Harsh Critic

space: outcome type: document links: [[@action/skills/SKILL-INDEX]]]

Harsh Critic — Completion Bias Defense

Purpose

Before Drewgent declares "done", Harsh Critic asks: "If the user saw this right now, what would they say?"

This defends against the 5 structural LLM completion biases:

1. Completion Bias — "let's call it done"
2. Error Loop Escape — delete the feature instead of fixing it
3. QA Skipping — "build passed" = "verified"
4. Silent Standard Lowering — "good enough for now"
5. Delegation to User — "please check this manually"

---

3-Tier Classification

EXTREME — HARD Block (exit 2)

These patterns mean the task is NOT done. Block immediately.

| Pattern | Detection | Example | |---------|-----------|---------| | Exception Violation | User said "don't X" but X happened | "I said don't push and you pushed" | | QA Without Evidence | "PASS" claimed but no .qa-evidence.json | "QA passed" without any test run | | Same Mistake Repeat | Same issue from last session detected again | "This is the 3rd time you..." | Fake Completion | "Done" declared but build/test still failing | "Completed!" with unfixed errors |

HIGH — FAIL + Mandatory Fix

These patterns need correction before proceeding.

| Pattern | Detection | Example | |---------|-----------|---------| | Delegation to User | "please check", "verify yourself", "確認してください" | Shifting verification burden | | Scope Expansion | One-time instruction applied globally | User said once → permanent rule created | | Unverified Claim | Statement without supporting evidence | "This is the pattern" with no diff/URL | | Design Regression | Visual quality dropped without acknowledgment | "It works" but design is broken |

MEDIUM — Warning

These are suboptimal but not blocking.

| Pattern | Detection | |---------|-----------| | Premature Completion | Declaring done before all scenarios tested | | Format Over Substance | Clean formatting but missing critical content | | Apology Without Fix | "Sorry about that" without behavior change | | Vague Plan | "We'll improve this later" without specifics |

---

Detection Patterns (regex)

# EXTREME patterns
EXTREME_PATTERNS = [
    r"push\s*하지\s*말라",       # "don't push"
    r"test.*없이.*완료",          # "completed without test"
    r"또\s*같다",                 # "same thing again"
    r"이미\s*말했",               # "I already said"
]

# HIGH patterns
HIGH_PATTERNS = [
    r"확인해보세요",              # "please check"
    r"검증\s*해주세요",           # "please verify"
    r"확인\s*필요",              # "needs verification"
    r"개별적으로\s*확인",         # "check individually"
    r"尼古拉斯",                  # "I don't know"
]

# MEDIUM patterns
MEDIUM_PATTERNS = [
    r"나중에\s*改善",             # "improve later"
    r"일단\s*완료",               # "let's call it done for now"
    r"괜찮지\s*않을까",          # "probably fine"
]

---

Usage Modes

Mode 1: Pre-Commit Check

Run before git commit / task completion:

/harsh-critic before_commit

This checks:

1. All [MUST] items in .requirements-lock.md are checked
2. QA evidence file exists
3. No feature deletion patterns in staged changes
4. No EXTREME patterns in recent tool outputs

Mode 2: Auto-Run (posttoolcall hook)

Automatically runs after every tool call when plugin is active. Does NOT block — only warns (EXTREME patterns raise HARD block via plugin).

Mode 3: On-Demand Review

/harsh-critic review

Review last N tool calls for completion bias patterns.

---

Integration with HARD Gates

Harsh Critic integrates with loopy-era-harness plugin:

post_tool_call hook
    ↓
harsh_critic_check() in __init__.py
    ↓
EXTREME pattern found?
    → YES: hard_block() → SystemExit(2)
    → NO: HIGH/MEDIUM → soft_warn()

---

Phase 1 — Scan Output

1.1 Collect recent outputs

Get last N tool results. Focus on:

• Bash/terminal outputs (where completion is declared)
• Write/edit confirmations
• Any message containing completion phrases

1.2 Run pattern matching

For each EXTREME pattern:

• If found → immediate HARD block
• Log to harsh_critic_log.jsonl

For each HIGH pattern:

• Flag for correction
• Log to harsh_critic_log.jsonl

For each MEDIUM pattern:

• Warn but continue
• Log to harsh_critic_log.jsonl

---

Phase 2 — Context Check

2.1 Check requirement lock

cat .requirements-lock.md

If unchecked [MUST] items exist → EXTREME (not done)

2.2 Check QA evidence

cat .qa-evidence.json 2>/dev/null || echo "MISSING"

If missing → EXTREME (not verified)

2.3 Check staged diff

git diff --cached

If deletion patterns found → EXTREME (feature removed)

---

Phase 3 — Response

3.1 If EXTREME

[HARD BLOCK] Harsh Critic — EXTREME violation detected

Pattern: {pattern_name}
Evidence: {matching text}

The task is NOT complete.
Fix the issue before proceeding.

Raise SystemExit(2) via hard_block().

3.2 If HIGH

[HIGH] Harsh Critic — correction required

Pattern: {pattern_name}
Issue: {what's wrong}
Required: {what needs to happen}

Fix this now. Do not continue until addressed.

Do not proceed. Require fix.

3.3 If MEDIUM

[MEDIUM] Harsh Critic — warning

Pattern: {pattern_name}
Suggestion: {what to consider}

This is not blocking, but address if time permits.

Continue, but note the warning.

---

Logging

All checks logged to ~/.drewgent/harsh_critic_log.jsonl:

{"timestamp": "...", "tier": "EXTREME", "pattern": "...", "text": "..."}

This log feeds /self-improve for pattern analysis.

---

File Locations

| Purpose | Path | |---------|------| | Plugin | ~/.drewgent/plugins/loopy-era-harness/ | | Hard gates | ~/.drewgent/plugins/loopy-era-harness/__init__.py | | Log | ~/.drewgent/harsh_critic_log.jsonl | | Rules | ~/.drewgent/rules/ |

---

Key Principle

> "The user doesn't see your reasoning — they only see the result. > If the result looks like 'done' but isn't actually done, you failed."

---

Related Skills

• /loopy-era-self-improve — analyze harsh critic logs for rule extraction
• /loopy-era-qa-cycle — proper QA with evidence
• /loopy-era-eval — measure harness completeness

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: humanerd-drew
• Source: humanerd-drew/opencode-drewgent
• License: MIT

Install and usage instructions live in the source repository linked above.