@tangle-network/agent-bench 0.7.1 → 0.8.0

package/src/trata-gate.mts

@@ -30,6 +30,7 @@
 import { appendFileSync } from 'node:fs'
 import { resolveAdapter } from './adapters'
 import type { BenchScore, BenchTask } from './benchmarks/types'
+import { runBenchRouterTurn } from './router-turn'
 import { runPool } from './run-pool'
 
 function must(name: string): string {
@@ -61,33 +62,32 @@ async function workerComplete(
   cfg: { routerBaseUrl: string; routerKey: string; model: string; timeoutMs: number },
 ): Promise<{ answer: string; inputTokens: number; outputTokens: number; durationMs: number }> {
   const startedAt = Date.now()
-  const res = await fetch(`${cfg.routerBaseUrl}/chat/completions`, {
-    method: 'POST',
-    signal: AbortSignal.timeout(cfg.timeoutMs),
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${cfg.routerKey}` },
-    body: JSON.stringify({
-      model: cfg.model,
-      temperature: 0,
-      max_tokens: 4096,
-      messages: [
-        { role: 'system', content: ANALYST_SYSTEM },
-        { role: 'user', content: task.prompt },
-      ],
-    }),
-  })
-  if (!res.ok) {
-    const body = (await res.text()).slice(0, 300)
-    throw new Error(`router ${res.status} for ${task.id}: ${body}`)
-  }
-  const j = (await res.json()) as {
-    choices?: Array<{ message?: { content?: string } }>
-    usage?: { prompt_tokens?: number; completion_tokens?: number }
-  }
-  const answer = j.choices?.[0]?.message?.content ?? ''
+  const result = await runBenchRouterTurn(
+    {
+      routerBaseUrl: cfg.routerBaseUrl,
+      routerKey: cfg.routerKey,
+      profile: {
+        name: 'trata-financial-analyst',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: cfg.model,
+          metadata: {
+            temperature: 0,
+            maxTokens: Number(process.env.WORKER_MAX_TOKENS ?? 4096),
+          },
+        },
+        prompt: { systemPrompt: ANALYST_SYSTEM },
+      },
+      timeoutMs: cfg.timeoutMs,
+    },
+    task.prompt,
+  )
+  if (result.usage.tokensKnown === false) throw new Error('worker provider omitted token usage')
   return {
-    answer,
-    inputTokens: j.usage?.prompt_tokens ?? 0,
-    outputTokens: j.usage?.completion_tokens ?? 0,
+    answer: result.finalText,
+    inputTokens: result.usage.input,
+    outputTokens: result.usage.output,
     durationMs: Date.now() - startedAt,
   }
 }

package/src/trata-gepa.mts

@@ -60,6 +60,7 @@ import {
   officialOptimizerModel,
   requiredTokenPricing,
 } from './official-optimizer-config.mjs'
+import { runBenchRouterTurn } from './router-turn'
 
 interface TrataScenario extends Scenario {
   task: BenchTask
@@ -122,23 +123,31 @@ async function chatComplete(
   messages: Array<{ role: string; content: string }>,
   signal: AbortSignal,
 ): Promise<{ content: string; usage?: { input: number; output: number } }> {
-  const res = await fetch(`${baseUrl}/chat/completions`, {
-    method: 'POST',
-    signal: AbortSignal.any([signal, AbortSignal.timeout(180_000)]),
-    headers: { 'content-type': 'application/json', authorization: `Bearer ${key}` },
-    body: JSON.stringify({ model, temperature: 0, max_tokens: maxTokens, messages }),
-  })
-  if (!res.ok) throw new Error(`router ${res.status}: ${(await res.text()).slice(0, 300)}`)
-  const j = (await res.json()) as {
-    choices?: Array<{ message?: { content?: string } }>
-    usage?: { prompt_tokens?: number; completion_tokens?: number }
+  const system = messages.find((message) => message.role === 'system')?.content
+  const result = await runBenchRouterTurn(
+    {
+      routerBaseUrl: baseUrl,
+      routerKey: key,
+      profile: {
+        name: 'trata-gepa-worker',
+        harness: 'cli-base',
+        model: {
+          provider: 'tangle-router',
+          default: model,
+          metadata: { temperature: 0, maxTokens },
+        },
+        ...(system ? { prompt: { systemPrompt: system } } : {}),
+      },
+      signal,
+    },
+    { messages: messages.filter((message) => message.role !== 'system') },
+  )
+  return {
+    content: result.finalText,
+    ...(result.usage.tokensKnown === false
+      ? {}
+      : { usage: { input: result.usage.input, output: result.usage.output } }),
   }
-  const content = j.choices?.[0]?.message?.content ?? ''
-  const usage =
-    j.usage?.prompt_tokens != null
-      ? { input: j.usage.prompt_tokens, output: j.usage.completion_tokens ?? 0 }
-      : undefined
-  return { content, usage }
 }
 
 async function main(): Promise<void> {

package/src/trata-hedge-solve.mts

@@ -0,0 +1,80 @@
+/** Single-shot lower-bound solver for trata-hedge-bench. */
+import { readdirSync, readFileSync, statSync, writeFileSync } from 'node:fs'
+import { join, relative, resolve } from 'node:path'
+import { runBenchRouterTurn } from './router-turn'
+
+const environment = process.argv[2]
+if (!environment) throw new Error('usage: solve.mts <env-dir> [out.txt]')
+const output = process.argv[3] ?? '/tmp/thb-answer.txt'
+const model = process.env.WORKER_MODEL
+if (!model) throw new Error('WORKER_MODEL is required')
+const dataBudget = Number(process.env.DATA_BUDGET ?? 160_000)
+const maxTokens = Number(process.env.MAX_TOKENS ?? 6_000)
+const temperature = Number(process.env.TEMPERATURE ?? 0.5)
+const routerKey = process.env.TANGLE_API_KEY
+if (!routerKey) throw new Error('TANGLE_API_KEY is required')
+const routerBaseUrl = process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1'
+
+for (const [name, value] of [
+  ['DATA_BUDGET', dataBudget],
+  ['MAX_TOKENS', maxTokens],
+] as const) {
+  if (!Number.isSafeInteger(value) || value <= 0) throw new Error(`${name} must be a positive integer`)
+}
+if (!Number.isFinite(temperature)) throw new Error('TEMPERATURE must be finite')
+
+const root = resolve(environment)
+const dataDir = join(root, 'environment', 'data')
+const instruction = readFileSync(join(root, 'instruction.md'), 'utf8')
+
+function filesUnder(dir: string): string[] {
+  return readdirSync(dir)
+    .flatMap((name) => {
+      const path = join(dir, name)
+      return statSync(path).isDirectory() ? filesUnder(path) : [path]
+    })
+}
+
+function rank(path: string): number {
+  if (path.includes('earnings_call')) return 0
+  if (path.includes('financials')) return 1
+  if (path.includes('company_profiles')) return 2
+  if (path.includes('press_releases')) return 3
+  return 4
+}
+
+let used = 0
+const blocks: string[] = []
+const files = filesUnder(dataDir).sort((a, b) => rank(a) - rank(b) || a.localeCompare(b))
+for (const path of files) {
+  const content = readFileSync(path, 'utf8')
+  if (used + content.length > dataBudget) continue
+  used += content.length
+  blocks.push(`\n=== FILE: data/${relative(dataDir, path)} ===\n${content}`)
+}
+process.stderr.write(`[solve] ${blocks.length}/${files.length} files in context (${used} chars)\n`)
+
+const prompt =
+  '--- AVAILABLE DATA (cite files by their `data/<path>` name inline) ---\n' +
+  blocks.join('') +
+  '\n\n--- END DATA ---\nWrite ONLY the full analysis (no preamble). Inline-cite every claim with its `data/<path>`.'
+
+const result = await runBenchRouterTurn(
+  {
+    routerBaseUrl,
+    routerKey,
+    profile: {
+      name: 'trata-hedge-solver',
+      harness: 'cli-base',
+      model: {
+        provider: 'tangle-router',
+        default: model,
+        metadata: { temperature, maxTokens },
+      },
+      prompt: { systemPrompt: instruction },
+    },
+  },
+  prompt,
+)
+writeFileSync(output, result.finalText)
+process.stderr.write(`[solve] wrote ${result.finalText.length} chars -> ${output} (model=${model})\n`)

package/src/worker-blender.ts

@@ -18,7 +18,7 @@ import type { Span } from '@tangle-network/agent-eval'
 import type { BenchTask } from './benchmarks/types'
 import { DEFAULT_BLENDER_DIRECTIVE } from './directives'
 import { runRefineLoop } from './refine-loop'
-import { routerChatWithUsage } from '@tangle-network/agent-runtime/kernel'
+import { runBenchRouterTurn } from './router-turn'
 
 export { DEFAULT_BLENDER_DIRECTIVE } from './directives'
 
@@ -177,17 +177,27 @@ export async function solveBlenderLocal(task: BenchTask, cfg: BlenderLocalConfig
     runShot: async (user, round, dir) => {
       const runnerPath = join(dir, 'runner.py')
       const scriptPath = join(dir, 'model.py')
-      const { content, usage: u } = await routerChatWithUsage(
-        cfg,
-        [
-          { role: 'system', content: directive },
-          { role: 'user', content: user },
-        ],
-        { temperature: 0.3 },
+      const turn = await runBenchRouterTurn(
+        {
+          routerBaseUrl: cfg.routerBaseUrl,
+          routerKey: cfg.routerKey,
+          profile: {
+            name: 'blender-worker',
+            harness: 'cli-base',
+            model: {
+              provider: 'tangle-router',
+              default: cfg.model,
+              metadata: { temperature: 0.3 },
+            },
+            prompt: { systemPrompt: directive },
+          },
+        },
+        user,
       )
-      if (u) {
-        usage.input += u.input
-        usage.output += u.output
+      const content = turn.finalText
+      if (turn.usage.tokensKnown !== false) {
+        usage.input += turn.usage.input
+        usage.output += turn.usage.output
       }
       const script = extractPy(content)
       trace.push({ spanId: `s-author-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: content.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)

package/src/worker-browser.ts

@@ -18,7 +18,7 @@
 import { readFile } from 'node:fs/promises'
 import type { Span } from '@tangle-network/agent-eval'
 import type { BenchTask } from './benchmarks/types'
-import { routerChatWithUsage } from '@tangle-network/agent-runtime/kernel'
+import { runBenchRouterTurn } from './router-turn'
 
 export interface BrowserLocalConfig {
   routerBaseUrl: string
@@ -67,10 +67,23 @@ export async function solveBrowserLocal(task: BenchTask, cfg: BrowserLocalConfig
 
   trace.push({ spanId: 's-task', runId, kind: 'llm', name: 'web task', model: cfg.model, messages: [{ role: 'user', content: goal }], startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)
 
-  const { content, usage } = await routerChatWithUsage(cfg, [
-    { role: 'system', content: directive },
-    { role: 'user', content: task.prompt },
-  ])
+  const turn = await runBenchRouterTurn(
+    {
+      routerBaseUrl: cfg.routerBaseUrl,
+      routerKey: cfg.routerKey,
+      profile: {
+        name: 'browser-local-worker',
+        model: { provider: 'tangle-router', default: cfg.model },
+        prompt: { systemPrompt: directive },
+      },
+    },
+    task.prompt,
+  )
+  const content = turn.finalText
+  const usage =
+    turn.usage.tokensKnown === false
+      ? undefined
+      : { input: turn.usage.input, output: turn.usage.output }
   const artifact = content.trim()
 
   const elementId = /ELEMENT:\s*\[?(\d+)\]?/i.exec(artifact)?.[1] ?? ''

package/src/worker-build123d.ts

@@ -21,7 +21,7 @@ import type { Span } from '@tangle-network/agent-eval'
 import type { BenchTask } from './benchmarks/types'
 import { DEFAULT_BUILD123D_DIRECTIVE } from './directives'
 import { runRefineLoop } from './refine-loop'
-import { routerChatWithUsage } from '@tangle-network/agent-runtime/kernel'
+import { runBenchRouterTurn } from './router-turn'
 
 export { DEFAULT_BUILD123D_DIRECTIVE } from './directives'
 
@@ -106,13 +106,22 @@ export async function solveBuild123dLocal(task: BenchTask, cfg: Build123dConfig)
     runShot: async (user, round, dir) => {
       const scriptPath = join(dir, 'build.py')
       const stepPath = join(dir, 'output.step')
-      const { content, usage: u } = await routerChatWithUsage(cfg, [
-        { role: 'system', content: sys },
-        { role: 'user', content: user },
-      ])
-      if (u) {
-        usage.input += u.input
-        usage.output += u.output
+      const turn = await runBenchRouterTurn(
+        {
+          routerBaseUrl: cfg.routerBaseUrl,
+          routerKey: cfg.routerKey,
+          profile: {
+            name: 'build123d-worker',
+            model: { provider: 'tangle-router', default: cfg.model },
+            prompt: { systemPrompt: sys },
+          },
+        },
+        user,
+      )
+      const content = turn.finalText
+      if (turn.usage.tokensKnown !== false) {
+        usage.input += turn.usage.input
+        usage.output += turn.usage.output
       }
       const source = extractPy(content)
       trace.push({ spanId: `s-author-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: content.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)

package/src/worker-cad.ts

@@ -17,12 +17,13 @@ import { mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'
 import { tmpdir } from 'node:os'
 import { join } from 'node:path'
 import { promisify } from 'node:util'
-import { acquireSandbox, routerChatWithUsage } from '@tangle-network/agent-runtime/kernel'
+import { acquireSandbox } from '@tangle-network/agent-runtime/kernel'
 import { Sandbox } from '@tangle-network/sandbox'
 import type { Span } from '@tangle-network/agent-eval'
 import type { BenchTask } from './benchmarks/types'
 import { DEFAULT_CAD_DIRECTIVE, DEFAULT_CAD_SANDBOX_DIRECTIVE } from './directives'
 import { runRefineLoop } from './refine-loop'
+import { runBenchRouterTurn } from './router-turn'
 
 export { DEFAULT_CAD_DIRECTIVE } from './directives'
 
@@ -138,13 +139,22 @@ export async function solveCadRefineLocal(task: BenchTask, cfg: CadLocalConfig):
       const scadPath = join(dir, 'model.scad')
       const stlPath = join(dir, 'model.stl')
       const pngPath = join(dir, 'model.png')
-      const { content, usage: u } = await routerChatWithUsage(cfg, [
-        { role: 'system', content: directive },
-        { role: 'user', content: user },
-      ])
-      if (u) {
-        usage.input += u.input
-        usage.output += u.output
+      const turn = await runBenchRouterTurn(
+        {
+          routerBaseUrl: cfg.routerBaseUrl,
+          routerKey: cfg.routerKey,
+          profile: {
+            name: 'cad-local-worker',
+            model: { provider: 'tangle-router', default: cfg.model },
+            prompt: { systemPrompt: directive },
+          },
+        },
+        user,
+      )
+      const content = turn.finalText
+      if (turn.usage.tokensKnown !== false) {
+        usage.input += turn.usage.input
+        usage.output += turn.usage.output
       }
       const scad = extractScad(content)
       trace.push({ spanId: `s-reply-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: content.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)
@@ -253,10 +263,19 @@ export async function solveCadRefine(task: BenchTask, cfg: CadRefineConfig): Pro
         ? task.prompt
         : `Your previous OpenSCAD had this problem:\n${lastErr}\n\nHere is the previous source:\n${history[history.length - 1]?.artifact ?? ''}\n\nFix it so it compiles AND better matches the brief:\n${task.prompt}`,
     runShot: async (user, round, box) => {
-      const { content: reply } = await routerChatWithUsage(cfg, [
-        { role: 'system', content: sys },
-        { role: 'user', content: user },
-      ])
+      const turn = await runBenchRouterTurn(
+        {
+          routerBaseUrl: cfg.routerBaseUrl,
+          routerKey: cfg.routerKey,
+          profile: {
+            name: 'cad-sandbox-worker',
+            model: { provider: 'tangle-router', default: cfg.model },
+            prompt: { systemPrompt: sys },
+          },
+        },
+        user,
+      )
+      const reply = turn.finalText
       const scad = extractScad(reply)
       trace.push({ spanId: `s-reply-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: reply.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)
       trace.push({ spanId: `s-write-${round}`, runId, kind: 'tool', name: 'write_file', toolName: 'create_file', args: { path: 'model.scad', content: scad }, startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)

package/src/worker.ts

@@ -89,7 +89,7 @@ export async function solveShot(
 
   // Cold-start-resilient via the shared lineage layer (a gateway-timed-out create is
   // recovered by name lookup). The inline profile + backend override is the same
-  // generic AgentRunSpec the runLoop kernel boots against the real sandbox.
+  // generic AgentRunSpec the runAgentRounds kernel boots against the real sandbox.
   const controller = new AbortController()
   const timer = cfg.timeoutMs ? setTimeout(() => controller.abort(), cfg.timeoutMs) : undefined
   const agentRun: AgentRunSpec<string> = {

package/tb_agents/opencode_router_agent.py

@@ -66,6 +66,15 @@ _USAGE_OUT_PATH = f"{_CONTAINER_DIR}/opencode-events.jsonl"
 _USAGE_ERR_PATH = f"{_CONTAINER_DIR}/opencode-run.err"
 
 
+REFINE_DIRECTIVE = (
+    "REFINE PASS — your previous attempt failed the task's automated tests. "
+    "Use the evidence below to correct the concrete failure without discarding work "
+    "that was already right. Verify the task's success criteria before finishing.\n\n"
+    "<previous_attempt>\n{prior}\n</previous_attempt>\n\n"
+    "--- ORIGINAL TASK ---\n{instruction}"
+)
+
+
 class OpenCodeRouterAgent(OpenCodeAgent):
     """`OpenCodeAgent` + Tangle-router routing + real token metering.
 
@@ -77,27 +86,71 @@ class OpenCodeRouterAgent(OpenCodeAgent):
     a real `AgentResult`.
     """
 
-    def __init__(self, model_name: str, *args, **kwargs):
-        # The stock __init__ does `self._provider, _ = model_name.split("/")`, which
-        # raises on a 3-segment or bare id. Derive the provider robustly ourselves,
-        # then hand the base class a value it can split without raising.
-        self._model_name = model_name
-        provider, _, sub = model_name.partition("/")
+    def __init__(
+        self,
+        model_name: str,
+        profile_path: str,
+        prior_attempt_hex: str | None = None,
+        *args,
+        **kwargs,
+    ):
+        self._profile = json.loads(Path(profile_path).read_text())
+        if self._profile.get("harness") != "opencode":
+            raise ValueError("AgentProfile.harness must be 'opencode'")
+        model = self._profile.get("model")
+        if not isinstance(model, dict):
+            raise ValueError("AgentProfile.model is required")
+        provider = model.get("provider")
+        declared = model.get("default")
+        if not isinstance(provider, str) or not provider.strip():
+            raise ValueError("AgentProfile.model.provider must be explicit")
+        if not isinstance(declared, str) or not declared.strip() or declared == "runtime-selected":
+            raise ValueError("AgentProfile.model.default must be concrete")
+        wire_model = declared if "/" in declared else f"{provider}/{declared}"
+        if model_name != wire_model:
+            raise ValueError(
+                f"Terminal-Bench model {model_name!r} conflicts with AgentProfile {wire_model!r}"
+            )
+        self._model_name = wire_model
         self._provider = provider
-        # Model id opencode sees, minus the provider segment (e.g. "glm-5.2").
-        self._provider_model = sub or provider
+        self._provider_model = declared.split("/", 1)[-1]
+        self._prior_attempt = self._decode_prior(prior_attempt_hex)
+        prompt = self._profile.get("prompt")
+        self._profile_instructions: list[str] = []
+        if isinstance(prompt, dict):
+            system = prompt.get("systemPrompt")
+            if isinstance(system, str) and system.strip():
+                self._profile_instructions.append(system.strip())
+            instructions = prompt.get("instructions")
+            if isinstance(instructions, list):
+                self._profile_instructions.extend(
+                    value.strip()
+                    for value in instructions
+                    if isinstance(value, str) and value.strip()
+                )
         self._router_base_url = (
             os.environ.get("OPENAI_BASE_URL")
             or os.environ.get("ROUTER_BASE_URL")
             or DEFAULT_ROUTER_BASE_URL
         )
         self._router_api_key = os.environ.get("OPENAI_API_KEY", "")
-        # Call the grandparent (AbstractInstalledAgent) init via OpenCodeAgent, but
-        # avoid the base split-crash by not delegating provider derivation to it.
-        super(OpenCodeAgent, self).__init__(*args, **kwargs)
-        self._version = kwargs.get("version", "latest")
+        super().__init__(wire_model, *args, **kwargs)
         self._logger = logger.getChild(__name__)
 
+    @staticmethod
+    def _decode_prior(value: str | None) -> str:
+        if not value:
+            return ""
+        raw = str(value).strip()
+        if raw.startswith("h"):
+            raw = raw[1:]
+        if not raw:
+            return ""
+        try:
+            return bytes.fromhex(raw).decode("utf-8").strip()
+        except ValueError as error:
+            raise ValueError("prior_attempt_hex must be 'h' + hex-encoded UTF-8") from error
+
     @staticmethod
     def name() -> str:
         return "agent-runtime-opencode-router"
@@ -136,16 +189,6 @@ class OpenCodeRouterAgent(OpenCodeAgent):
             "OPENAI_BASE_URL": self._router_base_url,
             "OPENCODE_CONFIG": _CONFIG_PATH,
         }
-        # Pass through any provider-native key the user already exported, so the same
-        # agent also works when pointed at a vendor endpoint rather than the router.
-        for passthrough in (
-            "ANTHROPIC_API_KEY",
-            "DEEPSEEK_API_KEY",
-            "GROQ_API_KEY",
-            "ZAI_API_KEY",
-        ):
-            if passthrough in os.environ:
-                env.setdefault(passthrough, os.environ[passthrough])
         return {k: v for k, v in env.items() if v}
 
     def _build_router_config(self) -> str:
@@ -159,14 +202,7 @@ class OpenCodeRouterAgent(OpenCodeAgent):
             "$schema": "https://opencode.ai/config.json",
             # Headless benchmark runs cannot answer interactive permission prompts.
             # Keep this identical for raw and supervisor arms.
-            "permission": {
-                "edit": "allow",
-                "bash": "allow",
-                "webfetch": "allow",
-                "read": "allow",
-                "write": "allow",
-                "external_directory": "allow",
-            },
+            "permission": self._profile.get("permission", {}),
             "provider": {
                 self._provider: {
                     "npm": "@ai-sdk/openai-compatible",
@@ -184,25 +220,24 @@ class OpenCodeRouterAgent(OpenCodeAgent):
         return json.dumps(config)
 
     def _run_agent_commands(self, instruction: str):
-        """Same command as the stock agent, but `--format json` with the event
-        stream captured to a file so we can meter token usage after the run."""
+        """Instrument Terminal-Bench's stock command without selecting a model here."""
         from terminal_bench.terminal.models import TerminalCommand
 
-        escaped_instruction = shlex.quote(instruction)
-        command = (
-            f"opencode --model {self._model_name} --format json "
-            f"run {escaped_instruction} "
-            f"> {_USAGE_OUT_PATH} 2> {_USAGE_ERR_PATH}"
-        )
-        return [
-            TerminalCommand(
+        instrumented = []
+        for base in super()._run_agent_commands(instruction):
+            marker = " run "
+            if marker not in base.command:
+                raise ValueError("stock OpenCodeAgent command has no run subcommand")
+            command = base.command.replace(marker, " --format json run ", 1)
+            command += f" > {_USAGE_OUT_PATH} 2> {_USAGE_ERR_PATH}"
+            instrumented.append(TerminalCommand(
                 command=command,
                 min_timeout_sec=0.0,
                 max_timeout_sec=float("inf"),
                 block=True,
                 append_enter=True,
-            )
-        ]
+            ))
+        return instrumented
 
     def _write_config_to_container(self, session: TmuxSession) -> None:
         config_json = self._build_router_config()
@@ -365,6 +400,13 @@ class OpenCodeRouterAgent(OpenCodeAgent):
                 "OPENAI_API_KEY (router key) is not set; opencode has no credential "
                 "to reach the router."
             )
+        if self._profile_instructions:
+            instruction = "\n\n".join([*self._profile_instructions, instruction])
+        if self._prior_attempt:
+            instruction = REFINE_DIRECTIVE.format(
+                prior=self._prior_attempt,
+                instruction=instruction,
+            )
         # Publish container id + inject router config BEFORE the base flow runs the
         # agent. The base perform_task copies the install script into _CONTAINER_DIR,
         # sources setup-env.sh (which exports OPENCODE_CONFIG), installs opencode, and

package/dist/cadbench-BrpwOU6A.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"cadbench-BrpwOU6A.js","names":[],"sources":["../src/worker-blender.ts","../src/benchmarks/cadbench.ts"],"sourcesContent":["/**\n * BlenderLLM / CADBench worker. The deliverable for a CADBench task is a Blender\n * `bpy` Python script that builds the described 3D model. We author it via the\n * router, execute it headless in Blender (Cycles CPU, no GPU), auto-frame the\n * produced geometry, and render N standardized views — the images the CADBench\n * criteria judge scores. The authoring directive is the GEPA-optimizable surface.\n *\n * Requires `blender` + `xvfb-run` on PATH (apt blender 4.x). No GPU: Cycles CPU\n * with denoising off (the apt build ships without OpenImageDenoise).\n */\n\nimport { execFile } from 'node:child_process'\nimport { mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { promisify } from 'node:util'\nimport type { Span } from '@tangle-network/agent-eval'\nimport type { BenchTask } from './benchmarks/types'\nimport { DEFAULT_BLENDER_DIRECTIVE } from './directives'\nimport { runRefineLoop } from './refine-loop'\nimport { routerChatWithUsage } from '@tangle-network/agent-runtime/kernel'\n\nexport { DEFAULT_BLENDER_DIRECTIVE } from './directives'\n\nconst execFileAsync = promisify(execFile)\n\nasync function runLocal(cmd: string, args: string[], cwd: string, timeoutMs = 180_000): Promise<{ code: number; stdout: string; stderr: string }> {\n  try {\n    const { stdout, stderr } = await execFileAsync(cmd, args, { cwd, maxBuffer: 1 << 26, timeout: timeoutMs })\n    return { code: 0, stdout, stderr }\n  } catch (err) {\n    const e = err as { code?: number; stdout?: string; stderr?: string; message?: string }\n    return { code: typeof e.code === 'number' ? e.code : 1, stdout: e.stdout ?? '', stderr: e.stderr ?? e.message ?? String(err) }\n  }\n}\n\n\n/** Strip markdown fences so we keep just the Python. */\nfunction extractPy(text: string): string {\n  const fence = /```(?:python|py)?\\s*\\n([\\s\\S]*?)```/i.exec(text)\n  return (fence ? fence[1] : text).trim()\n}\n\n/**\n * The standardized Blender runner (written to a temp file per run). It clears the\n * scene, executes the agent's bpy script, auto-frames the produced meshes, sets\n * up neutral lighting, and renders N azimuth views with Cycles CPU.\n */\nconst RUNNER_PY = `\nimport bpy, sys, math, mathutils, traceback, os\nagent_script, outdir, nviews = sys.argv[-3], sys.argv[-2], int(sys.argv[-1])\nbpy.ops.object.select_all(action='SELECT'); bpy.ops.object.delete()\nok=True\ntry:\n    g={'bpy':bpy,'math':math,'mathutils':mathutils,'__name__':'__main__'}\n    exec(compile(open(agent_script).read(), agent_script, 'exec'), g)\nexcept Exception as e:\n    traceback.print_exc(); print('AGENT_SCRIPT_ERROR:'+repr(e)); ok=False\nmeshes=[o for o in bpy.context.scene.objects if o.type=='MESH']\nif not meshes:\n    print('NO_MESH'); sys.exit(0 if ok else 3)\nmn=[1e18]*3; mx=[-1e18]*3\nfor o in meshes:\n    for c in o.bound_box:\n        w=o.matrix_world @ mathutils.Vector(c)\n        for i in range(3): mn[i]=min(mn[i],w[i]); mx[i]=max(mx[i],w[i])\ncenter=mathutils.Vector(((mn[0]+mx[0])/2,(mn[1]+mx[1])/2,(mn[2]+mx[2])/2))\nsize=max(mx[i]-mn[i] for i in range(3)) or 1.0\n# standardize: drop any agent-added cameras/lights\nfor o in list(bpy.context.scene.objects):\n    if o.type in ('CAMERA','LIGHT'): bpy.data.objects.remove(o, do_unlink=True)\nw=bpy.context.scene.world or bpy.data.worlds.new('W'); bpy.context.scene.world=w\nw.use_nodes=True\ntry: w.node_tree.nodes['Background'].inputs[1].default_value=0.6\nexcept Exception: pass\nbpy.ops.object.light_add(type='SUN'); sun=bpy.context.object; sun.data.energy=4.0; sun.rotation_euler=mathutils.Euler((0.6,0.2,0.4))\nbpy.ops.object.camera_add(); cam=bpy.context.object; bpy.context.scene.camera=cam\nsc=bpy.context.scene\nsc.render.engine='CYCLES'; sc.cycles.samples=20; sc.cycles.device='CPU'; sc.cycles.use_denoising=False\nsc.render.resolution_x=640; sc.render.resolution_y=640; sc.render.film_transparent=False\ndist=size*2.4\nel=math.radians(58)\nfor v in range(nviews):\n    az=math.radians(40 + v*360.0/nviews)\n    cam.location=center+mathutils.Vector((math.cos(az)*math.sin(el), math.sin(az)*math.sin(el), math.cos(el)))*dist\n    d=(center-cam.location); cam.rotation_euler=d.to_track_quat('-Z','Y').to_euler()\n    sc.render.filepath=os.path.join(outdir, 'view_%d.png'%v); bpy.ops.render.render(write_still=True)\nprint('RENDER_DONE')\n`.trim()\n\n/** Execute a bpy script headless + render N standardized views — no authoring.\n *  Used by the CADBench judge to render an artifact before vision-scoring it. */\nexport async function renderBpy(script: string, opts: { views?: number } = {}): Promise<{ built: boolean; renders: string[]; error?: string }> {\n  const views = Math.max(1, opts.views ?? 4)\n  const dir = await mkdtemp(join(tmpdir(), 'blender-judge-'))\n  const runnerPath = join(dir, 'runner.py')\n  const scriptPath = join(dir, 'model.py')\n  try {\n    await writeFile(runnerPath, RUNNER_PY)\n    await writeFile(scriptPath, script)\n    const run = await runLocal('xvfb-run', ['-a', 'blender', '--background', '--python', runnerPath, '--', scriptPath, dir, String(views)], dir)\n    const out = `${run.stdout}\\n${run.stderr}`\n    const built = /RENDER_DONE/.test(out)\n    if (!built) return { built: false, renders: [], error: (/(AGENT_SCRIPT_ERROR:.*|NO_MESH)/.exec(out)?.[0] ?? out.trim().slice(-400)) }\n    const renders: string[] = []\n    for (let v = 0; v < views; v++) {\n      const buf = await readFile(join(dir, `view_${v}.png`)).catch(() => undefined)\n      if (buf) renders.push(`data:image/png;base64,${buf.toString('base64')}`)\n    }\n    return { built: renders.length > 0, renders }\n  } finally {\n    await rm(dir, { recursive: true, force: true }).catch(() => {})\n  }\n}\n\nexport interface BlenderLocalConfig {\n  routerBaseUrl: string\n  routerKey: string\n  model: string\n  rounds?: number\n  /** N standardized views to render (CADBench uses 4). Default 4. */\n  views?: number\n  /** The bpy authoring directive — the GEPA-optimizable surface. */\n  directive?: string\n}\n\nexport interface BlenderShot {\n  /** The bpy script the agent wrote — the artifact. */\n  artifact: string\n  /** Rendered view PNGs as data URIs (the images the criteria judge scores). */\n  renders: string[]\n  trace: Span[]\n  usage: { input: number; output: number }\n  ok: boolean\n  /** True if the script executed and produced at least one mesh. */\n  built: boolean\n  detail?: string\n}\n\n/**\n * Author a bpy script for the task via the router, execute + render it headless\n * in Blender, refine on execution errors across rounds. Returns the script, the\n * rendered views, a screenshot-rich trace, and real token usage.\n */\nexport async function solveBlenderLocal(task: BenchTask, cfg: BlenderLocalConfig): Promise<BlenderShot> {\n  const rounds = Math.max(1, cfg.rounds ?? 2)\n  const views = Math.max(1, cfg.views ?? 4)\n  const directive = cfg.directive ?? DEFAULT_BLENDER_DIRECTIVE\n  const trace: Span[] = []\n  const runId = `cadbench-${task.id}`\n  let ts = Date.now()\n  const tick = () => (ts += 1)\n  const usage = { input: 0, output: 0 }\n  // Carried across rounds in closures (the round Artifact is the bpy script; the\n  // render PNGs + built flag + lastErr persist outside the loop). usage is REAL.\n  let renders: string[] = []\n  let built = false\n  let lastErr = ''\n\n  trace.push({ spanId: 's-brief', runId, kind: 'llm', name: 'brief', model: cfg.model, messages: [{ role: 'user', content: task.prompt }], startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n\n  // Migrated onto runRefineLoop: the mkdtemp dir (with the runner.py written once\n  // in setup) is the Ctx; built (RENDER_DONE + ≥1 collected view) is the early-stop,\n  // modeled as a judge so default-decide stops the loop. The round-2+ steer carries\n  // lastErr + the prior script verbatim.\n  const res = await runRefineLoop<string, string>({\n    rounds,\n    setup: async () => {\n      const dir = await mkdtemp(join(tmpdir(), 'blender-'))\n      await writeFile(join(dir, 'runner.py'), RUNNER_PY)\n      return dir\n    },\n    prompt: (round, history) =>\n      round === 1\n        ? task.prompt\n        : `Your previous bpy script failed:\\n${lastErr}\\n\\nPrevious script:\\n${history[history.length - 1]?.artifact ?? ''}\\n\\nFix it so it runs under \\`blender --background --python\\` and builds the object as mesh(es). Brief:\\n${task.prompt}`,\n    runShot: async (user, round, dir) => {\n      const runnerPath = join(dir, 'runner.py')\n      const scriptPath = join(dir, 'model.py')\n      const { content, usage: u } = await routerChatWithUsage(\n        cfg,\n        [\n          { role: 'system', content: directive },\n          { role: 'user', content: user },\n        ],\n        { temperature: 0.3 },\n      )\n      if (u) {\n        usage.input += u.input\n        usage.output += u.output\n      }\n      const script = extractPy(content)\n      trace.push({ spanId: `s-author-${round}`, runId, kind: 'llm', name: `author r${round}`, model: cfg.model, messages: [{ role: 'user', content: round === 1 ? task.prompt : 'refine' }], output: content.slice(0, 600), startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n      trace.push({ spanId: `s-write-${round}`, runId, kind: 'tool', name: 'write_file', toolName: 'create_file', args: { path: 'model.py', content: script }, startedAt: tick(), endedAt: tick(), status: 'ok' } as Span)\n\n      await writeFile(scriptPath, script)\n      const run = await runLocal('xvfb-run', ['-a', 'blender', '--background', '--python', runnerPath, '--', scriptPath, dir, String(views)], dir)\n      const out = `${run.stdout}\\n${run.stderr}`\n      built = /RENDER_DONE/.test(out)\n      lastErr = built ? '' : (/(AGENT_SCRIPT_ERROR:.*|NO_MESH)/.exec(out)?.[0] ?? out.trim().slice(-800))\n      trace.push({ spanId: `s-blender-${round}`, runId, kind: 'tool', name: `blender r${round}`, toolName: 'shell.exec', args: 'blender --background --python runner.py model.py', result: (built ? 'RENDER_DONE' : lastErr).slice(0, 1500), startedAt: tick(), endedAt: tick(), status: built ? 'ok' : 'error', error: built ? undefined : `exit ${run.code}` } as Span)\n\n      if (built) {\n        const collected: string[] = []\n        for (let v = 0; v < views; v++) {\n          const buf = await readFile(join(dir, `view_${v}.png`)).catch(() => undefined)\n          if (buf) collected.push(`data:image/png;base64,${buf.toString('base64')}`)\n        }\n        renders = collected\n        // first view carries the screen span (run-capsule reveal)\n        trace.push({ spanId: `s-render-${round}`, runId, kind: 'tool', name: 'render', toolName: 'render.screenshot', args: { action: 'rendered model', url: 'view_0.png' }, attributes: collected[0] ? { screenshot: collected[0] } : {}, startedAt: tick(), endedAt: tick(), status: collected.length ? 'ok' : 'error', error: collected.length ? undefined : 'render produced no image' } as Span)\n        built = collected.length > 0\n      }\n      return { artifact: script }\n    },\n    judge: async () => ({ valid: built }),\n    teardown: (dir) => rm(dir, { recursive: true, force: true }).then(() => {}, () => {}),\n  })\n\n  const script = res.final.artifact\n  return {\n    artifact: script,\n    renders,\n    trace,\n    usage,\n    ok: script.trim().length > 0,\n    built,\n    detail: built ? `built + rendered ${renders.length} views` : `did not build in ${rounds} rounds${lastErr ? `; last: ${lastErr.slice(0, 140)}` : ''}`,\n  }\n}\n","/**\n * CADBench / BlenderLLM adapter (FreedomIntelligence/CADBench, arXiv:2412.14203).\n * Task = NL instruction → a Blender `bpy` script. Score = the paper's criteria\n * eval: render the produced model to standardized views, then a vision judge\n * (GPT-4o-class) marks each per-task criterion bullet pass/fail against the\n * rendered images + the script text. score = fraction of criteria satisfied.\n *\n * Data: the published dataset's `criteria` flattened to a bullet list (700 tasks,\n * 500 Simulative + 200 Wild). Point CADBENCH_PATH at the cleaned JSONL\n * ({id,name,instruction,type,criteria:string[]} per line). Judge creds from\n * TANGLE_API_KEY / ROUTER_BASE / JUDGE_MODEL (default gpt-4o).\n */\n\nimport { readFile } from 'node:fs/promises'\nimport type { BenchScore, BenchTask, BenchmarkAdapter, LoadOptions } from './types'\nimport { renderBpy } from '../worker-blender'\n\ninterface CadBenchMeta {\n  name: string\n  type: string\n  criteria: string[]\n}\n\nfunction must(name: string): string {\n  const v = process.env[name]\n  if (!v) throw new Error(`env ${name} is required for the CADBench judge`)\n  return v\n}\n\n/** One batched vision call: rendered views + the bpy script + the numbered\n *  criteria → a JSON array of booleans (true = satisfied). Faithful to the\n *  paper's combined image+script evaluation. Throws on transport failure (never\n *  a silent zero); a parse miss falls back to \"all fail\" with a note. */\nasync function judgeCriteria(\n  instruction: string,\n  script: string,\n  criteria: string[],\n  renders: string[],\n): Promise<{ passed: boolean[]; note: string }> {\n  const base = (process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1').replace(/\\/$/, '')\n  const key = must('TANGLE_API_KEY')\n  const model = process.env.JUDGE_MODEL ?? 'deepseek-v4-flash'\n  const numbered = criteria.map((c, i) => `${i + 1}. ${c}`).join('\\n')\n  const text =\n    `You are strictly grading a 3D model that was built by a Blender bpy script for this instruction:\\n\"${instruction}\"\\n\\n` +\n    `Below are ${renders.length} rendered views of the produced model, and the script that built it. ` +\n    `For EACH numbered criterion, decide whether it is satisfied (judge geometry/shape/proportion/structure from the IMAGES; judge color/size/material reasonableness from the SCRIPT where the images are ambiguous). ` +\n    `Return ONLY a JSON array of exactly ${criteria.length} booleans (true=satisfied, false=not), in order, no prose.\\n\\nCRITERIA:\\n${numbered}\\n\\nSCRIPT:\\n\\`\\`\\`python\\n${script.slice(0, 6000)}\\n\\`\\`\\``\n  const content: unknown[] = [{ type: 'text', text }]\n  for (const url of renders) content.push({ type: 'image_url', image_url: { url } })\n  const res = await fetch(`${base}/chat/completions`, {\n    method: 'POST',\n    headers: { 'content-type': 'application/json', authorization: `Bearer ${key}` },\n    body: JSON.stringify({ model, max_tokens: 1500, temperature: 0, messages: [{ role: 'user', content }] }),\n  })\n  if (!res.ok) throw new Error(`judge ${model} ${res.status}: ${(await res.text()).slice(0, 200)}`)\n  const data = (await res.json()) as { choices?: Array<{ message?: { content?: string } }> }\n  const raw = data.choices?.[0]?.message?.content ?? ''\n  const m = /\\[\\s*(?:true|false)[\\s\\S]*?\\]/i.exec(raw)\n  if (!m) return { passed: criteria.map(() => false), note: `judge returned no parseable verdict: ${raw.slice(0, 80)}` }\n  let arr: unknown\n  try {\n    arr = JSON.parse(m[0].toLowerCase())\n  } catch {\n    return { passed: criteria.map(() => false), note: 'judge verdict not valid JSON' }\n  }\n  const bools = Array.isArray(arr) ? arr.map((x) => x === true) : []\n  // Pad/truncate to criteria length (a short array scores the missing as fail).\n  const passed = criteria.map((_, i) => bools[i] === true)\n  return { passed, note: `${passed.filter(Boolean).length}/${criteria.length} criteria` }\n}\n\nexport function createCadBenchAdapter(): BenchmarkAdapter {\n  let cache: Array<{ id: string; instruction: string; meta: CadBenchMeta }> | null = null\n\n  async function load(): Promise<typeof cache & object> {\n    if (cache) return cache\n    const path = process.env.CADBENCH_PATH\n    if (!path) throw new Error('CADBENCH_PATH must point at the cleaned CADBench JSONL ({id,instruction,type,criteria:[]} per line)')\n    const text = await readFile(path, 'utf8')\n    cache = text\n      .split('\\n')\n      .filter((l) => l.trim())\n      .map((l) => {\n        const r = JSON.parse(l) as { id: string; name?: string; instruction: string; type?: string; criteria: string[] }\n        return { id: r.id, instruction: r.instruction, meta: { name: r.name ?? '', type: r.type ?? '', criteria: r.criteria } }\n      })\n    return cache\n  }\n\n  return {\n    name: 'cadbench',\n\n    async preflight() {\n      const { execFile } = await import('node:child_process')\n      const { promisify } = await import('node:util')\n      const exec = promisify(execFile)\n      try {\n        await exec('xvfb-run', ['-a', 'blender', '--version'], { timeout: 30_000 })\n      } catch (err) {\n        throw new Error(\n          `cadbench preflight failed: ${(err instanceof Error ? err.message : String(err)).slice(0, 200)}\\n` +\n            `Fix: install Blender + Xvfb (sudo apt-get install -y blender xvfb). The judge runs \\`xvfb-run -a blender --background --python\\`.`,\n        )\n      }\n      await load()\n    },\n\n    async loadTasks(opts: LoadOptions = {}) {\n      let rows = await load()\n      if (opts.ids) rows = rows.filter((r) => opts.ids!.includes(r.id))\n      // TYPE filter (Simulative|Wild) via env, applied before limit.\n      const t = process.env.CADBENCH_TYPE\n      if (t) rows = rows.filter((r) => r.meta.type.toLowerCase() === t.toLowerCase())\n      if (opts.limit != null) rows = rows.slice(0, opts.limit)\n      return rows.map((r): BenchTask => ({ id: r.id, prompt: r.instruction, metadata: r.meta as unknown as Record<string, unknown> }))\n    },\n\n    async goldArtifact() {\n      return undefined // no reference bpy script ships with the benchmark\n    },\n\n    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n      const meta = task.metadata as unknown as CadBenchMeta\n      const criteria = meta.criteria ?? []\n      if (!artifact.trim()) return { resolved: false, score: 0, detail: 'empty artifact' }\n      if (criteria.length === 0) return { resolved: false, score: 0, detail: 'task has no criteria' }\n      const r = await renderBpy(artifact, { views: 4 })\n      if (!r.built) return { resolved: false, score: 0, detail: `did not build/render: ${r.error ?? 'no mesh'}` }\n      const { passed, note } = await judgeCriteria(task.prompt, artifact, criteria, r.renders)\n      const score = passed.filter(Boolean).length / criteria.length\n      return { resolved: score === 1, score, detail: note }\n    },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAwBA,MAAM,gBAAgB,UAAU,QAAQ;AAExC,eAAe,SAAS,KAAa,MAAgB,KAAa,YAAY,MAAoE;CAChJ,IAAI;EACF,MAAM,EAAE,QAAQ,WAAW,MAAM,cAAc,KAAK,MAAM;GAAE;GAAK,WAAW,KAAK;GAAI,SAAS;EAAU,CAAC;EACzG,OAAO;GAAE,MAAM;GAAG;GAAQ;EAAO;CACnC,SAAS,KAAK;EACZ,MAAM,IAAI;EACV,OAAO;GAAE,MAAM,OAAO,EAAE,SAAS,WAAW,EAAE,OAAO;GAAG,QAAQ,EAAE,UAAU;GAAI,QAAQ,EAAE,UAAU,EAAE,WAAW,OAAO,GAAG;EAAE;CAC/H;AACF;;;;;;AAcA,MAAM,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAwChB,KAAK;;;AAIP,eAAsB,UAAU,QAAgB,OAA2B,CAAC,GAAmE;CAC7I,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,SAAS,CAAC;CACzC,MAAM,MAAM,MAAM,QAAQ,KAAK,OAAO,GAAG,gBAAgB,CAAC;CAC1D,MAAM,aAAa,KAAK,KAAK,WAAW;CACxC,MAAM,aAAa,KAAK,KAAK,UAAU;CACvC,IAAI;EACF,MAAM,UAAU,YAAY,SAAS;EACrC,MAAM,UAAU,YAAY,MAAM;EAClC,MAAM,MAAM,MAAM,SAAS,YAAY;GAAC;GAAM;GAAW;GAAgB;GAAY;GAAY;GAAM;GAAY;GAAK,OAAO,KAAK;EAAC,GAAG,GAAG;EAC3I,MAAM,MAAM,GAAG,IAAI,OAAO,IAAI,IAAI;EAElC,IAAI,CADU,cAAc,KAAK,GACxB,GAAG,OAAO;GAAE,OAAO;GAAO,SAAS,CAAC;GAAG,OAAQ,kCAAkC,KAAK,GAAG,CAAC,GAAG,MAAM,IAAI,KAAK,CAAC,CAAC,MAAM,IAAI;EAAG;EACpI,MAAM,UAAoB,CAAC;EAC3B,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,KAAK;GAC9B,MAAM,MAAM,MAAM,SAAS,KAAK,KAAK,QAAQ,EAAE,KAAK,CAAC,CAAC,CAAC,YAAY,KAAA,CAAS;GAC5E,IAAI,KAAK,QAAQ,KAAK,yBAAyB,IAAI,SAAS,QAAQ,GAAG;EACzE;EACA,OAAO;GAAE,OAAO,QAAQ,SAAS;GAAG;EAAQ;CAC9C,UAAU;EACR,MAAM,GAAG,KAAK;GAAE,WAAW;GAAM,OAAO;EAAK,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;CAChE;AACF;;;;;;;;;;;;;;;AC1FA,SAAS,KAAK,MAAsB;CAClC,MAAM,IAAI,QAAQ,IAAI;CACtB,IAAI,CAAC,GAAG,MAAM,IAAI,MAAM,OAAO,KAAK,oCAAoC;CACxE,OAAO;AACT;;;;;AAMA,eAAe,cACb,aACA,QACA,UACA,SAC8C;CAC9C,MAAM,QAAQ,QAAQ,IAAI,eAAe,iCAAA,CAAkC,QAAQ,OAAO,EAAE;CAC5F,MAAM,MAAM,KAAK,gBAAgB;CACjC,MAAM,QAAQ,QAAQ,IAAI,eAAe;CACzC,MAAM,WAAW,SAAS,KAAK,GAAG,MAAM,GAAG,IAAI,EAAE,IAAI,GAAG,CAAC,CAAC,KAAK,IAAI;CAMnE,MAAM,UAAqB,CAAC;EAAE,MAAM;EAAQ,MAAA,sGAJ4D,YAAY,iBACrG,QAAQ,OAAO,6TAEW,SAAS,OAAO,2EAA2E,SAAS,6BAA6B,OAAO,MAAM,GAAG,GAAI,EAAE;CAC/I,CAAC;CAClD,KAAK,MAAM,OAAO,SAAS,QAAQ,KAAK;EAAE,MAAM;EAAa,WAAW,EAAE,IAAI;CAAE,CAAC;CACjF,MAAM,MAAM,MAAM,MAAM,GAAG,KAAK,oBAAoB;EAClD,QAAQ;EACR,SAAS;GAAE,gBAAgB;GAAoB,eAAe,UAAU;EAAM;EAC9E,MAAM,KAAK,UAAU;GAAE;GAAO,YAAY;GAAM,aAAa;GAAG,UAAU,CAAC;IAAE,MAAM;IAAQ;GAAQ,CAAC;EAAE,CAAC;CACzG,CAAC;CACD,IAAI,CAAC,IAAI,IAAI,MAAM,IAAI,MAAM,SAAS,MAAM,GAAG,IAAI,OAAO,KAAK,MAAM,IAAI,KAAK,EAAA,CAAG,MAAM,GAAG,GAAG,GAAG;CAEhG,MAAM,OAAM,MADQ,IAAI,KAAK,EAAA,CACZ,UAAU,EAAE,EAAE,SAAS,WAAW;CACnD,MAAM,IAAI,iCAAiC,KAAK,GAAG;CACnD,IAAI,CAAC,GAAG,OAAO;EAAE,QAAQ,SAAS,UAAU,KAAK;EAAG,MAAM,wCAAwC,IAAI,MAAM,GAAG,EAAE;CAAI;CACrH,IAAI;CACJ,IAAI;EACF,MAAM,KAAK,MAAM,EAAE,EAAE,CAAC,YAAY,CAAC;CACrC,QAAQ;EACN,OAAO;GAAE,QAAQ,SAAS,UAAU,KAAK;GAAG,MAAM;EAA+B;CACnF;CACA,MAAM,QAAQ,MAAM,QAAQ,GAAG,IAAI,IAAI,KAAK,MAAM,MAAM,IAAI,IAAI,CAAC;CAEjE,MAAM,SAAS,SAAS,KAAK,GAAG,MAAM,MAAM,OAAO,IAAI;CACvD,OAAO;EAAE;EAAQ,MAAM,GAAG,OAAO,OAAO,OAAO,CAAC,CAAC,OAAO,GAAG,SAAS,OAAO;CAAW;AACxF;AAEA,SAAgB,wBAA0C;CACxD,IAAI,QAA+E;CAEnF,eAAe,OAAuC;EACpD,IAAI,OAAO,OAAO;EAClB,MAAM,OAAO,QAAQ,IAAI;EACzB,IAAI,CAAC,MAAM,MAAM,IAAI,MAAM,qGAAqG;EAEhI,SAAQ,MADW,SAAS,MAAM,MAAM,EAAA,CAErC,MAAM,IAAI,CAAC,CACX,QAAQ,MAAM,EAAE,KAAK,CAAC,CAAC,CACvB,KAAK,MAAM;GACV,MAAM,IAAI,KAAK,MAAM,CAAC;GACtB,OAAO;IAAE,IAAI,EAAE;IAAI,aAAa,EAAE;IAAa,MAAM;KAAE,MAAM,EAAE,QAAQ;KAAI,MAAM,EAAE,QAAQ;KAAI,UAAU,EAAE;IAAS;GAAE;EACxH,CAAC;EACH,OAAO;CACT;CAEA,OAAO;EACL,MAAM;EAEN,MAAM,YAAY;GAChB,MAAM,EAAE,aAAa,MAAM,OAAO;GAClC,MAAM,EAAE,cAAc,MAAM,OAAO;GACnC,MAAM,OAAO,UAAU,QAAQ;GAC/B,IAAI;IACF,MAAM,KAAK,YAAY;KAAC;KAAM;KAAW;IAAW,GAAG,EAAE,SAAS,IAAO,CAAC;GAC5E,SAAS,KAAK;IACZ,MAAM,IAAI,MACR,+BAA+B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAA,CAAG,MAAM,GAAG,GAAG,EAAE,oIAEjG;GACF;GACA,MAAM,KAAK;EACb;EAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;GACtC,IAAI,OAAO,MAAM,KAAK;GACtB,IAAI,KAAK,KAAK,OAAO,KAAK,QAAQ,MAAM,KAAK,IAAK,SAAS,EAAE,EAAE,CAAC;GAEhE,MAAM,IAAI,QAAQ,IAAI;GACtB,IAAI,GAAG,OAAO,KAAK,QAAQ,MAAM,EAAE,KAAK,KAAK,YAAY,MAAM,EAAE,YAAY,CAAC;GAC9E,IAAI,KAAK,SAAS,MAAM,OAAO,KAAK,MAAM,GAAG,KAAK,KAAK;GACvD,OAAO,KAAK,KAAK,OAAkB;IAAE,IAAI,EAAE;IAAI,QAAQ,EAAE;IAAa,UAAU,EAAE;GAA2C,EAAE;EACjI;EAEA,MAAM,eAAe,CAErB;EAEA,MAAM,MAAM,MAAiB,UAAuC;GAElE,MAAM,WADO,KAAK,SACI,YAAY,CAAC;GACnC,IAAI,CAAC,SAAS,KAAK,GAAG,OAAO;IAAE,UAAU;IAAO,OAAO;IAAG,QAAQ;GAAiB;GACnF,IAAI,SAAS,WAAW,GAAG,OAAO;IAAE,UAAU;IAAO,OAAO;IAAG,QAAQ;GAAuB;GAC9F,MAAM,IAAI,MAAM,UAAU,UAAU,EAAE,OAAO,EAAE,CAAC;GAChD,IAAI,CAAC,EAAE,OAAO,OAAO;IAAE,UAAU;IAAO,OAAO;IAAG,QAAQ,yBAAyB,EAAE,SAAS;GAAY;GAC1G,MAAM,EAAE,QAAQ,SAAS,MAAM,cAAc,KAAK,QAAQ,UAAU,UAAU,EAAE,OAAO;GACvF,MAAM,QAAQ,OAAO,OAAO,OAAO,CAAC,CAAC,SAAS,SAAS;GACvD,OAAO;IAAE,UAAU,UAAU;IAAG;IAAO,QAAQ;GAAK;EACtD;CACF;AACF"}