the same session) or a separate reviewer, ensuring a detached/clean review context. Same-session review by a fresh subagent is structurally sound and explicitly allowed. |
Including remediation proposals in the report (recusal) | Stop at facts, structural context, and impact. Propose fixes directly in the codebase (if permitted) but keep the filed issue strictly forensic. |
| Citing a single session as justification for a new mechanism | Recurrence is the evidence base for framework change, not the salience of a single transcript. |
Review multiple sessions to identify systemic effectiveness and trends.
Analyze aggregate true/false rates, temporal trends, coverage, and cost-benefit. Produce a report in this format:
# Trend Review: <Component Name>
**Question**: <review question> **Date**: <today>
**Corpus**: <N> files, <date range> **Sample**: <N> files (criteria: ...)
## Executive Summary
[3-5 sentences summarizing if the system is working and the main trend/issues.]
## Objectives Verdict
| Objective | Verdict | Evidence |
| --------- | -------------------------------------------- | -------- |
| [obj 1] | ANSWERED / PARTIALLY ANSWERED / UNANSWERABLE | [why] |
## Individual Assessments
### <filename> (<date>)
- **Context**: ... **Component behavior**: ... **Accuracy**: [TP/FP/FN] **Impact**: ...
## Aggregate Analysis
[Signal Quality / Temporal Trends / Coverage Map / Cost-Benefit]
## Recommendations
[Actionable, prioritized recommendations with evidence citations.]
## Confidence and Limitations
Save the report to ~/.aops/sessions/reviews/<component>-trend-<date>.md.
Triage and process open issues on nicsuzor/academicOps (batch limit: ≤20 issues).
Fetch open issues, focusing on the focal issue first if a directed focus is provided. Classify using this rubric:
| Disposition | Criterion | Action | Label |
|---|---|---|---|
close-as-stale | >90d old, no activity, or describes retired behavior. | Close issue with explanatory comment. | triaged-stale |
consolidate-duplicate | Describes the same underlying problem. | Merge details into canonical issue; close duplicate with cross-link comment (#N). | triaged-duplicate |
aggregate | Multiple issues sharing one root cause/fix surface. | Fold into one fix-epic task (leave queued); close source issues with pointer comments. | triaged-aggregate |
evidence-bump | Accumulates evidence for a related open issue/epic. | Leave open; add comment citing canonical issue (#N). | triaged-evidence-bump |
single-task | Atomic task (AC clear, ≤3 files, obvious fix). | File polecat task with Closes #N. | triaged-single |
fix-epic | Multi-step, multi-file, or design-required work. | Create epic task and decompose, leave queued. | triaged-epic |
defer | Real but blocked or low-criticality. | Apply defer label and revisit-by date. | triaged-defer |
For proposals that add or escalate a rule, perform this review before assigning the disposition:
specs/ENFORCEMENT-MAP.md rows as a pipeline gap.Log results in the following format:
## Cycle <N> — applied (open before: <K>; batch: <M>)
### Applied autonomously (done)
- Consolidate-duplicate: #R → unique detail merged into #S, #R verified state:closed
- Aggregate: #A, #B, #C → folded into fix-epic <id> (queued)
- Evidence bump: #T → bumped #U
- Close (stale): #P
- Defer: #Q (revisit-by YYYY-MM-DD)
- Single-task: #X → "<title>"
- Fix-epic (queued): <title> ← #D, #E
### Needs human triage / decision (waiting)
- #Z (rubric ambiguous: <reason>)
- <add-or-escalate proposal>: cost-ladder reasoning + ≥3 recurrence links
severity (or set severity=0) on tasks created during the sweep. Assigning non-zero severity to ordinary tasks is prohibited; severity belongs exclusively on target milestones (see [[../remember/references/TAXONOMY.md#severity-target-boundary]]).priority=0 (P0) on swept tasks unless it is deliberately calibrated under canonical rules (see [[../remember/references/TAXONOMY.md#p0-calibration-bar]]) and explicitly requested/justified.state: closed.epic-a0523a25 and append the cycle log details.Skill(skill="verify", args="Verify cycle <N> of /issue-sweep on epic-a0523a25.")Core academicOps skill — institutional memory, strategic coordination, workflow routing, and framework governance. Merges butler (chief-of-staff) with framework development conventions.
Daily note lifecycle — compose and maintain a factual daily note. Reports the state of the day; does not prioritise or recommend. SSoT for daily note structure.
Emergency session bail — fast resume task + short handover, no commit/PR/reflection. For when you (or the user) need a clean context now. Use /end-session for canonical close.