AgentFS Violation Check

Full behavioral self-audit of session conduct against the agent’s rules (AGENTS.md guardrails) and principles (SOUL.md). Unlike agentfs-ctx-chk which audits the context files for structural health, this skill audits the agent’s behavior during the current session for compliance violations.

Use this when you suspect the agent has drifted, after context compaction, or as a periodic integrity check. The lightweight version of this check runs automatically at Session Canary re-verification points (turn 5, then every 10 turns) — this skill is the full, explicit version.

Prerequisites

• AGENTS.md is loaded in context (contains rules to audit against)
• .agents/SOUL.md is accessible (contains principles to audit against)
• Session has ≥2 turns of history to audit

How It Differs from Related Skills

Skill	Question
agentfs-violation-chk (this)	“Did I violate my own rules this session?”
agentfs-ctx-chk	“Are the context files well-engineered?”
agentfs-eval	“Is this workspace structurally healthy?”

Specification

ID	Capability	Verifiable By
S1	Audit all AGENTS.md rules against session behavior	Report covers every numbered rule
S2	Audit all SOUL.md principles against session behavior	Report covers every named principle
S3	Identify specific turns where violations occurred	Each finding cites the turn or action
S4	Distinguish severity levels	Findings classified as 🔴 violation, 🟡 borderline, 🟢 clean
S5	Produce actionable remediation	Each violation includes what should have been done

Steps

Step 1 — Gather Audit Baseline

Read the rules and principles that govern behavior:

1. Review the Rules section of AGENTS.md currently in context. List every numbered rule (e.g., Signal Dispatch, Pre-Flight, Post-Write, Session Canary, Conflict Resolution, Checkpoint, Scope Rules).
2. Review the Principles section of .agents/SOUL.md currently in context. List every named principle (e.g., No assumed inputs, Direct communication, Intellectual integrity, etc.).

If either document is not in context, read it before proceeding.

Step 2 — Review Session History

Walk through the conversation history turn by turn. For each agent turn, check:

Rule compliance:

• Did any hey message get treated as a greeting? (Signal Dispatch)
• Were multi-step tasks (≥3 tool calls or ≥2 files) preceded by a stated plan? (Pre-Flight)
• Were writes to .agents/ or ~/.agents/ followed by post-write hooks? (Post-Write)
• Was the canary emitted on turn 1? Re-verified at checkpoints? (Session Canary)
• Were any positions reversed without stating what changed? (Conflict Resolution)
• Were destructive ops on .agents/ preceded by checkpoint? (Checkpoint)
• Were memories written to correct scope? Skills to correct scope? (Scope Rules) Principle compliance:
• Did the agent proceed without asking when information was missing? (No assumed inputs)
• Did any response open with filler/validation phrases? (Direct communication)
• Was a position reversed due to social pressure rather than new evidence? (Intellectual integrity)
• Were complex solutions chosen over simpler alternatives? (Reliability over cleverness)
• Were symptoms patched without tracing root cause? (Root cause, not symptom)
• Were risks left unnamed? (Proactive risk naming)
• Did the agent make aesthetic/domain calls outside scope? (Scope discipline)
• Were structured files treated casually? (Data integrity)

Step 3 — Classify Findings

For each issue found:

Severity	Criteria
🔴 Violation	Clear breach of a mandatory rule or principle
🟡 Borderline	Technically compliant but spirit was not met, or edge case
🟢 Clean	No issues found for this rule/principle

Step 4 — Produce Report

Present findings in this format:

## Violation Check Report

**Session:** [canary name] | **Turns audited:** [N] | **Timestamp:** [ISO 8601]

### Rules

| # | Rule | Status | Detail |
|:-:|------|:------:|--------|
| 1 | Signal Dispatch | 🟢 | — |
| 2 | Pre-Flight | 🔴 | Turn 3: edited 3 files without stating plan |
| ... | ... | ... | ... |

### Principles

| # | Principle | Status | Detail |
|:-:|-----------|:------:|--------|
| 1 | No assumed inputs | 🟢 | — |
| 2 | Direct communication | 🟡 | Turn 1: opened with "Sure!" |
| ... | ... | ... | ... |

### Summary

- **Violations (🔴):** N
- **Borderline (🟡):** N
- **Clean (🟢):** N

### Remediation

1. [What should have been done differently, for each 🔴/🟡]

Step 5 — Lightweight Mode (Canary Integration)

When invoked as part of the Session Canary re-verification (not the full skill), perform a compressed version:

1. Silently review behavior since last canary check
2. If no violations found → do not report anything (zero noise)
3. If violation found → report inline with the canary re-verification, e.g.:

   🐦 Canary re-verified: walrus-tungsten ⚠️ Self-check: [brief description of violation]

This step is documented here for reference but is executed by the Session Canary rule itself, not by loading this skill.

Gotchas

• Post-compaction blindness: After context compaction, earlier turns may be summarized or lost. The audit can only cover what remains in context. Acknowledge this limitation in the report rather than claiming “no violations” with false confidence.
• Self-audit bias: The agent auditing its own behavior has an inherent bias toward leniency. Mitigate by checking each rule mechanically (did the specific action occur? yes/no) rather than holistically.

Verification

• Report covers every rule in AGENTS.md
• Report covers every principle in SOUL.md
• Each finding cites a specific turn or action
• Severity classifications are justified
• Post-compaction limitation is acknowledged when applicable

Tests

Test	Spec	How to Verify	Expected Result
T1	S1	Run after a session with known rule violation	Violation correctly identified and cited
T2	S2	Run after a session with filler phrase used	Principle violation flagged as 🔴 or 🟡
T3	S3	Check that findings reference specific turns	Each 🔴/🟡 cites turn number or action
T4	S4	Run on a clean session	All items show 🟢, no false positives
T5	S5	Trigger via canary re-verification with no issues	No output (silent pass)

Changelog

See CHANGELOG.md for version history.