a01cac0657
Plafonul general saptamanal Max 5x se consuma in ~2 zile, urmate de ~5 zile blocate. Masurat pe 3 masini (Windows, claude-agent, moltbot), 2026-07-29 -> 2026-09-01: 11,15 miliarde tokeni, 435 sesiuni, 64.211 cereri API, dedup pe message.id + requestId. Deficit +32% (~570 prompturi/saptamana). Concluzia nu e cea asteptata: bugetele plafonate in dolari (OpenCode Zen $20, OpenCode Go) se evapora la acest volum -- Go a tinut 1-2 zile in practica. Planurile Z.AI se contorizeaza in prompturi, nu in dolari, si rezista la un workload cu 7,1 cereri API per prompt. Z.AI Coding Lite ($18/luna) acopera 70% din deficit; Pro ($72) il acopera integral. Verificat empiric ca subagentii nu pot rula pe GLM (in proces, moștenesc auth), dar sesiunile separate pot: endpoint fals local a primit cererea cu tokenul alternativ, fara sa atinga OAuth-ul Max. Masurat: ~186MB RSS si 20-60s pornire per sesiune -- de aici regulile de orchestrare (sarcini mari, max 4 procese, fan-out in interiorul unei sesiuni GLM). Ramane neverificat multiplicatorul de credite pentru GLM-5.3-Flash; la 3x in loc de 1x, Lite pica si decizia se muta pe Pro. - docs/supliment-glm-zai.md: rezumat de implementare (indexat in CLAUDE.md) - claude_usage_report.md + CSV-uri: analiza completa si datele brute - tools/claude_usage_*.py: pipeline de masurare read-only, re-rulabil remote - tools/claude-glm.sh: wrapper de sesiune GLM (task/resume/shell) Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4uzvgm7AyJch5WH8QHRhY
37 lines
1.7 KiB
Bash
Executable File
37 lines
1.7 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Sesiune Claude Code separata pe GLM (Z.AI). NU consuma cota Claude Max.
|
|
#
|
|
# tools/claude-glm.sh task "implementeaza T3 din TODO.md" # one-shot, JSON pe stdout
|
|
# tools/claude-glm.sh resume <session_id> "continua" # predare / continuare
|
|
# tools/claude-glm.sh shell # sesiune interactiva
|
|
# tools/claude-glm.sh -- <argumente claude brute>
|
|
#
|
|
# Env optional: GLM_MODEL (default glm-5.3-flash[1m]), GLM_CONTEXT (default 250000)
|
|
set -euo pipefail
|
|
|
|
KEY_FILE="${ZAI_KEY_FILE:-$HOME/.config/zai/key}"
|
|
[[ -r "$KEY_FILE" ]] || { echo "Lipseste cheia Z.AI in $KEY_FILE (chmod 600)" >&2; exit 1; }
|
|
|
|
export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
|
|
export ANTHROPIC_AUTH_TOKEN="$(tr -d '\r\n' < "$KEY_FILE")"
|
|
# [1m] = fereastra reala de 1M; fara el Claude Code presupune 200k si compacteaza devreme
|
|
MODEL="${GLM_MODEL:-glm-5.3-flash[1m]}"
|
|
export ANTHROPIC_DEFAULT_SONNET_MODEL="$MODEL"
|
|
export ANTHROPIC_DEFAULT_OPUS_MODEL="$MODEL"
|
|
export ANTHROPIC_DEFAULT_HAIKU_MODEL="glm-5.3-flash[1m]"
|
|
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
|
|
export API_TIMEOUT_MS=3000000
|
|
# config separat: transcripturile GLM nu se amesteca cu cele Anthropic in statistica de consum
|
|
export CLAUDE_CONFIG_DIR="${CLAUDE_CONFIG_DIR_GLM:-$HOME/.claude-glm}"
|
|
|
|
CTX="${GLM_CONTEXT:-250000}"
|
|
COMMON=(--model sonnet --autocompact "$CTX" --output-format json)
|
|
|
|
case "${1:-shell}" in
|
|
task) shift; exec claude -p "$*" "${COMMON[@]}" < /dev/null ;;
|
|
resume) sid="$2"; shift 2; exec claude -p "$*" --resume "$sid" "${COMMON[@]}" < /dev/null ;;
|
|
shell) shift || true; exec claude --model sonnet --autocompact "$CTX" "$@" ;;
|
|
--) shift; exec claude "$@" ;;
|
|
*) exec claude "$@" ;;
|
|
esac
|