dissertation-agent-architecture-audit

name	dissertation-agent-architecture-audit
description	Audit the dissertation agent's rule stack, memory layers, skill routing, tool discipline, window separation, persistence, and output gates for conflicts, stale context, hidden assumptions, or maintenance risk.

Dissertation Agent Architecture Audit

Use this skill for deeper maintenance audits of the dissertation agent system, especially when rules conflict, behavior degrades, the user reports unreliable outputs, or new skills/workflows are added.

Purpose

Adapt ECC agent-architecture audit to a local dissertation workflow. This is not a software production audit. It checks whether the current rules, skills, memory files, and window roles are coherent and source-grounded.

Architecture Layers To Audit

Layer	Dissertation Version	Common Failure
1	System/developer instructions	conflicts with project rules or window role
2	Current chat history	old task dominates current request
3	AGENTS.md	skill list or workflow outdated
4	PROJECT_AGENT_PREFERENCES.md	preferences duplicate or conflict with skills
5	.agents/skills/	overlapping triggers, missing frontmatter, stale rules
6	research-wiki/TASK_STATE.md	stale next action or missing handoff
7	knowledge-base / Obsidian	duplicate facts, broken links, unclear source of record
8	tool surface	browser, Word renderer, LMS, Obsidian, connectors assumed but not verified
9	output style rules	final answer shape conflicts with task need
10	persistence	local files, memory notes, and generated documents disagree

Audit Workflow

1. Define the symptom or audit scope.
2. Read only relevant files first:
   • AGENTS.md
   • PROJECT_AGENT_PREFERENCES.md
   • research-wiki/TASK_STATE.md
   • research-wiki/WINDOW_WORKFLOW_PROMPTS.md
   • relevant SKILL.md files
3. Map findings to the layer above.
4. Rank findings by severity:
   • critical: can produce wrong formal or ethical output
   • high: likely to cause repeated workflow failure
   • medium: creates confusion or duplicated effort
   • low: minor maintainability issue
5. Recommend minimal fixes.
6. Update TASK_STATE.md after confirmed rule changes.

Findings Format

Finding:
- Severity:
- Layer:
- Evidence:
- Risk:
- Fix:
- Affects Production Window: yes/no

Default Fix Order

1. Remove or clarify conflicting rules.
2. Restore source-first and document-quality gates.
3. Clarify window role separation.
4. Update skill routing in agent-orchestration.
5. Update shared memory and task-state handoff.
6. Record residual risk.

Guardrails

• Do not blame the model before checking wrapper/rule/memory layers.
• Do not rewrite many skills at once unless the defect is systemic.
• Do not edit dissertation content files unless the audit specifically concerns them and the user authorizes it.
• Do not treat a clean current state as proof that an earlier failure never happened.