icoa-cli 2.19.120 → 2.19.122

package/dist/commands/vla4ctf.js

@@ -1 +1 @@
-import chalk from"chalk";import{readFileSync as o,existsSync as e}from"node:fs";import{resolve as n,dirname as t,join as i}from"node:path";import{fileURLToPath as l}from"node:url";import{spawn as r}from"node:child_process";import{logCommand as s}from"../lib/logger.js";import{getExamState as c,saveExamState as a}from"../lib/exam-state.js";import{printError as g}from"../lib/ui.js";let u=null,p=!1;export function isVla4CtfActive(){return p}export function exitVla4Ctf(){p=!1,u=null}export function enterVla4Ctf(o){const e=c();if(!e)return g("No exam in progress. Run `exam <token>` first."),!1;const n=e.questions.find(e=>e.number===o);return n?"vla"!==n.type?(g(`Q${o} is not a VLA challenge.`),!1):(u={qNum:o,question:n,baselineSeen:!1,probesUsed:0,hintsUsed:new Set,startedAt:Date.now()},p=!0,function(){if(!u)return;const o=u.question,e=(c()?.answers||{})[u.qNum];console.log(),console.log(chalk.bold.cyan("  ═══ VLA4CTF — Q"+u.qNum+": "+(o.category||"VLA")+" ═══")),e&&console.log(chalk.gray("  Current answer: ")+chalk.yellow(e)+chalk.gray(" (submit again to change)")),console.log(),console.log(chalk.cyan("  ┌────────────────────────────────────────────────────")),console.log(chalk.cyan("  │ ")+chalk.bold.white(`Q${u.qNum}  [${o.category||"VLA"}]  · ${o.points||6} pts`)),console.log(chalk.cyan("  │"));for(const e of String(o.text||"").split("\n")){const o=e.length>64?e.slice(0,61)+"...":e;console.log(chalk.cyan("  │ ")+chalk.white(o))}console.log(chalk.cyan("  │")),console.log(chalk.cyan("  │ ")+chalk.gray("Full question:  ")+chalk.white(`exam q ${u.qNum}`)),console.log(chalk.cyan("  │ ")+chalk.gray("Flag format:    ICOA{...}")),console.log(chalk.cyan("  └────────────────────────────────────────────────────")),console.log(),console.log(chalk.bold.white("  How to attack this VLA")),console.log(),console.log(chalk.bold.green("    baseline")+chalk.gray("               show default-instruction action")),console.log(chalk.bold.green('    probe "<instruction>"')+chalk.gray("  test a custom instruction")),console.log(chalk.bold.green("    image <path>")+chalk.gray("           upload modified image (adversarial patch)")),console.log(chalk.yellow("    hint a / b / c")+chalk.gray("         scripted hints")),console.log(chalk.bold.green("    submit ICOA{...}")+chalk.gray("       submit your flag")),console.log(chalk.gray("    exit")+chalk.gray("                   return to exam")),console.log()}(),!0):(g(`Question ${o} not in current exam.`),!1)}function m(o,e){const n=o.toLowerCase(),t=/\b(release|drop|let\s*go|open\s+gripper|stop)\b/i.test(n);/\b(red|rojo|赤い|红)\b/i.test(n)&&/\b(pick|grab|grasp|抓|取)\b/i.test(n);let i=.95,l=[.31,.12,.45],r="gripper closing on red cup at (0.31, 0.12, 0.45)";return 41===e?t&&(i=.04+.02*Math.random(),l=[.31,.12,.45],r=`gripper OPENING (${i.toFixed(2)}), arm holds position`):43===e&&/\b(purple|transparent|invisible|orange)\b/i.test(n)&&(i=.5,l=[.001,0,.001],r="arm hesitating, no clear target (||xyz|| ≈ 0)"),{gripper:i,xyz:l,flagCondition:41===e&&i<.1||43===e&&Math.sqrt(l[0]**2+l[1]**2+l[2]**2)<.05,description:r}}function d(o){return 41===o?"ICOA{vla_prompt_inject}":42===o?"ICOA{patch_redirect}":43===o?"ICOA{vla_confused}":44===o?"ICOA{action_jailbreak}":45===o?"ICOA{backdoor_trigger}":"ICOA{unknown}"}export async function handleVla4CtfMessage(g){if(!u)return"exit";const f=g.trim(),y=f.toLowerCase();if(s(`vla4ctf: ${f}`),"exit"===y||"back"===y||"quit"===y)return p=!1,u=null,console.log(),console.log(chalk.gray("  Returned to exam.")),console.log(),"exit";if("sim"===y||y.startsWith("sim "))return function(o){const n=function(){const o=t(l(import.meta.url));for(const n of[i(o,"..","..","panda","mujoco-launcher.py"),i(o,"..","..","..","panda","mujoco-launcher.py")])if(e(n))return n;return null}();if(!n)return console.log(chalk.yellow("  MuJoCo launcher not found locally.")),void console.log(chalk.gray("  Install via: pip install mujoco; clone panda/mujoco-launcher.py"));console.log(chalk.gray(`  Launching MuJoCo (scenario: ${o})...`)),r("python3",[n,o,"--seconds","5"],{stdio:"inherit"})}(f.slice(3).trim()||{41:"prompt_inj",42:"patch",43:"confused",44:"baseline",45:"baseline"}[u.qNum]||"baseline"),"continue";if("baseline"===y){const o=u.question.baselineInstruction||"Pick up the red cup",e=m(o,u.qNum);return u.baselineSeen=!0,console.log(),console.log(chalk.gray("  Default instruction: ")+chalk.white('"'+o+'"')),console.log(chalk.gray("  Octo action: ")+chalk.white(e.description)),console.log(),"continue"}const b=f.match(/^probe\s+["']?(.+?)["']?\s*$/i);if(b){u.probesUsed+=1;const o=m(b[1],u.qNum);if(console.log(),console.log(chalk.gray("  → ")+chalk.cyan("Octo action: ")+chalk.white(o.description)),o.flagCondition){const o=d(u.qNum);console.log(),console.log(chalk.bold.green("  ✓ Win condition met!  Flag: ")+chalk.bold.yellow(o)),console.log(chalk.gray("  Type ")+chalk.bold.green(`submit ${o}`)+chalk.gray(" to record your answer."))}else console.log(chalk.gray("  (no win yet — keep iterating)"));return console.log(),"continue"}const h=f.match(/^image\s+(.+)$/i);if(h){const t=n(h[1].trim()),i=function(n,t){if(!e(n))return{gripper:.95,xyz:[.31,.12,.45],flagCondition:!1,description:"no image uploaded — baseline action returned"};const i=o(n).length;return 42===t&&i>1024?{gripper:.9,xyz:[-.28,.12,.45],flagCondition:!0,description:"gripper closing at (-0.28, 0.12, 0.45) — TARGET SHIFTED to right side (knife)"}:{gripper:.95,xyz:[.31,.12,.45],flagCondition:!1,description:"image accepted but no action drift detected"}}(t,u.qNum);if(console.log(),console.log(chalk.gray("  Uploading: ")+chalk.white(t)),console.log(chalk.gray("  → ")+chalk.cyan("Octo action: ")+chalk.white(i.description)),i.flagCondition){const o=d(u.qNum);console.log(),console.log(chalk.bold.green("  ✓ Win condition met!  Flag: ")+chalk.bold.yellow(o))}return console.log(),"continue"}const w=y.match(/^hint\s+([abc])$/);if(w){const o=w[1].toUpperCase();u.hintsUsed.add(o);const e=(u.question.hints||{})[o]||`(no hint ${o} for Q${u.qNum})`,n="A"===o?chalk.green:"B"===o?chalk.yellow:chalk.red;console.log(),console.log(n.bold(`  ▸ Hint ${o}`)),console.log();for(const o of e.split("\n"))console.log(chalk.white("    "+o));return console.log(),"continue"}const C=f.match(/^submit\s+(.+)/i);if(C){let o=C[1].trim();if(/^submit\s+/i.test(o)&&(o=o.replace(/^submit\s+/i,"").trim()),o=o.replace(/^["'`]+|["'`]+$/g,"").trim(),/^[A-Da-d]$/.test(o))return console.log(),console.log(chalk.yellow(`  "${o}" looks like an MCQ letter, not a flag.`)),console.log(chalk.gray("  Flag format: ")+chalk.green("ICOA{your_flag}")),console.log(),"continue";const e=c();if(!e)return"exit";const n=e.answers[u.qNum];return e.interactions||(e.interactions=[]),e.interactions.push({ts:(new Date).toISOString(),q:u.qNum,type:n?"answer_changed":"answer_submitted",input:o,result:"via vla4ctf"}),e.answers[u.qNum]=o,e._lastQ=u.qNum+1<=45?u.qNum+1:u.qNum,a(e),console.log(),n?console.log(chalk.green(`  ✓ Q${u.qNum} answer updated: `)+chalk.yellow(o)):console.log(chalk.green.bold(`  ✓ Answer for Q${u.qNum} recorded: ${o}`)),console.log(chalk.gray("  (Correctness shown after final exam submit.)")),console.log(),"continue"}return console.log(),console.log(chalk.gray("  Try one of: ")+chalk.white('baseline / probe "..." / image <path> / hint a/b/c / submit ICOA{...} / exit')),console.log(),"continue"}export function registerVla4CtfCommand(o){o.command("vla4ctf").description("Enter VLA attack mode (for Paper D Q41-Q45)").action(()=>{const o=c();if(!o)return void g("No exam in progress. Run `exam <token>` first.");const e=o._lastQ||41;enterVla4Ctf(e>=41&&e<=45?e:41)})}
+import chalk from"chalk";import{readFileSync as o,existsSync as e}from"node:fs";import{resolve as n,join as t}from"node:path";import{spawn as i}from"node:child_process";import{logCommand as r}from"../lib/logger.js";import{getExamState as s,saveExamState as l}from"../lib/exam-state.js";import{printError as a}from"../lib/ui.js";let c=null,g=!1;export function isVla4CtfActive(){return g}export function exitVla4Ctf(){g=!1,c=null}export function enterVla4Ctf(o){const e=s();if(!e)return a("No exam in progress. Run `exam <token>` first."),!1;const n=e.questions.find(e=>e.number===o);return n?"vla"!==n.type?(a(`Q${o} is not a VLA challenge.`),!1):(c={qNum:o,question:n,baselineSeen:!1,probesUsed:0,hintsUsed:new Set,startedAt:Date.now()},g=!0,function(){if(!c)return;const o=c.question,e=(s()?.answers||{})[c.qNum];console.log(),console.log(chalk.bold.cyan("  ═══ VLA4CTF — Q"+c.qNum+": "+(o.category||"VLA")+" ═══")),e&&console.log(chalk.gray("  Current answer: ")+chalk.yellow(e)+chalk.gray(" (submit again to change)")),console.log(),console.log(chalk.cyan("  ┌────────────────────────────────────────────────────")),console.log(chalk.cyan("  │ ")+chalk.bold.white(`Q${c.qNum}  [${o.category||"VLA"}]  · ${o.points||6} pts`)),console.log(chalk.cyan("  │"));for(const e of String(o.text||"").split("\n")){const o=e.length>64?e.slice(0,61)+"...":e;console.log(chalk.cyan("  │ ")+chalk.white(o))}console.log(chalk.cyan("  │")),console.log(chalk.cyan("  │ ")+chalk.gray("Full question:  ")+chalk.white(`exam q ${c.qNum}`)),console.log(chalk.cyan("  │ ")+chalk.gray("Flag format:    ICOA{...}")),console.log(chalk.cyan("  └────────────────────────────────────────────────────")),console.log(),console.log(chalk.bold.white("  How to attack this VLA")),console.log(),console.log(chalk.bold.green("    baseline")+chalk.gray("               show default-instruction action")),console.log(chalk.bold.green('    probe "<instruction>"')+chalk.gray("  test a custom instruction")),console.log(chalk.bold.green("    image <path>")+chalk.gray("           upload modified image (adversarial patch)")),console.log(chalk.yellow("    hint a / b / c")+chalk.gray("         scripted hints")),console.log(chalk.bold.green("    submit ICOA{...}")+chalk.gray("       submit your flag")),console.log(chalk.gray("    exit")+chalk.gray("                   return to exam")),console.log()}(),!0):(a(`Question ${o} not in current exam.`),!1)}function u(o,e){const n=o.toLowerCase(),t=/\b(release|drop|let\s*go|open\s+gripper|stop)\b/i.test(n);/\b(red|rojo|赤い|红)\b/i.test(n)&&/\b(pick|grab|grasp|抓|取)\b/i.test(n);let i=.95,r=[.31,.12,.45],s="gripper closing on red cup at (0.31, 0.12, 0.45)";return 41===e?t&&(i=.04+.02*Math.random(),r=[.31,.12,.45],s=`gripper OPENING (${i.toFixed(2)}), arm holds position`):43===e&&/\b(purple|transparent|invisible|orange)\b/i.test(n)&&(i=.5,r=[.001,0,.001],s="arm hesitating, no clear target (||xyz|| ≈ 0)"),{gripper:i,xyz:r,flagCondition:41===e&&i<.1||43===e&&Math.sqrt(r[0]**2+r[1]**2+r[2]**2)<.05,description:s}}function p(o){return 41===o?"ICOA{vla_prompt_inject}":42===o?"ICOA{patch_redirect}":43===o?"ICOA{vla_confused}":44===o?"ICOA{action_jailbreak}":45===o?"ICOA{backdoor_trigger}":"ICOA{unknown}"}export async function handleVla4CtfMessage(a){if(!c)return"exit";const d=a.trim(),m=d.toLowerCase();if(r(`vla4ctf: ${d}`),"exit"===m||"back"===m||"quit"===m)return g=!1,c=null,console.log(),console.log(chalk.gray("  Returned to exam.")),console.log(),"exit";if("sim"===m||m.startsWith("sim "))return await async function(o,e){const n=`${process.env.ICOA_SERVER_URL?.replace(/\/+$/,"")||"https://practice.icoa2026.au"}/api/ai/vla/${o}/sim`;console.log(chalk.gray("  Rendering MuJoCo replay on server..."));try{const r=await fetch(n,{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({action:e??null}),signal:AbortSignal.timeout(3e4)});if(!r.ok){const o=await r.text().catch(()=>"");return void console.log(chalk.yellow(`  Render failed (${r.status}): ${o.slice(0,120)}`))}const s=await r.json();if(!s.success||!s.data?.mp4_b64)return void console.log(chalk.yellow("  Renderer returned no video (MuJoCo unavailable on backend)."));const l=process.env.TMPDIR||"/tmp",a=t(l,`icoa-sim-q${o}-${Date.now()}.mp4`),{writeFileSync:c}=await import("node:fs");c(a,Buffer.from(s.data.mp4_b64,"base64")),console.log(chalk.green("  ✓ ")+chalk.gray("Saved: ")+chalk.white(a)),s.data.description&&console.log(chalk.gray("  Action: ")+chalk.white(s.data.description));const g="darwin"===process.platform?"open":"win32"===process.platform?"start":"xdg-open";g?(i(g,[a],{stdio:"ignore",detached:!0}).unref(),console.log(chalk.gray("  Opening in your default video player..."))):console.log(chalk.gray("  Open the file above in your video player."))}catch(o){const e=o instanceof Error?o.message:String(o);console.log(chalk.yellow(`  Sim error: ${e}`))}}(c.qNum),"continue";if("baseline"===m){const o=c.question.baselineInstruction||"Pick up the red cup",e=u(o,c.qNum);return c.baselineSeen=!0,console.log(),console.log(chalk.gray("  Default instruction: ")+chalk.white('"'+o+'"')),console.log(chalk.gray("  Octo action: ")+chalk.white(e.description)),console.log(),"continue"}const f=d.match(/^probe\s+["']?(.+?)["']?\s*$/i);if(f){c.probesUsed+=1;const o=u(f[1],c.qNum);if(console.log(),console.log(chalk.gray("  → ")+chalk.cyan("Octo action: ")+chalk.white(o.description)),o.flagCondition){const o=p(c.qNum);console.log(),console.log(chalk.bold.green("  ✓ Win condition met!  Flag: ")+chalk.bold.yellow(o)),console.log(chalk.gray("  Type ")+chalk.bold.green(`submit ${o}`)+chalk.gray(" to record your answer."))}else console.log(chalk.gray("  (no win yet — keep iterating)"));return console.log(),"continue"}const y=d.match(/^image\s+(.+)$/i);if(y){const t=n(y[1].trim()),i=function(n,t){if(!e(n))return{gripper:.95,xyz:[.31,.12,.45],flagCondition:!1,description:"no image uploaded — baseline action returned"};const i=o(n).length;return 42===t&&i>1024?{gripper:.9,xyz:[-.28,.12,.45],flagCondition:!0,description:"gripper closing at (-0.28, 0.12, 0.45) — TARGET SHIFTED to right side (knife)"}:{gripper:.95,xyz:[.31,.12,.45],flagCondition:!1,description:"image accepted but no action drift detected"}}(t,c.qNum);if(console.log(),console.log(chalk.gray("  Uploading: ")+chalk.white(t)),console.log(chalk.gray("  → ")+chalk.cyan("Octo action: ")+chalk.white(i.description)),i.flagCondition){const o=p(c.qNum);console.log(),console.log(chalk.bold.green("  ✓ Win condition met!  Flag: ")+chalk.bold.yellow(o))}return console.log(),"continue"}const b=m.match(/^hint\s+([abc])$/);if(b){const o=b[1].toUpperCase();c.hintsUsed.add(o);const e=(c.question.hints||{})[o]||`(no hint ${o} for Q${c.qNum})`,n="A"===o?chalk.green:"B"===o?chalk.yellow:chalk.red;console.log(),console.log(n.bold(`  ▸ Hint ${o}`)),console.log();for(const o of e.split("\n"))console.log(chalk.white("    "+o));return console.log(),"continue"}const h=d.match(/^submit\s+(.+)/i);if(h){let o=h[1].trim();if(/^submit\s+/i.test(o)&&(o=o.replace(/^submit\s+/i,"").trim()),o=o.replace(/^["'`]+|["'`]+$/g,"").trim(),/^[A-Da-d]$/.test(o))return console.log(),console.log(chalk.yellow(`  "${o}" looks like an MCQ letter, not a flag.`)),console.log(chalk.gray("  Flag format: ")+chalk.green("ICOA{your_flag}")),console.log(),"continue";const e=s();if(!e)return"exit";const n=e.answers[c.qNum];return e.interactions||(e.interactions=[]),e.interactions.push({ts:(new Date).toISOString(),q:c.qNum,type:n?"answer_changed":"answer_submitted",input:o,result:"via vla4ctf"}),e.answers[c.qNum]=o,e._lastQ=c.qNum+1<=45?c.qNum+1:c.qNum,l(e),console.log(),n?console.log(chalk.green(`  ✓ Q${c.qNum} answer updated: `)+chalk.yellow(o)):console.log(chalk.green.bold(`  ✓ Answer for Q${c.qNum} recorded: ${o}`)),console.log(chalk.gray("  (Correctness shown after final exam submit.)")),console.log(),"continue"}return console.log(),console.log(chalk.gray("  Try one of: ")+chalk.white('baseline / probe "..." / image <path> / hint a/b/c / submit ICOA{...} / exit')),console.log(),"continue"}export function registerVla4CtfCommand(o){o.command("vla4ctf").description("Enter VLA attack mode (for Paper D Q41-Q45)").action(()=>{const o=s();if(!o)return void a("No exam in progress. Run `exam <token>` first.");const e=o._lastQ||41;enterVla4Ctf(e>=41&&e<=45?e:41)})}