#!/usr/bin/env bash
# claudecodex — launch Claude Code with GPT-5.6 via a local Codex-auth proxy.
#
# Plain `claude` is untouched: every override below is exported only inside
# this process, then we exec claude. Nothing leaks into your shell or your
# normal Anthropic-authenticated sessions.
#
# Proxy: https://github.com/raine/claude-code-proxy (Anthropic-compatible
# endpoint on :18765, backed by ChatGPT Plus/Pro Codex auth). Needs a proxy
# build that knows the GPT-5.6 models (v0.1.8+).
#
# Model switching (GPT-5.6 family: sol = best coding, terra = mid, luna = budget):
#   claudecodex           → gpt-5.6-sol    (default / "big")
#   claudecodex --mini    → gpt-5.6-luna   (fast/cheap; alias --luna)
#   claudecodex --big     → gpt-5.6-sol    (explicit; alias --sol)
#   claudecodex --terra   → gpt-5.6-terra  (mid tier)
#   In-session: /model → pick gpt-5.6-sol or gpt-5.6-luna, then press 's'
#               (session-only) so plain `claude`'s default stays untouched.
#   The big + small models are always listed in the /model picker.
#
# Thinking level: gpt-5.6-sol launches at MAX effort by default — the highest
# level reachable through claude-code-proxy today. (Sol's "ultra" parallel-
# subagent mode exists in the official Codex app but is not plumbed through the
# proxy yet; the proxy's ceiling is "max". Setting CLAUDECODEX_EFFORT=ultra is
# auto-downgraded to max until a proxy build supports it.) luna/terra keep
# Claude Code's own effort. Override any launch with CLAUDECODEX_EFFORT or
# --effort; change it live with /effort.
#
# One-time setup:
#   curl -fsSL https://raw.githubusercontent.com/raine/claude-code-proxy/main/scripts/install.sh | bash
#   claude-code-proxy codex auth login
#
# Permissions: starts in --dangerously-skip-permissions (bypass) mode by default,
# so tool calls run without prompting. Set CLAUDECODEX_SKIP_PERMISSIONS=0 to start
# with normal permission prompts instead.
#
# Tunables (set in your environment if you want non-defaults):
#   CLAUDECODEX_PORT               proxy port           (default: 18765)
#   CLAUDECODEX_MODEL              "big" model          (default: gpt-5.6-sol)
#   CLAUDECODEX_SMALL_MODEL        fast/cheap model     (default: gpt-5.6-luna)
#   CLAUDECODEX_EFFORT             thinking level        (default: max on sol)
#   CLAUDECODEX_CONTEXT            context window        (default: 372000)
#   CLAUDECODEX_SKIP_PERMISSIONS   1=bypass, 0=normal    (default: 1)
#   CLAUDECODEX_PROXY_BIN          proxy binary          (default: claude-code-proxy)
#   CLAUDECODEX_AGENT_VIEW=1       show jobs dashboard on start
set -euo pipefail

PROXY_PORT="${CLAUDECODEX_PORT:-18765}"
PROXY_BIN="${CLAUDECODEX_PROXY_BIN:-claude-code-proxy}"
BIG_MODEL="${CLAUDECODEX_MODEL:-gpt-5.6-sol}"
SMALL_MODEL="${CLAUDECODEX_SMALL_MODEL:-gpt-5.6-luna}"
MID_MODEL="gpt-5.6-terra"
PROXY_LOG="${XDG_STATE_HOME:-$HOME/.local/state}/claudecodex/proxy.log"

# Pick the active model from sugar flags, then strip them so the rest passes
# through to claude untouched. We use claude's own --model flag, which is
# session-scoped and never written to ~/.claude/settings.json — so switching
# here can't change plain `claude`'s default model.
ACTIVE_MODEL=""
ARGS=()
for arg in "$@"; do
  case "$arg" in
    --mini|--luna|--gpt-5.6-luna) ACTIVE_MODEL="$SMALL_MODEL" ;;
    --big|--sol|--gpt-5.6-sol)    ACTIVE_MODEL="$BIG_MODEL" ;;
    --terra|--gpt-5.6-terra)      ACTIVE_MODEL="$MID_MODEL" ;;
    *)                            ARGS+=("$arg") ;;
  esac
done

port_open() {
  (exec 3<>"/dev/tcp/127.0.0.1/${PROXY_PORT}") 2>/dev/null
}

if ! port_open; then
  if ! command -v "$PROXY_BIN" >/dev/null 2>&1; then
    echo "claudecodex: proxy binary '$PROXY_BIN' not found." >&2
    echo "  Install: curl -fsSL https://raw.githubusercontent.com/raine/claude-code-proxy/main/scripts/install.sh | bash" >&2
    echo "  Login:   claude-code-proxy codex auth login" >&2
    exit 1
  fi
  mkdir -p "$(dirname "$PROXY_LOG")"
  echo "claudecodex: starting proxy on 127.0.0.1:${PROXY_PORT} (log: ${PROXY_LOG})" >&2
  PORT="$PROXY_PORT" nohup "$PROXY_BIN" serve >>"$PROXY_LOG" 2>&1 &
  disown
  for _ in $(seq 1 50); do
    port_open && break
    sleep 0.1
  done
  if ! port_open; then
    echo "claudecodex: proxy did not come up on :${PROXY_PORT} — check ${PROXY_LOG}" >&2
    exit 1
  fi
fi

export ANTHROPIC_BASE_URL="http://localhost:${PROXY_PORT}"
export ANTHROPIC_AUTH_TOKEN="claudecodex-local"
# Both models are always exported so both appear in the /model picker: the big
# one as the main model, the small one as the Haiku/fast slot. The active model
# for this launch is chosen via the --model flag below (default = big).
export ANTHROPIC_MODEL="$BIG_MODEL"
export ANTHROPIC_SMALL_FAST_MODEL="$SMALL_MODEL"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="$SMALL_MODEL"

# Open the normal chat screen, not the background-jobs dashboard ("agent view").
# Set CLAUDECODEX_AGENT_VIEW=1 to get the dashboard back. Note: while disabled,
# background-agent features (--bg, claude agents) are unavailable in this session.
if [ "${CLAUDECODEX_AGENT_VIEW:-0}" != "1" ]; then
  export CLAUDE_CODE_DISABLE_AGENT_VIEW=1
fi

# Context window. gpt-5.6-sol's Codex window is 372K tokens (~353K effective
# after the 95% multiplier). The 272K figure is only the pricing knee — input
# past 272K bills at ~2x, it is NOT the hard cap. Claude Code otherwise assumes
# 200K for unknown model names, which skews the gauge and compacts too early;
# MAX_CONTEXT_TOKENS sets the real window (applies to non-claude-* models only).
# Claude Code's built-in autocompact buffer keeps a margin below this, so it
# won't overflow past the effective window into "exceeds context window" errors.
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="${CLAUDECODEX_CONTEXT:-372000}"
export CLAUDE_CODE_AUTO_COMPACT_WINDOW="${CLAUDECODEX_CONTEXT:-372000}"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
export CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1

# Build the flags claude launches with. --model is session-scoped (chosen above).
# --dangerously-skip-permissions is added by default; skipped if you set
# CLAUDECODEX_SKIP_PERMISSIONS=0, or if you already passed any permission flag.
LAUNCH_FLAGS=()
[ -n "$ACTIVE_MODEL" ] && LAUNCH_FLAGS+=(--model "$ACTIVE_MODEL")

# Default thinking level: the "big" model (gpt-5.6-sol) launches at max effort —
# the highest reachable through the proxy. Other models keep Claude Code's own
# effort. CLAUDECODEX_EFFORT overrides. --effort is session-scoped, so plain
# `claude`'s effort setting is never touched.
RESOLVED_MODEL="${ACTIVE_MODEL:-$BIG_MODEL}"
DEFAULT_EFFORT=""
[ "$RESOLVED_MODEL" = "$BIG_MODEL" ] && DEFAULT_EFFORT="max"
EFFORT="${CLAUDECODEX_EFFORT:-$DEFAULT_EFFORT}"
# 'ultra' (Sol's parallel-subagent mode) isn't exposed by claude-code-proxy yet
# (its ceiling is max), and Claude Code has no ultra effort — passing it clamps
# to xhigh. Map it to max, the real ceiling, so the intent still gets the most.
if [ "$EFFORT" = "ultra" ]; then
  echo "claudecodex: 'ultra' effort isn't reachable through the proxy yet — using 'max'." >&2
  EFFORT="max"
fi

user_set_perms=0
user_set_effort=0
for arg in "${ARGS[@]}"; do
  case "$arg" in
    --dangerously-skip-permissions|--allow-dangerously-skip-permissions|--permission-mode|--permission-mode=*)
      user_set_perms=1 ;;
    --effort|--effort=*) user_set_effort=1 ;;
  esac
done
[ -n "$EFFORT" ] && [ "$user_set_effort" = "0" ] && LAUNCH_FLAGS+=(--effort "$EFFORT")
if [ "${CLAUDECODEX_SKIP_PERMISSIONS:-1}" = "1" ] && [ "$user_set_perms" = "0" ]; then
  LAUNCH_FLAGS+=(--dangerously-skip-permissions)
fi

exec claude "${LAUNCH_FLAGS[@]}" "${ARGS[@]}"