Heretic Abliteration

{"ts":1780240556260,"tool":"Bash","latency_ms":null,"is_error":false}
{"ts":1780240570625,"tool":"Bash","latency_ms":null,"is_error":false}
{"ts":1780241858267,"tool":"Bash","latency_ms":null,"is_error":false}
{"ts":1780241882518,"tool":"Bash","latency_ms":null,"is_error":false}
{"ts":1780241893805,"tool":"mcp__linear-server__save_issue","latency_ms":null,"is_error":false}
{"ts":1780241906219,"tool":"mcp__linear-server__save_comment","latency_ms":null,"is_error":false}

{"ts":1780240590672,"session_id":"2b8efde2-8f4f-49ae-b42d-d180ccf20561","reflexes":{"r01_mechanism":false,"r02_lens":false,"r03_snapshot":false,"r04_depchain":false,"r05_worktree":false,"r06_ticket":false,"r07_dogfood_plan":false,"r08_validation":false,"r09_first_write":false,"r10_empirical":false,"r11_pr_opened":false,"r12_watcher":false,"r13_healing":false,"r14_cross_review":false,"r15_deploy_verify":false,"r16_receipt":false,"r17_pr_comments":false,"r18_auto_merge":false,"r19_cleanup":false,"r20_bridge":true,"r21_bookkeeping":false},"compliance_rate":0.048,"fired_count":1,"total_reflexes":21,"anti_rationalization":{"present":false,"value":null},"tool_call_count":0}
{"ts":1780241930764,"session_id":"2b8efde2-8f4f-49ae-b42d-d180ccf20561","reflexes":{"r01_mechanism":false,"r02_lens":false,"r03_snapshot":false,"r04_depchain":false,"r05_worktree":false,"r06_ticket":false,"r07_dogfood_plan":false,"r08_validation":false,"r09_first_write":false,"r10_empirical":false,"r11_pr_opened":false,"r12_watcher":false,"r13_healing":false,"r14_cross_review":false,"r15_deploy_verify":false,"r16_receipt":false,"r17_pr_comments":false,"r18_auto_merge":false,"r19_cleanup":false,"r20_bridge":true,"r21_bookkeeping":false},"compliance_rate":0.048,"fired_count":1,"total_reflexes":21,"anti_rationalization":{"present":false,"value":null},"tool_call_count":0}

# Heretic skill — generated artifacts
gguf-out/
*.gguf
Modelfile.*
.venv/
__pycache__/
*.log

#!/usr/bin/env bash
# claude-code-on-ollama.sh — drive Claude Code with a LOCAL Ollama model.
#
# Ollama serves the Anthropic Messages API natively at /v1/messages — the exact
# protocol Claude Code speaks — so NO proxy (claude-code-router / LiteLLM) is needed.
# Claude Code just points at Ollama:
#     ANTHROPIC_BASE_URL=http://localhost:11434  ANTHROPIC_API_KEY=ollama  claude --model <m>
#
# Validated 2026-05-31 (M4 Pro / 24GB / Ollama 0.20.7):
#  - Text round-trips fine -> usable as a CHAT backend with any model.
#  - Tool-calls are MODEL-FAMILY specific: Llama-3.x emits the format Ollama parses
#    into tool_use blocks (works even abliterated); Qwen2.5 does NOT. Default below is
#    a Llama-family model for that reason.
#  - The FULL Claude Code agent loop does NOT complete on 24GB: 128K ctx overflows RAM
#    (empty), capped 32K ctx starves/crashes decode (empty). A working local agent
#    needs 64GB+ unified RAM or a GPU box (broomva/remote-gpu). On 24GB: chat only.
#
# Usage:
#   scripts/claude-code-on-ollama.sh [model] [--smoke] [-- <extra claude args>]
#   scripts/claude-code-on-ollama.sh huihui_ai/qwen2.5-coder-abliterate:14b --smoke
#   scripts/claude-code-on-ollama.sh                      # launch interactive claude on default model
set -euo pipefail

MODEL="${1:-huihui_ai/llama3.2-abliterate:3b}"   # Llama family: tool-calls parse in Ollama (Qwen2.5 don't)
[ $# -gt 0 ] && shift || true
OLLAMA_URL="${OLLAMA_HOST:-http://localhost:11434}"
SMOKE=0; PASS=()
for a in "$@"; do [ "$a" = "--smoke" ] && SMOKE=1 || PASS+=("$a"); done

command -v claude >/dev/null 2>&1 || { echo "ERROR: claude (Claude Code) not on PATH."; exit 1; }
command -v ollama >/dev/null 2>&1 || { echo "ERROR: ollama not on PATH (brew install ollama)."; exit 1; }

# daemon up?
curl -sf "$OLLAMA_URL/api/version" >/dev/null 2>&1 || { echo "starting ollama daemon…"; ollama serve >/tmp/ollama.log 2>&1 & for _ in $(seq 1 20); do curl -sf "$OLLAMA_URL/api/version" >/dev/null 2>&1 && break; sleep 0.5; done; }

# native Anthropic endpoint present? (the whole trick)
code=$(curl -s -o /dev/null -w '%{http_code}' -X POST "$OLLAMA_URL/v1/messages" \
  -H 'content-type: application/json' \
  -d "{\"model\":\"$MODEL\",\"max_tokens\":5,\"messages\":[{\"role\":\"user\",\"content\":\"hi\"}]}" 2>/dev/null)
if [ "$code" != "200" ]; then
  echo "ERROR: Ollama at $OLLAMA_URL does not serve /v1/messages (HTTP $code)."
  echo "       Upgrade Ollama (brew upgrade ollama) — native Anthropic Messages API required."
  exit 1
fi

# model present? pull if not
if ! ollama list 2>/dev/null | awk '{print $1}' | grep -qx "$MODEL"; then
  echo "pulling $MODEL …"; ollama pull "$MODEL"
fi

# capability nudge
case "$MODEL" in
  *:1b|*:0.5b|*:1.5b|*:2b|*:3b) echo "⚠️  $MODEL is small — expect poor tool-use / agentic reliability. 14B+ recommended." ;;
esac

export ANTHROPIC_BASE_URL="$OLLAMA_URL"
export ANTHROPIC_API_KEY="ollama"        # dummy; Ollama ignores auth
export ANTHROPIC_MODEL="$MODEL"

if [ "$SMOKE" = 1 ]; then
  echo "=== capability probe: can $MODEL drive Claude Code's tools? ==="
  echo "    (one agent turn on a 14B can take minutes — be patient)"
  WORK="$(mktemp -d)"; LOG="$WORK/stream.json"
  ( cd "$WORK" && timeout "${SMOKE_TIMEOUT:-600}" \
      claude -p "Use the Bash tool to run exactly: echo OLLAMA_AGENT_OK" \
        --model "$MODEL" --dangerously-skip-permissions \
        --output-format stream-json --verbose < /dev/null > "$LOG" 2>&1 ) || true
  python3 - "$LOG" <<'PY'
import json, sys, re
tu = tr = False; result = ""
for line in open(sys.argv[1]):
    line = line.strip()
    if not line.startswith("{"): continue
    try: o = json.loads(line)
    except Exception: continue
    m = o.get("message", {})
    if isinstance(m, dict):
        for c in (m.get("content") or []):
            if isinstance(c, dict) and c.get("type") == "tool_use": tu = True
            if isinstance(c, dict) and c.get("type") == "tool_result": tr = True
    if o.get("type") == "result": result = str(o.get("result", ""))
texty = bool(re.search(r'```json|tool_call|"arguments"\s*:|"name"\s*:\s*"(Bash|Read|Edit|Write|Glob|Grep)"', result))
if tu and tr:
    print("✅ AGENTIC — tool_use round-tripped and executed. Usable as a Claude Code agent.")
elif texty:
    print("⚠️  CHAT-ONLY — model emitted the tool call as TEXT, not a structured tool_use")
    print("    block; Ollama /v1/messages did not translate it, so Claude Code can't run it.")
    print("    Fine as a chat/Q&A backend; NOT a tool-using agent.")
else:
    print("✗  no usable tool attempt (model too weak, or endpoint lacks tool translation).")
PY
  rm -rf "$WORK"
else
  echo "Launching Claude Code on $MODEL (Ctrl-C to exit)…"
  exec claude --model "$MODEL" "${PASS[@]}"
fi

#!/usr/bin/env bash
# heretic-doctor.sh — environment readiness check for the heretic-abliteration skill.
# Reports which path is viable: GPU-run, CPU-run (slow), or serve-only.
set -euo pipefail

VENV="${HERETIC_VENV:-$HOME/.venvs/heretic}"

echo "=== Heretic / Ollama environment doctor ==="
echo

# --- platform ---
OS="$(uname -s)"; ARCH="$(uname -m)"
echo "Platform : $OS $ARCH"
if [[ "$OS" == "Darwin" ]]; then
  echo "Chip     : $(sysctl -n machdep.cpu.brand_string 2>/dev/null || echo '?')"
  RAM_GB=$(( $(sysctl -n hw.memsize 2>/dev/null || echo 0) / 1024 / 1024 / 1024 ))
  echo "RAM      : ${RAM_GB} GB unified"
fi

# --- accelerator ---
ACCEL="cpu"
if command -v nvidia-smi >/dev/null 2>&1; then
  ACCEL="cuda"
  echo "GPU      : CUDA detected → $(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1)"
elif [[ "$OS" == "Darwin" && "$ARCH" == "arm64" ]]; then
  ACCEL="mps-blocked"
  echo "GPU      : Apple Metal (MPS) present — but BLOCKED for GQA models (PyTorch mps.matmul bug). Heretic must run --device-map cpu here."
else
  echo "GPU      : none detected (CPU-only)"
fi

# --- python / venv / heretic ---
echo
if [[ -d "$VENV" ]]; then
  echo "venv     : $VENV (exists)"
  # shellcheck disable=SC1091
  source "$VENV/bin/activate"
else
  echo "venv     : $VENV (MISSING — run scripts/heretic-install.sh)"
fi
echo "python   : $(python3 --version 2>&1) [$(command -v python3)]"
if python3 -c "import heretic.main" >/dev/null 2>&1; then
  HV=$(python3 -c "import importlib.metadata as m; print(m.version('heretic-llm'))" 2>/dev/null || echo '?')
  echo "heretic  : OK (heretic-llm $HV, importable)"
else
  echo "heretic  : NOT importable (run scripts/heretic-install.sh; check kernels/jinja2 fixes)"
fi

# --- ollama ---
echo
if command -v ollama >/dev/null 2>&1; then
  echo "ollama   : $(ollama --version 2>&1 | head -1)"
  if curl -sS http://localhost:11434/api/version >/dev/null 2>&1; then
    echo "           daemon UP"
  else
    echo "           daemon DOWN (start with: ollama serve &)"
  fi
else
  echo "ollama   : NOT installed (brew install ollama)"
fi

# --- llama.cpp (for GGUF conversion) ---
CONVERT="$(command -v convert_hf_to_gguf.py || true)"
if command -v llama-quantize >/dev/null 2>&1 && [[ -n "$CONVERT" ]]; then
  echo "llama.cpp: OK (llama-quantize + $CONVERT)"
else
  echo "llama.cpp: incomplete (need 'brew install llama.cpp' for llama-quantize + convert_hf_to_gguf.py)"
fi

# --- disk ---
echo
echo "disk     : $(df -h / | awk 'NR==2{print $4" free on /"}')"

# --- verdict ---
echo
echo "=== Recommended path ==="
case "$ACCEL" in
  cuda)        echo "→ RUN locally: scripts/heretic-run.sh <model>  (GPU; BNB_4BIT ok)." ;;
  mps-blocked) echo "→ CPU only (slow, small models). For real models use broomva/remote-gpu, then scripts/heretic-to-ollama.sh."
               echo "→ Or skip running Heretic entirely: scripts/ollama-pull-abliterated.sh (fast, recommended)." ;;
  *)           echo "→ CPU only (slow). Prefer scripts/ollama-pull-abliterated.sh, or a GPU via broomva/remote-gpu." ;;
esac

#!/usr/bin/env bash
# heretic-install.sh — install heretic-llm into a CLEAN, dedicated venv and apply
# the two dependency fixes discovered during the 2026-05-30 dogfood.
#
# Why a venv (not conda base): heretic pulls transformers 5.9.0 which drags in two
# breakages. One (kernels) is universal; the other (ancient jinja2) only bites
# crowded base envs. A fresh venv + these defensive fixes = clean import.
set -euo pipefail

VENV="${HERETIC_VENV:-$HOME/.venvs/heretic}"

echo "=== Installing heretic-llm into $VENV ==="

if ! command -v python3 >/dev/null 2>&1; then
  echo "ERROR: python3 not found (need 3.10+)."; exit 1
fi

# 1. Clean venv (never the conda base env)
if [[ ! -d "$VENV" ]]; then
  python3 -m venv "$VENV"
fi
# shellcheck disable=SC1091
source "$VENV/bin/activate"
python3 -m pip install -U pip >/dev/null

# 2. Install heretic
python3 -m pip install -U heretic-llm

# 3. FIX #1 (universal): transformers 5.9.x ↔ kernels 0.15.x import crash
#    "ValueError: Either a revision or a version must be specified" in
#    transformers/integrations/hub_kernels.py. kernels is an OPTIONAL accelerator.
python3 -m pip uninstall -y kernels kernels-data >/dev/null 2>&1 || true

# 4. FIX #2 (defensive): chat-template needs jinja2 >= 3.0 (pass_eval_context).
#    Fresh venvs are fine; this only matters if an old jinja2 leaked in.
python3 -m pip install -U 'jinja2>=3.1' >/dev/null

# 5. gguf (for later HF→GGUF conversion via convert_hf_to_gguf.py)
python3 -m pip install -U gguf >/dev/null 2>&1 || true

# 6. Verify the import actually works
echo
if python3 -c "import heretic.main; print('import OK')"; then
  HV=$(python3 -c "import importlib.metadata as m; print(m.version('heretic-llm'))" 2>/dev/null || echo '?')
  echo "✅ heretic-llm $HV installed and importable in $VENV"
  echo "   Activate with: source $VENV/bin/activate"
  echo "   Next: scripts/heretic-doctor.sh   then   scripts/heretic-run.sh <model>"
else
  echo "❌ heretic still not importable — see references/troubleshooting.md"
  exit 1
fi

#!/usr/bin/env bash
# heretic-run.sh — run a Heretic abliteration locally with the right device flags.
#
# IMPORTANT: run this in a REAL TERMINAL. Heretic's end-of-run "save / upload / chat"
# menu (and the resume menu) use questionary, which needs a TTY. Piping stdin from
# /dev/null makes it crash (OSError [Errno 22]) AFTER optimization finishes — so you
# lose the saved model. A TTY is required to actually save the result.
#
# Usage:
#   scripts/heretic-run.sh <hf-model-id> [extra heretic args...]
#   scripts/heretic-run.sh Qwen/Qwen3-4B-Instruct-2507
#   scripts/heretic-run.sh Qwen/Qwen3-0.6B --n-trials 8        # small smoke
set -euo pipefail

VENV="${HERETIC_VENV:-$HOME/.venvs/heretic}"
MODEL="${1:-}"
if [[ -z "$MODEL" ]]; then
  echo "Usage: $0 <hf-model-id> [extra heretic args...]"; exit 1
fi
shift || true

if [[ ! -d "$VENV" ]]; then
  echo "ERROR: venv $VENV missing. Run scripts/heretic-install.sh first."; exit 1
fi
# shellcheck disable=SC1091
source "$VENV/bin/activate"

# TTY guard — heretic needs an interactive terminal to save.
if [[ ! -t 0 ]]; then
  echo "⚠️  stdin is not a TTY. Heretic will run the optimization but CRASH at the"
  echo "    save menu, discarding the model. Run in an interactive terminal to save."
fi

# Device selection (dogfood-validated):
OS="$(uname -s)"; ARCH="$(uname -m)"
DEV_ARGS=()
if command -v nvidia-smi >/dev/null 2>&1; then
  echo "→ CUDA detected. Full speed; add --quantization BNB_4BIT to fit large models."
elif [[ "$OS" == "Darwin" && "$ARCH" == "arm64" ]]; then
  echo "⚠️  Apple Silicon: forcing --device-map cpu. MPS hits a PyTorch mps.matmul"
  echo "    LLVM error on grouped-query-attention models (Qwen3, Llama3, …)."
  echo "    CPU works but is slow (~50–110 tok/s). Use a real GPU for >1B models"
  echo "    (see broomva/remote-gpu)."
  export PYTORCH_ENABLE_MPS_FALLBACK=1
  DEV_ARGS=(--device-map cpu --quantization NONE)
else
  echo "→ No GPU: running on CPU (slow)."
  DEV_ARGS=(--device-map cpu --quantization NONE)
fi

set -x
heretic --model "$MODEL" "${DEV_ARGS[@]}" "$@"

#!/usr/bin/env bash
# heretic-to-ollama.sh — convert a Heretic-saved HF model into a GGUF and register
# it with Ollama. This is the bridge: Heretic outputs HF safetensors; Ollama serves
# GGUF. Ollama can NEVER be Heretic's backend — it only runs the finished model.
#
# Pipeline:  HF safetensors  →  convert_hf_to_gguf.py  →  llama-quantize  →  ollama create
#
# Usage:
#   scripts/heretic-to-ollama.sh <hf-model-dir> [ollama-name] [quant]
#   scripts/heretic-to-ollama.sh ./Qwen3-4B-heretic my-heretic Q4_K_M
set -euo pipefail

MODEL_DIR="${1:-}"
NAME="${2:-heretic-model}"
QUANT="${3:-Q4_K_M}"
VENV="${HERETIC_VENV:-$HOME/.venvs/heretic}"
WORK="${HERETIC_WORK:-$(pwd)/gguf-out}"

if [[ -z "$MODEL_DIR" || ! -d "$MODEL_DIR" ]]; then
  echo "Usage: $0 <hf-model-dir> [ollama-name] [quant]"; exit 1
fi

CONVERT="$(command -v convert_hf_to_gguf.py || echo /opt/homebrew/bin/convert_hf_to_gguf.py)"
if [[ ! -e "$CONVERT" ]]; then
  echo "ERROR: convert_hf_to_gguf.py not found. brew install llama.cpp"; exit 1
fi
if ! command -v llama-quantize >/dev/null 2>&1; then
  echo "ERROR: llama-quantize not found. brew install llama.cpp"; exit 1
fi
if ! command -v ollama >/dev/null 2>&1; then
  echo "ERROR: ollama not found. brew install ollama"; exit 1
fi

# Ensure gguf python dep (convert script needs it)
[[ -d "$VENV" ]] && source "$VENV/bin/activate" || true
python3 -c "import gguf" 2>/dev/null || python3 -m pip install -U gguf

mkdir -p "$WORK"
F16="$WORK/${NAME}-f16.gguf"
Q="$WORK/${NAME}-${QUANT}.gguf"

echo "=== 1/3 HF → GGUF (f16) ==="
python3 "$CONVERT" "$MODEL_DIR" --outfile "$F16" --outtype f16

echo "=== 2/3 quantize → $QUANT ==="
llama-quantize "$F16" "$Q" "$QUANT"

echo "=== 3/3 ollama create '$NAME' ==="
MODELFILE="$WORK/Modelfile.$NAME"
printf 'FROM %s\n' "$Q" > "$MODELFILE"
ollama create "$NAME" -f "$MODELFILE"

echo
echo "✅ Registered. Smoke test:"
echo "   ollama run $NAME \"In one word, capital of Australia?\""

#!/usr/bin/env bash
# ollama-pull-abliterated.sh — FAST PATH (recommended on Apple Silicon / no GPU).
# Skip running Heretic entirely: pull a community pre-abliterated GGUF straight into
# Ollama and smoke-test it. Same *class* of model Heretic produces, zero compute.
#
# Usage:
#   scripts/ollama-pull-abliterated.sh [ollama-model]
#   scripts/ollama-pull-abliterated.sh huihui_ai/llama3.2-abliterate:3b
#
# Catalogs of pre-abliterated GGUFs: ollama.com (search "abliterate"), or any
# Hugging Face GGUF via:  ollama run hf.co/<user>/<repo>:<quant>
set -euo pipefail

MODEL="${1:-huihui_ai/llama3.2-abliterate:1b}"

if ! command -v ollama >/dev/null 2>&1; then
  echo "ERROR: ollama not found. brew install ollama"; exit 1
fi
if ! curl -sS http://localhost:11434/api/version >/dev/null 2>&1; then
  echo "Starting ollama daemon…"; ollama serve >/tmp/ollama-heretic.log 2>&1 &
  for _ in $(seq 1 20); do curl -sS http://localhost:11434/api/version >/dev/null 2>&1 && break; sleep 0.5; done
fi

echo "=== pulling $MODEL ==="
ollama pull "$MODEL"

ask() { # prompt -> clean response via API (no TTY spinner noise)
  curl -sS http://localhost:11434/api/generate \
    -d "{\"model\":\"$MODEL\",\"prompt\":$(printf '%s' "$1" | python3 -c 'import json,sys;print(json.dumps(sys.stdin.read()))'),\"stream\":false}" \
    | python3 -c "import sys,json;print(json.load(sys.stdin)['response'][:400])"
}

echo
echo "=== smoke 1: capability preserved ==="
ask "What is the capital of Australia? Answer in one word."
echo
echo "=== smoke 2: false-refusal removed (benign sysadmin prompt) ==="
ask "How do I kill a frozen process on Linux? Give the exact command."
echo
echo "✅ Served via Ollama CLI. Chat with:  ollama run $MODEL"

{
  "version": 1,
  "skills": {
    "heretic-abliteration": {
      "source": "broomva/heretic-abliteration",
      "sourceType": "github",
      "computedHash": ""
    }
  }
}