| name | aops |
| category | instruction |
| description | Core academicOps skill — institutional memory, strategic coordination, workflow routing, and framework governance. Merges butler (chief-of-staff) with framework development conventions. |
academicOps Core Skill
Coordinate framework operations, maintain institutional memory in the PKB, route workflows, and enforce framework conventions.
Institutional Memory via PKB
The Personal Knowledge Base (PKB) is the single source of truth.
- On first invocation, load the framework state document:
mcp__plugin_aops-core_pkb__get_document(id="aops-state").
- Update this document with visions, decisions, open questions, and roadmaps. Use
mcp__plugin_aops-core_pkb__append for incremental updates.
- Use
list_tasks with project=<project-id> to find tasks scoped to a single project. Do not infer project membership from IDs.
Framework Governance
Instruction Maintenance
If corrected, encountering a process gap, or seeing a learning-tagged task:
- Locate the responsible instruction file (
CORE.md or SKILL.md).
- Update the file in the same turn.
- Validate that agents follow the updated instructions.
Enforced-behaviour carve-out: if the instruction is already enforced via a gate, hook, or ENFORCEMENT-MAP row — or if satisfying it would add, modify, or retire any such mechanism — do NOT update in place. Route to the Workflow Router "Change enforced behaviour" row and follow specs/enforcement/enforcement.md §4 instead.
Dogfooding & Process Verification
When failures occur, identify and fix the process gap rather than just applying a local patch. Do not declare victory until you verify success with evidence.
Supervising Delegated Work — invoke the supervisor skill (mandatory)
When you are delegating work to other agents and verifying it gets done — including running as
the main conversation orchestrator with background workers — invoke the supervisor skill;
never hand-roll supervision inline. The mandate, the "conversational orchestrator is not an
exemption" carve-out, and the full proof discipline are canonical at
[[../supervisor/SKILL.md#when-to-invoke-mandatory]].
Handover
- Ensure the next strategic step is filed as a PKB task before exiting.
- If a design decision produces implementation work affecting multiple files/projects, run
/planner decompose first.
Verification of Assumptions
Verify capability and constraints of framework mechanisms (plugins, model context, environment) against primary codebase files before making claims or routing tasks.
Strategic Guidance
- Prioritize features serving actual academic workflows.
- Recommend automation in order of maturity: Manual -> Assisted -> Supervised (Default) -> Autonomous. Move to autonomous only after multiple successful supervised runs.
Workflow Router
Route tasks based on scope:
Categorical Conventions
File Boundaries
$AOPS/*: Modification permitted.$ACA_DATA/*: Direct file operations forbidden; delegate writes/updates to PKB MCP tools.
Core Conventions
- One Spec Per Feature: Specifications are timeless.
- Single Source of Truth: Keep data in exactly one location.
- Self-Documenting (P#10): Embed documentation in code; avoid parallel docs.
- Always Dogfooding (P#22): Use real projects as development guides.
- Skills are Read-Only (P#23): Skills must not store state. State lives in
$ACA_DATA.
- Trust Version Control (P#24): No backup files (
.bak, _new). Edit files directly, commit, and push.
- Plan-First Development (P#41): Code only after plan approval, except for trivial typos/formatting.
- Just-in-Time Context (P#43): Automatically surface context at decision points.
- Memory Model (P#46): Semantic notes (synthesized) must be understandable without reading sources. Episodic notes are immutable after creation.
- Agents Execute Workflows (P#47): Workflow procedures belong in workflow files, not agent prompts.
- No Shitty NLP (P#49): Use LLM judgment for semantic decisions. Do not write programmatic scripts wrapping LLM APIs; use the agentic platforms directly.
Task Lifecycle & Execution
Follow the Plan -> Act -> Validate cycle:
- Plan: Formulate approach. Invoke critic review (
rbg or pauli) only if uncertainty or blast radius requires it.
- Act: Execute changes.
- Validate: Test locally, commit, push, and release task.
Terminal Move after Dispatch
When a coordinator executes a fire-and-forget dispatch with no follow-up or babysit instruction:
- The default terminal move is to close the session promptly (via
/dump or /end_session) rather than idling.
- Do not promise or wait for a completion notification for a non-notifying dispatch form. If the session has nothing to watch and no babysit instruction, close it immediately.
- Do not create a tracking task per dispatch (dispatch operations should not spawn overhead tracking tasks).
HALT Protocol
When unable to derive a decision:
- STOP: Halt execution.
- STATE: Record what cannot be determined and why.
- ASK: Prompt the user using
AskUserQuestion.
- DOCUMENT: Record the resolved rule.
Rules & Anti-Patterns
Anti-Pattern: Asking Permission for Safe Actions
Do not ask permission to file tasks, fix bugs, or execute safe operations (e.g., retitling, graph hygiene, canceling superseded tasks). Perform the action and report the outcome. Only ask permission for destructive or externally visible actions.
Anti-Pattern: Recording Actions Instead of Executing
Perform safe administrative actions (e.g. repointing dependencies, status flips, hygiene) immediately in the same turn instead of documenting them as future TODOs.
Reporting Norms
Report actions and states in plain English. Avoid using internal taxonomy (DECIDE-class, Externalisation Heuristic, P#) in user-facing output.
- Wrong: "DECIDE-class supersession recorded: cancel task-A."
- Right: "Cancelled task-A (superseded by task-B)."
- Wrong: "Applied DEFER-class treatment to decision."
- Right: "Decision deferred pending run data."
