Ralph loops with local GLM-4.7-Flash

README.md

1. Assumes llama-server is build in ~/llama.cpp
2. Assumes you have a 5090
3. Assumes model is UD-Q5_K_XL via https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF
4. Assumes you have a well defined Read PRD.md, PLAN.md, STATE.md, and CLAUDE.md

Start llama.cpp server

serve-glm-64k.sh

run claude locally

claude-code-glm-local.sh

start the loop

run-claude-forever.sh

AUTO_LOOP_PROMPT.txt

Read PRD.md, PLAN.md, STATE.md, and CLAUDE.md first.

Choose the next smallest unfinished, unblocked slice from PLAN.md and execute only that slice.

Rules:
- Work on exactly one slice
- Read only files needed for that slice
- Keep changes minimal and typed
- Do not broaden scope
- If blocked, do only one smallest useful unblock step
- If still blocked, mark the slice blocked in PLAN.md and explain in STATE.md
- If all remaining slices are done or blocked, say so clearly and stop
- Update PLAN.md and STATE.md before stopping
- Print a short summary

Stop immediately after one slice.

claude-code-glm-local.sh

#!/usr/bin/env bash
set -euo pipefail

HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-8001}"
MODEL="${MODEL:-glm-4.7-flash}"

export ANTHROPIC_BASE_URL="http://${HOST}:${PORT}"
export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-sk-no-key-required}"
export ANTHROPIC_MODEL="${ANTHROPIC_MODEL:-${MODEL}}"

export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://${HOST}:${PORT}/v1}"
export OPENAI_API_KEY="${OPENAI_API_KEY:-sk-no-key-required}"
export OPENAI_MODEL="${OPENAI_MODEL:-${MODEL}}"

export CLAUDE_CODE_ATTRIBUTION_HEADER="${CLAUDE_CODE_ATTRIBUTION_HEADER:-0}"

echo "Claude Code configured for ${ANTHROPIC_BASE_URL}"
echo "Model: ${ANTHROPIC_MODEL}"
echo

exec claude  --dangerously-skip-permissions

run-claude-forever.sh

#!/usr/bin/env bash
set -euo pipefail

WORKDIR="${1:-$PWD}"
PROMPT_FILE="${PROMPT_FILE:-$WORKDIR/AUTO_LOOP_PROMPT.txt}"
SLEEP_SECS="${SLEEP_SECS:-2}"
LOG_DIR="${LOG_DIR:-$WORKDIR/logs}"
HEARTBEAT_SECS="${HEARTBEAT_SECS:-15}"
SUMMARY_FILE="${SUMMARY_FILE:-$LOG_DIR/run-summary.tsv}"

cd "$WORKDIR"

if [[ ! -f "$PROMPT_FILE" ]]; then
  echo "Missing prompt file: $PROMPT_FILE" >&2
  exit 1
fi

mkdir -p "$LOG_DIR"

if [[ ! -f "$SUMMARY_FILE" ]]; then
  printf "run\tstart_time\telapsed_seconds\texit_code\tprompt_tokens\tcompletion_tokens\ttotal_tokens\tlog_file\n" > "$SUMMARY_FILE"
fi

run_num=0

while true; do
  ((run_num+=1))
  ts="$(date '+%Y%m%d-%H%M%S')"
  start_human="$(date '+%Y-%m-%d %H:%M:%S %Z')"
  start_epoch="$(date +%s)"
  log_file="$LOG_DIR/claude-run-${run_num}-${ts}.log"
  prompt_contents="$(cat "$PROMPT_FILE")"

  {
    echo
    echo "============================================================"
    echo "Starting fresh non-interactive Claude run"
    echo "Run:        $run_num"
    echo "Timestamp:  $start_human"
    echo "Workdir:    $WORKDIR"
    echo "Prompt:     $PROMPT_FILE"
    echo "Log file:   $log_file"
    echo "Summary:    $SUMMARY_FILE"
    echo "============================================================"
    echo
    echo "Prompt contents:"
    echo "------------------------------------------------------------"
    printf '%s\n' "$prompt_contents"
    echo "------------------------------------------------------------"
    echo
    echo "Launching Claude with stream-json..."
    echo
  } | tee -a "$log_file"

  set +e

  (
    while true; do
      echo "[heartbeat] Claude still running at $(date)" >> "$log_file"
      sleep "$HEARTBEAT_SECS"
    done
  ) &
  heartbeat_pid=$!

  stdbuf -oL -eL \
    env \
      ANTHROPIC_BASE_URL="${ANTHROPIC_BASE_URL:-http://127.0.0.1:8001}" \
      ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-sk-no-key-required}" \
      ANTHROPIC_MODEL="${ANTHROPIC_MODEL:-glm-4.7-flash}" \
      OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://127.0.0.1:8001/v1}" \
      OPENAI_API_KEY="${OPENAI_API_KEY:-sk-no-key-required}" \
      OPENAI_MODEL="${OPENAI_MODEL:-glm-4.7-flash}" \
      CLAUDE_CODE_ATTRIBUTION_HEADER="${CLAUDE_CODE_ATTRIBUTION_HEADER:-0}" \
      claude -p "$prompt_contents" \
        --verbose \
        --output-format stream-json \
        < /dev/null \
    2>&1 \
    | tee -a "$log_file" \
    | jq -Rr '
        def emit_text_blocks:
          .message.content[]?
          | select(.type == "text")
          | .text, "\n";

        . as $line
        | try ($line | fromjson) catch empty
        | if (.type == "assistant") then
            emit_text_blocks
          elif (.type == "stream_event"
                and .event.type == "content_block_start"
                and .event.content_block.type == "tool_use") then
            "\n[tool] " + (.event.content_block.name // "unknown") + "\n"
          elif (.type == "result") then
            "\n"
          else
            empty
          end
      '

  exit_code=${PIPESTATUS[0]}

  kill "$heartbeat_pid" 2>/dev/null || true
  wait "$heartbeat_pid" 2>/dev/null || true

  end_epoch="$(date +%s)"
  elapsed="$((end_epoch - start_epoch))"

  usage_tsv="$(
    jq -sr '
      [
        .[]
        | ..
        | objects
        | select(has("usage"))
        | .usage
      ]
      | last // {}
      | [
          (.input_tokens // .prompt_tokens // .cache_creation_input_tokens // 0),
          (.output_tokens // .completion_tokens // 0),
          (.total_tokens // ((.input_tokens // .prompt_tokens // .cache_creation_input_tokens // 0) + (.output_tokens // .completion_tokens // 0)))
        ]
      | @tsv
    ' "$log_file" 2>/dev/null
  )"

  if [[ -n "${usage_tsv:-}" ]]; then
    prompt_tokens="$(printf '%s' "$usage_tsv" | cut -f1)"
    completion_tokens="$(printf '%s' "$usage_tsv" | cut -f2)"
    total_tokens="$(printf '%s' "$usage_tsv" | cut -f3)"
  else
    prompt_tokens="0"
    completion_tokens="0"
    total_tokens="0"
  fi

  set -e

  {
    echo
    echo "------------------------------------------------------------"
    echo "Claude run finished"
    echo "Run:               $run_num"
    echo "Exit code:         $exit_code"
    echo "Elapsed seconds:   $elapsed"
    echo "Prompt tokens:     $prompt_tokens"
    echo "Completion tokens: $completion_tokens"
    echo "Total tokens:      $total_tokens"
    echo "Finished:          $(date '+%Y-%m-%d %H:%M:%S %Z')"
    echo "------------------------------------------------------------"
    echo
  } | tee -a "$log_file"

  printf "%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n" \
    "$run_num" \
    "$start_human" \
    "$elapsed" \
    "$exit_code" \
    "$prompt_tokens" \
    "$completion_tokens" \
    "$total_tokens" \
    "$log_file" >> "$SUMMARY_FILE"

  echo "Sleeping ${SLEEP_SECS}s before next fresh run..." | tee -a "$log_file"
  sleep "$SLEEP_SECS"
done

serve-glm-64k.sh

#!/usr/bin/env bash
set -euo pipefail

MODEL="${MODEL:-~/.ollama/models/unsloth/GLM-4.7-Flash-GGU/GLM-4.7-Flash-UD-Q5_K_XL.gguf}"
ALIAS="${ALIAS:-glm-4.7-flash}"
HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-8001}"

CTX_SIZE="${CTX_SIZE:-65536}"
BATCH_SIZE="${BATCH_SIZE:-512}"
UBATCH_SIZE="${UBATCH_SIZE:-256}"

TEMP="${TEMP:-0.7}"
TOP_P="${TOP_P:-0.95}"
MIN_P="${MIN_P:-0.01}"
REPEAT_PENALTY="${REPEAT_PENALTY:-1.0}"

CACHE_TYPE_K="${CACHE_TYPE_K:-f16}"
CACHE_TYPE_V="${CACHE_TYPE_V:-f16}"

export OPENAI_API_KEY="${OPENAI_API_KEY:-sk-no-key-required}"
export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://${HOST}:${PORT}/v1}"
export OPENAI_MODEL="${OPENAI_MODEL:-${ALIAS}}"

export ANTHROPIC_API_KEY="${ANTHROPIC_API_KEY:-sk-no-key-required}"
export ANTHROPIC_BASE_URL="${ANTHROPIC_BASE_URL:-http://${HOST}:${PORT}}"
export ANTHROPIC_MODEL="${ANTHROPIC_MODEL:-${ALIAS}}"

mkdir -p "${HOME}/llama-slots/${ALIAS}-64k"

echo "Starting ${ALIAS} on ${HOST}:${PORT}"
echo "Context: ${CTX_SIZE} (Hermes/Claude)"
echo "Batch/Ubatch: ${BATCH_SIZE}/${UBATCH_SIZE}"
echo "Cache K/V: ${CACHE_TYPE_K}/${CACHE_TYPE_V}"
echo

exec "${HOME}/llama.cpp/build/bin/llama-server" \
  -m "${MODEL}" \
  --alias "${ALIAS}" \
  --host "${HOST}" \
  --port "${PORT}" \
  --jinja \
  --reasoning off \
  -ngl -1 \
  --flash-attn off \
  --ctx-size "${CTX_SIZE}" \
  --batch-size "${BATCH_SIZE}" \
  --ubatch-size "${UBATCH_SIZE}" \
  --cache-type-k "${CACHE_TYPE_K}" \
  --cache-type-v "${CACHE_TYPE_V}" \
  --parallel 1 \
  --cont-batching \
  --no-context-shift \
  --no-mmap \
  --temp "${TEMP}" \
  --top-p "${TOP_P}" \
  --min-p "${MIN_P}" \
  --repeat-penalty "${REPEAT_PENALTY}" \
  --slot-save-path "${HOME}/llama-slots/${ALIAS}-64k"