cloverleaf-larry/lib/oauth.sh

#!/usr/bin/env bash
# oauth.sh — OAuth login flow against Claude.ai for Larry-Anywhere.
#
# Uses the same OAuth client/flow Anthropic's Claude Code CLI uses, so calls
# bill against your Claude Max / Pro subscription quota instead of pay-as-you-go
# API metering. Public client_id; PKCE; out-of-band code paste (no localhost
# server required, works behind any firewall).
#
# Subcommands:
#   login         start the auth flow; print URL; prompt for code
#   refresh       refresh the access token using the stored refresh token
#   ensure        print a valid access token (auto-refreshes if near-expired)
#   status        show current auth state + expiry
#   debug         dump full OAuth diagnostic state (file, parsed times, jq path,
#                 cygpath translation, truncated token previews) — safe to share
#   logout        delete the stored tokens
#
# Storage: $LARRY_HOME/.oauth.json (mode 0600)
#
# Env vars:
#   LARRY_OAUTH_DEBUG=1   emit per-call decision trace from cmd_ensure to stderr.
#                         Default off so the chat log stays clean; flip on when
#                         diagnosing why ensure returns empty.
#
# This is community/unofficial use of Anthropic's OAuth flow. Anthropic could
# tighten it at any time. If OAuth stops working, Larry transparently falls
# back to the API-key path stored in $LARRY_HOME/.env.
set -u
set -o pipefail

LARRY_HOME="${LARRY_HOME:-$HOME/.larry}"
OAUTH_FILE="$LARRY_HOME/.oauth.json"

# Anthropic Claude Code's publicly-visible OAuth client_id. Used by claude-code
# and several community CLI tools (droidrun/mobilerun, motiful/cc-gateway, ...).
#
# Endpoints migrated 2025: claude.ai/oauth/authorize → claude.com/cai/oauth/authorize,
# console.anthropic.com/v1/oauth/token → platform.claude.com/v1/oauth/token,
# console.anthropic.com/oauth/code/callback → platform.claude.com/oauth/code/callback.
# The OLD endpoints return a misleading "rate_limit_error" for any request.
# Scopes also expanded with user:sessions:claude_code, user:mcp_servers,
# user:file_upload — required by the new flow.
CLIENT_ID="${LARRY_OAUTH_CLIENT_ID:-9d1c250a-e61b-44d9-88ed-5944d1962f5e}"
AUTHORIZE_URL="${LARRY_OAUTH_AUTHORIZE_URL:-https://claude.com/cai/oauth/authorize}"
TOKEN_URL="${LARRY_OAUTH_TOKEN_URL:-https://platform.claude.com/v1/oauth/token}"
REDIRECT_URI="${LARRY_OAUTH_REDIRECT_URI:-https://platform.claude.com/oauth/code/callback}"
SCOPE="${LARRY_OAUTH_SCOPE:-org:create_api_key user:profile user:inference user:sessions:claude_code user:mcp_servers user:file_upload}"

die() { printf 'oauth: %s\n' "$*" >&2; exit 1; }

# Dependency check
command -v curl    >/dev/null 2>&1 || die "curl required"
command -v jq      >/dev/null 2>&1 || die "jq required"
command -v openssl >/dev/null 2>&1 || die "openssl required (for PKCE sha256)"

b64url() { base64 | tr '/+' '_-' | tr -d '=' | tr -d '\n'; }

# jqf — run jq against a file, but pipe the file via stdin so bash handles
# the path translation. Needed because on MobaXterm/Cygwin the bundled jq
# may be a Windows-native binary that doesn't understand Cygwin paths like
# /home/mobaxterm/... when they come in as argv. Stdin redirection always
# works because bash does the open() itself.
# Usage: jqf <file> <jq-args...>
jqf() {
  local file="$1"; shift
  jq "$@" < "$file"
}

urlenc() {
  # Minimal RFC3986-ish URL encoder for the bits we need (spaces, /, :)
  local s="$1"
  s="${s// /%20}"
  s="${s//:/%3A}"
  s="${s//\//%2F}"
  printf '%s' "$s"
}

gen_pkce() {
  local verifier challenge
  verifier=$(LC_ALL=C tr -dc 'a-zA-Z0-9-._~' </dev/urandom | head -c 64)
  challenge=$(printf '%s' "$verifier" | openssl dgst -sha256 -binary | b64url)
  printf '%s|%s' "$verifier" "$challenge"
}

cmd_login() {
  mkdir -p "$LARRY_HOME"
  local pkce verifier challenge state
  pkce=$(gen_pkce)
  verifier="${pkce%%|*}"
  challenge="${pkce##*|}"
  state=$(LC_ALL=C tr -dc 'a-zA-Z0-9' </dev/urandom | head -c 32)

  local url
  url="${AUTHORIZE_URL}?code=true"
  url="${url}&client_id=${CLIENT_ID}"
  url="${url}&response_type=code"
  url="${url}&redirect_uri=$(urlenc "$REDIRECT_URI")"
  url="${url}&scope=$(urlenc "$SCOPE")"
  url="${url}&code_challenge=${challenge}"
  url="${url}&code_challenge_method=S256"
  url="${url}&state=${state}"

  cat <<EOF

=== Larry-Anywhere — Claude subscription login ===

This binds Larry to your Claude.ai Max/Pro subscription quota (same flow
Claude Code uses). No API key needed.

1. Open this URL in any browser on any device:

   ${url}

2. Sign in with your Claude account.
3. Approve the app. You'll land on a page that displays a string in the form
       <CODE>#<STATE>
   (Anthropic uses a URL fragment, not a query param, to deliver them.)
   Copy the WHOLE string — both halves and the '#' between them.

4. Paste it here:

EOF
  printf 'authorization code (CODE#STATE): '
  read -r code_input
  [ -z "$code_input" ] && die "no code entered"

  # Split CODE#STATE. If the user pasted only the code (no '#'), keep the
  # state we generated; otherwise verify the returned state matches.
  local code returned_state
  if [[ "$code_input" == *"#"* ]]; then
    code="${code_input%%#*}"
    returned_state="${code_input#*#}"
    if [ -n "$returned_state" ] && [ "$returned_state" != "$state" ]; then
      die "state mismatch — got '$returned_state', expected '$state' (possible CSRF or stale URL; rerun login)"
    fi
  else
    code="$code_input"
  fi

  local resp
  resp=$(curl -sS -X POST "$TOKEN_URL" \
    -H "Content-Type: application/json" \
    -H "Accept: application/json" \
    -H "anthropic-beta: oauth-2025-04-20" \
    -H "User-Agent: claude-cli/2.1.85 (larry-anywhere)" \
    -d "$(jq -n \
        --arg cid       "$CLIENT_ID" \
        --arg code      "$code" \
        --arg verifier  "$verifier" \
        --arg redirect  "$REDIRECT_URI" \
        --arg state     "$state" \
        '{client_id:$cid, grant_type:"authorization_code", code:$code, code_verifier:$verifier, redirect_uri:$redirect, state:$state}')")

  if ! printf '%s' "$resp" | jq -e '.access_token' >/dev/null 2>&1; then
    printf '\nauth failed. server response:\n' >&2
    printf '%s\n' "$resp" | jq . >&2 2>/dev/null || printf '%s\n' "$resp" >&2
    cat >&2 <<EOF

Hints:
  - The callback delivers the code as CODE#STATE (fragment, not query).
    Paste the WHOLE string including '#'. Just CODE alone also works.
  - The code is single-use; if you used it already (even on a failed attempt),
    run 'larry-auth.sh login' again to get a fresh URL.
  - 'rate_limit_error' on a fresh code is the server's misleading mask for
    'malformed/used code' OR 'dead endpoint'. If you JUST upgraded and saw
    that error, double-check TOKEN_URL points at platform.claude.com — old
    console.anthropic.com URLs return rate_limit_error for everything.
    Current (as of 2026-05): https://platform.claude.com/v1/oauth/token
  - If OAuth is genuinely broken, fall back to the API key by deleting any
    oauth file and creating $LARRY_HOME/.env with ANTHROPIC_API_KEY=sk-ant-...
EOF
    exit 1
  fi

  local now; now=$(date +%s)
  umask 077
  # Write to a .new sidecar first, validate it parses + has the required keys,
  # then atomically mv into place. If jq fails mid-pipe, $OAUTH_FILE is never
  # half-written — preventing the "ensure returns empty because the file is
  # corrupt" failure mode that's otherwise invisible to the caller.
  printf '%s' "$resp" | jq --arg now "$now" '. + {fetched_at: ($now | tonumber)}' > "$OAUTH_FILE.new" || {
    rm -f "$OAUTH_FILE.new"
    die "failed to write oauth token (jq pipeline error)"
  }
  if ! jq -e '.access_token and .refresh_token and .fetched_at' < "$OAUTH_FILE.new" >/dev/null 2>&1; then
    rm -f "$OAUTH_FILE.new"
    die "wrote oauth file is missing access_token/refresh_token/fetched_at — aborting"
  fi
  mv "$OAUTH_FILE.new" "$OAUTH_FILE"
  chmod 600 "$OAUTH_FILE"
  printf '\n✓ logged in. Tokens saved to %s (mode 0600).\n' "$OAUTH_FILE"
  cmd_status
}

cmd_refresh() {
  [ -f "$OAUTH_FILE" ] || die "no oauth file at $OAUTH_FILE — run 'larry-auth.sh login' first"
  local refresh_token; refresh_token=$(jqf "$OAUTH_FILE" -r '.refresh_token // empty')
  [ -n "$refresh_token" ] || die "no refresh_token in $OAUTH_FILE — please run login again"

  local resp
  resp=$(curl -sS -X POST "$TOKEN_URL" \
    -H "Content-Type: application/json" \
    -H "Accept: application/json" \
    -H "anthropic-beta: oauth-2025-04-20" \
    -H "User-Agent: claude-cli/2.1.85 (larry-anywhere)" \
    -d "$(jq -n --arg cid "$CLIENT_ID" --arg rt "$refresh_token" \
        '{client_id:$cid, grant_type:"refresh_token", refresh_token:$rt}')")

  if ! printf '%s' "$resp" | jq -e '.access_token' >/dev/null 2>&1; then
    printf 'refresh failed:\n%s\n' "$resp" >&2
    return 1
  fi

  local now; now=$(date +%s)
  # Pre-read the old refresh_token so we don't need --slurpfile (which would
  # take a file path argv-style and break on MobaXterm's Windows-native jq).
  local prev_refresh; prev_refresh=$(jqf "$OAUTH_FILE" -r '.refresh_token // empty')
  if ! printf '%s' "$resp" \
    | jq --arg now "$now" --arg prev "$prev_refresh" \
        '. + {fetched_at: ($now|tonumber), refresh_token: (.refresh_token // $prev)}' \
    > "$OAUTH_FILE.new"; then
    rm -f "$OAUTH_FILE.new"
    printf 'refresh: failed to write new oauth file (jq pipeline error)\n' >&2
    return 1
  fi
  # Validate before clobbering the existing file. If validation fails the old
  # token file stays intact and ensure can still serve the soon-to-expire token.
  if ! jq -e '.access_token and .refresh_token and .fetched_at' < "$OAUTH_FILE.new" >/dev/null 2>&1; then
    rm -f "$OAUTH_FILE.new"
    printf 'refresh: new oauth file is missing required keys — keeping previous file intact\n' >&2
    return 1
  fi
  mv "$OAUTH_FILE.new" "$OAUTH_FILE"
  chmod 600 "$OAUTH_FILE"
  jqf "$OAUTH_FILE" -r '.access_token'
}

# dbg — gated decision-trace logger. Only fires when LARRY_OAUTH_DEBUG=1 is
# set in the env. Goes to stderr so it doesn't pollute the token printed on
# stdout. Caller of `oauth.sh ensure` should capture stderr and surface it
# if the token is empty.
dbg() {
  [ "${LARRY_OAUTH_DEBUG:-0}" = "1" ] || return 0
  printf 'oauth.dbg: %s\n' "$*" >&2
}

cmd_ensure() {
  if [ ! -f "$OAUTH_FILE" ]; then
    dbg "ensure: $OAUTH_FILE does not exist — returning empty"
    printf 'ensure: no oauth file at %s\n' "$OAUTH_FILE" >&2
    return 1
  fi
  # Sanity-check the file actually parses as JSON. If it doesn't (truncated,
  # half-written, mid-disk-full), every jqf below silently returns empty which
  # propagates as "token unavailable" with zero diagnostic info upstream.
  if ! jq -e . < "$OAUTH_FILE" >/dev/null 2>&1; then
    dbg "ensure: $OAUTH_FILE failed to parse as JSON — corrupted?"
    printf 'ensure: %s is not valid JSON (corrupted? rerun login)\n' "$OAUTH_FILE" >&2
    return 1
  fi
  local fetched_at expires_in
  fetched_at=$(jqf "$OAUTH_FILE" -r '.fetched_at // 0')
  expires_in=$(jqf "$OAUTH_FILE" -r '.expires_in // 3600')
  local now; now=$(date +%s)
  local expires_at=$((fetched_at + expires_in))
  local left=$((expires_at - now))
  dbg "ensure: fetched_at=$fetched_at expires_in=$expires_in now=$now expires_at=$expires_at left=${left}s"
  if [ "$now" -ge $((expires_at - 300)) ]; then
    dbg "ensure: within 300s of expiry — refreshing"
    if ! cmd_refresh >/dev/null 2>&1; then
      dbg "ensure: cmd_refresh failed; calling again with stderr exposed for trace"
      # Re-run with stderr visible so the upstream call_api capture sees WHY.
      cmd_refresh >/dev/null || true
      printf 'ensure: refresh failed (see above)\n' >&2
      return 1
    fi
    dbg "ensure: refresh OK — emitting new access_token"
    jqf "$OAUTH_FILE" -r '.access_token'
  else
    dbg "ensure: token still valid — emitting cached access_token"
    local tok; tok=$(jqf "$OAUTH_FILE" -r '.access_token // empty')
    if [ -z "$tok" ]; then
      dbg "ensure: .access_token is empty in the file"
      printf 'ensure: .access_token missing from %s\n' "$OAUTH_FILE" >&2
      return 1
    fi
    printf '%s' "$tok"
  fi
}

cmd_status() {
  if [ ! -f "$OAUTH_FILE" ]; then
    echo "OAuth: not authenticated (no $OAUTH_FILE)"
    return 1
  fi
  local fetched_at expires_in scope
  fetched_at=$(jqf "$OAUTH_FILE" -r '.fetched_at // 0')
  expires_in=$(jqf "$OAUTH_FILE" -r '.expires_in // 3600')
  scope=$(jqf "$OAUTH_FILE" -r '.scope // "(unknown)"')
  local now; now=$(date +%s)
  local expires_at=$((fetched_at + expires_in))
  local left=$((expires_at - now))
  printf 'OAuth status:\n'
  printf '  file:       %s\n' "$OAUTH_FILE"
  printf '  scope:      %s\n' "$scope"
  printf '  fetched_at: %s\n' "$(date -r "$fetched_at" 2>/dev/null || date -d "@$fetched_at" 2>/dev/null)"
  printf '  expires_in: %d s\n' "$expires_in"
  if [ "$left" -le 0 ]; then
    printf '  state:      EXPIRED (%ds ago) — will auto-refresh on next call\n' "$((-left))"
  else
    printf '  state:      valid for %d more seconds (~%d min)\n' "$left" "$((left/60))"
  fi
}

# cmd_debug — dump everything Mack would need to figure out why ensure is
# returning empty. SAFE to share: token previews are truncated to first 20
# chars only. Goes to stdout (so the user can copy-paste).
cmd_debug() {
  printf '=== Larry OAuth diagnostic ===\n'
  printf 'LARRY_HOME=%s\n' "$LARRY_HOME"
  printf 'OAUTH_FILE=%s\n' "$OAUTH_FILE"

  # jq binary identity — helps distinguish Cygwin jq from Windows-native jq
  # when paths look weird on MobaXterm.
  printf '\n[jq]\n'
  local jq_path; jq_path=$(command -v jq 2>/dev/null || true)
  printf '  path:    %s\n' "${jq_path:-(not found)}"
  if [ -n "$jq_path" ]; then
    printf '  version: %s\n' "$(jq --version 2>&1 | head -n1)"
    # On Cygwin, `file` may not exist but the path tells us if it's a .exe.
    case "$jq_path" in
      *.exe) printf '  flavor:  Windows-native (.exe — uses Windows paths in argv)\n' ;;
      *)     printf '  flavor:  Unix-style (POSIX paths in argv)\n' ;;
    esac
  fi

  # cygpath translation — only relevant on Cygwin/MobaXterm. If present, show
  # how the OAUTH_FILE looks to a Windows binary.
  if command -v cygpath >/dev/null 2>&1; then
    printf '\n[cygpath]\n'
    printf '  unix:    %s\n' "$OAUTH_FILE"
    printf '  windows: %s\n' "$(cygpath -w "$OAUTH_FILE" 2>&1 || echo '(cygpath failed)')"
  fi

  # File state
  printf '\n[file]\n'
  if [ ! -f "$OAUTH_FILE" ]; then
    printf '  status:  MISSING — run larry-auth.sh login\n'
    return 0
  fi
  # Mode, size, mtime — across mac (BSD stat) and linux (GNU stat)
  local stat_line
  stat_line=$(stat -f 'mode=%Sp size=%z mtime=%Sm' "$OAUTH_FILE" 2>/dev/null \
           || stat -c 'mode=%A size=%s mtime=%y' "$OAUTH_FILE" 2>/dev/null \
           || echo '(stat unavailable)')
  printf '  %s\n' "$stat_line"
  if ! jq -e . < "$OAUTH_FILE" >/dev/null 2>&1; then
    printf '  status:  CORRUPTED — not valid JSON\n'
    printf '  first 80 bytes (sanitized): '
    head -c 80 "$OAUTH_FILE" | tr -d '\n' | head -c 80
    printf '\n'
    return 0
  fi
  printf '  status:  parses as JSON\n'

  # Token math
  printf '\n[token math]\n'
  local fetched_at expires_in scope
  fetched_at=$(jqf "$OAUTH_FILE" -r '.fetched_at // 0')
  expires_in=$(jqf "$OAUTH_FILE" -r '.expires_in // 3600')
  scope=$(jqf "$OAUTH_FILE" -r '.scope // "(missing)"')
  local now; now=$(date +%s)
  local expires_at=$((fetched_at + expires_in))
  local left=$((expires_at - now))
  local refresh_at=$((expires_at - 300))
  local would_refresh="NO"
  [ "$now" -ge "$refresh_at" ] && would_refresh="YES (within 300s of expiry)"
  # Human-readable timestamps that work on both BSD and GNU date.
  local fetched_human expires_human now_human
  fetched_human=$(date -r "$fetched_at" 2>/dev/null || date -d "@$fetched_at" 2>/dev/null || echo "(date unavailable)")
  expires_human=$(date -r "$expires_at" 2>/dev/null || date -d "@$expires_at" 2>/dev/null || echo "(date unavailable)")
  now_human=$(date 2>/dev/null || echo "(date unavailable)")
  printf '  fetched_at:     %s (%s)\n' "$fetched_at" "$fetched_human"
  printf '  expires_in:     %s seconds\n' "$expires_in"
  printf '  expires_at:     %s (%s)\n' "$expires_at" "$expires_human"
  printf '  now:            %s (%s)\n' "$now" "$now_human"
  printf '  seconds_left:   %s (~%d min)\n' "$left" "$((left/60))"
  printf '  would_refresh:  %s\n' "$would_refresh"
  printf '  scope:          %s\n' "$scope"

  # Token previews — first 20 chars only, safe to share for "is it present"
  printf '\n[tokens (truncated)]\n'
  local at rt
  at=$(jqf "$OAUTH_FILE" -r '.access_token // empty')
  rt=$(jqf "$OAUTH_FILE" -r '.refresh_token // empty')
  if [ -z "$at" ]; then
    printf '  access_token:   (EMPTY/MISSING)\n'
  else
    printf '  access_token:   %s... (len=%d)\n' "$(printf '%s' "$at" | head -c 20)" "${#at}"
  fi
  if [ -z "$rt" ]; then
    printf '  refresh_token:  (EMPTY/MISSING)\n'
  else
    printf '  refresh_token:  %s... (len=%d)\n' "$(printf '%s' "$rt" | head -c 20)" "${#rt}"
  fi

  # Live ensure dry-run with stderr exposed so we see every decision.
  printf '\n[live ensure trace]\n'
  printf '  Running: LARRY_OAUTH_DEBUG=1 cmd_ensure (stderr below, stdout suppressed)\n'
  local trace_token trace_err trace_rc
  trace_err=$(mktemp)
  trace_token=$(LARRY_OAUTH_DEBUG=1 cmd_ensure 2>"$trace_err"); trace_rc=$?
  printf '  exit_code: %d\n' "$trace_rc"
  printf '  token_emitted: %s\n' "$([ -n "$trace_token" ] && echo "YES (len=${#trace_token})" || echo "NO (empty)")"
  printf '  stderr trace:\n'
  sed 's/^/    /' "$trace_err"
  rm -f "$trace_err"

  printf '\n=== end diagnostic ===\n'
}

cmd_logout() {
  if [ -f "$OAUTH_FILE" ]; then
    rm -f "$OAUTH_FILE"
    echo "logged out (removed $OAUTH_FILE)"
  else
    echo "no token file to remove"
  fi
}

case "${1:-status}" in
  login)   cmd_login ;;
  refresh) cmd_refresh ;;
  ensure)  cmd_ensure ;;
  status)  cmd_status ;;
  debug)   cmd_debug ;;
  logout)  cmd_logout ;;
  -h|--help|help) sed -n '2,30p' "$0" ;;
  *)       die "unknown subcommand: ${1:-} (try 'login|refresh|ensure|status|debug|logout')" ;;
esac