wrangler.jsonc

{
  "$schema": "node_modules/wrangler/config-schema.json",
  "name": "fly-dev",
  "main": "worker/src/index.ts",
  "compatibility_date": "2026-05-24",
  "compatibility_flags": ["nodejs_compat"],
  "assets": {
    "directory": "dist",
    "binding": "ASSETS",
    // Run the Worker BEFORE the assets layer so the auth gate fires on page
    // navigations (otherwise "/" -> index.html is served by assets directly,
    // bypassing the gate). The catch-all serves static sub-resources via the
    // ASSETS binding after the gate check.
    "run_worker_first": true,
    // SPA routing: the default "auto-trailing-slash" 302-redirects /index.html and
    // any client route (e.g. /command-center) toward "/", so deep links bounced to
    // the dashboard. "none" stops those redirects; "single-page-application" serves
    // index.html (200) for any unmatched path so the client router renders the route.
    "html_handling": "none",
    "not_found_handling": "single-page-application"
  },
  "routes": [
    {
      "pattern": "dev.fly.pm",
      "custom_domain": true
    }
  ],
  "workers_dev": false,
  "preview_urls": false,
  // Self-healing reaper + autonomous Linear project watcher: both run every 5 minutes.
  "triggers": {
    "crons": ["*/5 * * * *"]
  },
  "tail_consumers": [
    {
      "service": "fly-dev-tail"
    }
  ],
  "observability": {
    "enabled": true,
    "head_sampling_rate": 1,
    "logs": {
      "enabled": true,
      "head_sampling_rate": 1,
      "persist": true,
      "invocation_logs": true
    },
    "traces": {
      "enabled": true,
      "persist": true,
      "head_sampling_rate": 1
    }
  },
  "ai": {
    "binding": "AI"
  },
  "browser": {
    "binding": "MYBROWSER"
  },
  "dispatch_namespaces": [
    {
      "binding": "USER_WORKERS",
      "namespace": "fly-dev-production"
    }
  ],
  "kv_namespaces": [
    {
      "binding": "CACHE",
      "id": "486f8ccff1744020a1d8427f1d60ccdc"
    },
    {
      "binding": "SESSION_CACHE",
      "id": "53f23a4907da4f3e9a97d40c1c4f1f9a"
    }
  ],
  "r2_buckets": [
    {
      "binding": "ARTIFACTS",
      "bucket_name": "fly-dev-artifacts"
    },
    {
      "binding": "TEMPLATE_ASSETS",
      "bucket_name": "fly-dev-template-assets"
    }
  ],
  "d1_databases": [
    {
      "binding": "DB",
      "database_name": "fly-dev",
      "database_id": "3f03fde5-d0ba-4f35-940e-318567b45cab",
      "migrations_dir": "migrations"
    }
  ],
  "queues": {
    "producers": [
      {
        "binding": "WORK_QUEUE",
        "queue": "fly-dev-work"
      },
      {
        "binding": "MEMORY_QUEUE",
        "queue": "fly-dev-memory"
      }
    ],
    "consumers": [
      {
        // Agent-memory retains, off the run hot path. Generous retries because a
        // brief Hindsight/VM blip should redeliver, not dead-letter, a memory.
        "queue": "fly-dev-memory",
        "max_batch_size": 10,
        "max_batch_timeout": 5,
        "max_retries": 5,
        "dead_letter_queue": "fly-dev-memory-dlq"
      },
      {
        "queue": "fly-dev-work",
        // Serialize queue ADMISSION (max_concurrency: 1) so getActiveRunCount()
        // can't be read stale by parallel autoscaled consumer invocations racing
        // the count-then-create capacity gate. Throughput ceiling is 8 containers
        // anyway, so a single admission consumer is acceptable. Combined with the
        // per-batch reservation in queue(), this closes the cross-invocation race.
        "max_concurrency": 1,
        "max_batch_size": 8,
        "max_batch_timeout": 5,
        "max_retries": 3,
        "dead_letter_queue": "fly-dev-work-dlq"
      }
    ]
  },
  "workflows": [
    {
      "name": "fly-dev-run-workflow",
      "binding": "RUN_WORKFLOW",
      "class_name": "RunWorkflow"
    }
  ],
  "durable_objects": {
    "bindings": [
      {
        "name": "PROJECT_CONDUCTORS",
        "class_name": "ProjectConductor"
      },
      {
        "name": "USER_WORKER_CONTROLLERS",
        "class_name": "UserWorkerController"
      },
      {
        "name": "DEV_ORCHESTRATOR",
        "class_name": "DevOrchestratorAgent"
      },
      {
        "name": "SANDBOX_CONTAINER",
        "class_name": "SandboxContainer"
      }
    ]
  },
  "migrations": [
    {
      "tag": "v1",
      "new_sqlite_classes": [
        "ProjectConductor",
        "UserWorkerController",
        "DevOrchestratorAgent",
        "SandboxContainer"
      ]
    }
  ],
  "containers": [
    {
      "class_name": "SandboxContainer",
      "image": "./container/Dockerfile",
      // standard-3 specs (2 vCPU / 8 GiB / 16 GB disk), set EXPLICITLY as a custom
      // instance type because the named "standard-3" string did not apply the larger
      // resources on this wrangler/platform version (the deployed instance stayed at
      // 1 vCPU / 6 GiB). More CPU/disk for clone + agent + test gate, and a different
      // capacity pool than the constrained standard-2 ("no free slot").
      "instance_type": { "vcpu": 2, "memory_mib": 8192, "disk_mb": 16000 },
      "max_instances": 8
    }
  ],
  "vars": {
    "APP_ENV": "production",
    "APP_URL": "https://dev.fly.pm",
    "AI_GATEWAY_ID": "x",
    "CODEX_GATEWAY_MODEL": "openai/gpt-5.5",
    "CLOUDFLARE_GATEWAY_MODEL": "@cf/openai/gpt-oss-120b",
    "APPLE_TEMPLATE_REPO": "aloewright/warp-template-apple-multiplatform",
    "CLOUDFLARE_TEMPLATE_REPO": "aloewright/aloe-template-cloudflare-fullstack",
    "FIRECRAWL_WORKER_URL": "https://firecrawl-cf.lazee.workers.dev",
    "FLY_AUTH_ORIGIN": "https://fly.pm",
    "PLAYWRIGHT_WORKER_URL": "",
    "REQUIRE_HUMAN_APPROVAL": "false",
    "GITHUB_APP_ID": "3899646",
    "AUTH_HUB_URL": "https://auth.fly.pm",
    "CONTINUE_AUTONOMY": "true",
    "CONTINUE_EXECUTE_CAP": "3",
    // Queue-admission gate: how many sandboxes may run concurrently. Measured: with
    // self-exit + orphans cleared, the account still provisions only ~1-2 standard-3
    // containers at once (shared with ~6 other container apps), and gate=4 produced
    // capacity_exhausted churn. Set to 2 (real concurrency, minimal overshoot); the
    // bounded-requeue absorbs the rest. Raise only if 2 provisions cleanly with headroom.
    "MAX_CONTAINER_INSTANCES": "2",
    // Agent memory (self-hosted Hindsight). Flip MEMORY_ENABLED to "true" to turn on
    // recall/retain/reflect; secrets (HINDSIGHT_API_KEY, HINDSIGHT_CF_ACCESS_CLIENT_ID/
    // SECRET) are set via `wrangler secret put`. Server lives in infra/hindsight/.
    "MEMORY_ENABLED": "true",
    "MEMORY_REFLECT_MODE": "auto",
    "HINDSIGHT_BASE_URL": "https://hindsight.fly.pm"
  }
}